Markus: LLMs for LL.Ms

From campus to courtroom

Tue, 19 May 2026 08:16:36 GMT

Newsletter #27

Read to the end for some great Dutch input on what to do with AI

A word from the person behind the laptop

For the last few years I've been thinking about how you actually close the adoption gap for lawyers beyond just buying expensive licenses and hoping people use them.

I don't think lawyers are uniquely bad at adoption, but I am curious how we solve it specifically for them.

That’s the newsletter today.

Enjoy!

Wednesday night in Aarhus

As I walk into Aarhus University, a bunch of law students are already there, but they’re not here for the workshop we’re about to run. I’m curious to see how many actually show up to a “Vibecoding Masterclass” on a normal Wednesday with the sun shining outside.

I speak to Philip, the chairman of the local legaltech organisation: “We’re expecting quite a few people here today,” he says. The association has a few “hardcore” members who show up for pretty much every event he puts together.

“The interest is definitely there, but I think there are actually quite a few students out there working on their own AI thing.”

Mads Høegh talking about real life AI use cases

When the workshop starts, 12-15 students had booked their Wednesday evening to vibecode with their peers, which was a pleasant surprise. I know how hard it is to get students together for anything not directly linked to grades or a specific class, so I have to hand it to Philip for making this happen.

Now the question is what we’ll be able to achieve in the time we have together.

Getting started

After a one-hour intro covering AI in the legal sector, the upsides and downsides of vibe coding, and a framework for getting started on projects, we moved on to the actual building.

Some initial problems presented themselves when Google AI Studio turned out to be down. My choice for the day, mainly because the interface is easy to deal with and you can get pretty far without a paid subscription.

We quickly switched everyone over to Codex. The students were already familiar enough with the interface to get going, even though it looks a bit different from regular ChatGPT.

About half the teams started working on study-related problems. The other half worked on legal solutions tied to their student jobs. They all had reasonably good data to kick things off, though some datasets turned out to be a bit too large for the context window to handle.

I was surprised that almost none of the students had been vibe coding their own projects in their spare time. But it confirmed something I’d already suspected: people don’t have time.

There’s a widespread assumption that students have more bandwidth than people with full-time jobs. I’d push back on that based on my experience.

What they actually built

With two of us supporting the teams, and despite a few technical hiccups, everyone presented something genuinely interesting after just two to three hours.

My personal favourite was a dashboard for students studying for a specific exam. The app presented which topics had been most prominent in past exams and guided students on where to spend their time. I thought it was legitimately brilliant, and honestly something I’d have used myself as a student.

Here’s what the solution looks like. Amazing job by Jonas, Oskar and Magnus. Unfortunately, its only available in Danish, but you can access the full version here.

Other strong solutions included a tool supporting legal interns with drafting standard inheritance documents for clients, and a system for organising knowledge from a particularly unstructured professor. If you’ve been to university, you know exactly the type.

The adoption gap will persist

I’ve been teaching this stuff for years, and I still get surprised by how large the gap is between what the models can do and what actually gets built.

That’s not a criticism of what the students developed. Quite the opposite. They’re working at the frontier. But there’s still a significant distance between what we were doing that day and what’s genuinely achievable with these tools.

I say this as someone who feels behind himself. There’s so much to test and so many assumptions to pressure-check that it’s hard to know where to start.

Numbers show that even among the top 20% of ChatGPT users, total messages sent in all of 2025 topped out at around 1,500. That’s less than five prompts a day.

This means that AI probably isn’t a daily essential for most people yet. Fitting AI adoption (whatever that means) into an already full life is harder than the hype suggests unless you’re dedicating a vast amount of time to this.

“The consumer use is a mile wide and an inch deep”. Same goes for implementation with big businesses incl. law firms.

Numbers from the latest presentation from Benedict Evans

This got me wondering: does the same dynamic hold for all lawyers or just the students. I reached out to a person that know this much better than I do.

From students to practitioners: a conversation with Elgar Weijtmans

My experience is primarily with students, so I reached out to Elgar Weijtmans. Weijtmans is a former lawyer from the Netherlands who now supports legal professionals with AI implementation.

He’s run a series of hackathons for practising lawyers, across two days and three separate sessions, and I wanted to hear what he’d seen on the other side of graduation.

The participants ranged from in-house counsel to associates at firms, with one session dedicated to students at a university of applied sciences. The common thread, Weijtmans thinks, was curiosity.

“Most of these people have had ideas brewing for a while, things like ‘we really need a better way to handle X,’ or frustrations with a specific process. AI has made it impossible to look away, but reading about it only gets you so far. This was the first time many of them could actually try to build something themselves without learning to code first. That tends to get people through the door.”

The ideas that came out of the room were all over the place, which he takes as a good sign. They encouraged people to pick low-stakes topics where a mistake wouldn’t hurt anyone, just to get the momentum going.

One team built “Legally Fit,” a wellness app for better work-life balance among lawyers. Another made a tool to review settlement agreements for employees.

“Basically a quick sanity check on whether anything unusual is hiding in the fine print. That one I thought was excellent, a real everyday pain point turned into a working prototype in a few hours.”

Pizza is mandatory for hackathons.

What surprised him the most was the attitude in the room. “Lawyers can be cautious by nature, and rightly so, but everyone showed up with real enthusiasm.”

A deliberate design choice helped: they sent a video beforehand covering all the risks and limitations of vibe coding. We did the same in Aarhus btw.

With lawyers, if you don’t address that upfront, your entire hackathon becomes a risk workshop. By getting it out of the way early, people felt free to just explore. “What really got me is that everyone left energised and inspired. Not most people. Everyone. That doesn’t happen often.”

On the tooling side, the choice of Claude Code was completely deliberate. “In my view it’s the most complete coding agent available right now for this kind of work. There are plenty of alternatives, Cursor, Lovable, and others, but the pace of development around Anthropic’s tools is remarkable and the quality is consistently high.”

And the simple truth, he says, is that it’s what he uses every single day. If you’re going to teach people how to vibe code, you’d rather show them how you actually do it than demo something you only know in theory.

As for how much of what people built was actually usable the following Monday, Weijtmans draws a clear line between the skills and the artifacts. The skills, understanding what an agent is, how prompting works, what the building process feels like, are immediately transferable.

The apps themselves were never meant to be production-ready. But most teams ended up with something that could genuinely serve as a foundation. “Some of the simpler, lower-risk projects were honestly an afternoon away from being deployable.”

The wrong question

That said, he thinks the fixation on immediate usability misses the point. “Vibe coding has only been feasible since roughly December. The technology is moving so fast that building intuition now is more valuable than shipping something today.

The moment where non-developers can create genuinely useful and safe tools is coming, and it’s coming soon. So the better question isn’t ‘can I use this on Monday?’ but ‘will I be ready when that moment arrives?’”

The question right now is not whether AI will change how you work because it already is. The more interesting finding from these sessions is how quickly lawyers develop intuition for what it means for their specific work once they actually try something.

That's what a Wednesday night in Aarhus or a legal hackathon in the Netherlands gives you.

And that’s why it’s worth it in the long run.

Subscribe now

You can follow Elgar’s work and thinking at Unfiltered Bits.

You can see a short video about the Aarhus hackathon here (English Subtitles):

Law students don't care about your AI strategy

Thu, 16 Apr 2026 08:29:47 GMT

Newsletter #26

Read to the end for an idea

A word from the person behind the laptop

Every time someone asks me about AI adoption among law students, I’ve given the same answer: they’re ahead of the average. Younger people adopt faster. Studies back it up. My own conversations with students backed it up.

Then I actually conducted a survey to find out.

A few weeks ago, Harvey and Legora both announced partnerships with top US and UK law schools. Stanford, Cornell, Oxford, King’s College London. Free access, curriculum integration, the works. Harvey is now in 25+ American law schools. Legora just launched its Legal AI Scholars Program at nine more.

You’d think AI is about to eat legal education whole if you look at those partnerships.

But then I surveyed 53 Danish law students. I talked to student associations, posted on Facebook like a proper boomer, and joined the friday bar at KU.

The gap between those two realities is what this story is about.

How I collected the data

I surveyed 53 Danish law students across semesters and campuses and had a more in depth qualitative convo with 5 more.

The survey was in Danish, distributed through student networks, AI-focused groups, and yes, in person at a friday bar. I wanted a mix. Not just the AI-curious crowd, but the people who don’t think about this stuff at all. Semester distribution covers everything from second-semester undergrads to students finishing their master’s degree.

I’m not going to pretend 53 responses is a representative sample of all Danish law students. It isn’t. But it’s enough to spot patterns and confirm some assumptions I had.

AI is shadow IT in the classroom

The first question was simple: have you received any form of AI training as part of your law degree?

70% said no.

That tracks with what I saw during my own time as an external lecturer, and with what Philip Dalmose told me when I interviewed him for the newsletter back in March. He founded a student AI association in Aarhus specifically because there were zero technology courses in the curriculum. “There are currently no courses related to AI or technology available,” he said at the time.

It doesnt surprise me that this hasn't changed much. The 30% who said yes is probably inflated by the fact that some respondents came through AI-focused student networks. If I’d only surveyed at the friday bar, I suspect it would be closer to 20%. In other words around the general population average.

A win for Sam Altman

Sam Altman has had a rough few weeks with Anthropic launching new functions every 3 hours and winning a bigger chunk of the market plus some wild accusations in this piece by The New Yorker. When I asked which tools students use, the answer was overwhelmingly one thing: ChatGPT.

Perhaps he would be happy to know that he’s winning student consumer in Denmark. Or perhaps he doesn’t care. I do though.

43 out of 53 respondents use ChatGPT. That’s 81%. The next most common tools were Microsoft Copilot and legal-specific platforms like Legora, Harvey or Pandektes (a local DK company), both around 25%. Claude came in at about 21%.

But look at it from the other direction: Roughly a third, use ChatGPT and nothing else. Three respondents don’t use any AI tools at all. The “multi-tool” user who has ChatGPT, Claude, Copilot, and a legal AI platform open in different tabs? That person exists in this dataset, but they’re the exception.

One student told me “AI just doesn’t take up that much attention”.

For many of these students, AI is ChatGPT, and ChatGPT is something you occasionally ask a question when you’re stuck. Either as part of studying or something else, but the legal workflows are not top of mind for many students.

The frequency picture is less dramatic than you’d think

About 43% of respondents say they use AI daily. That sounds high until you remember that “daily use” for a student might mean asking ChatGPT to put together a recipe for a nice spag bol.

It doesn’t necessarily mean they’re running contract analysis or doing legal research with it.

23% use AI weekly. Another 23% say “once in a while.” And 8% never use AI tools at all.

When I was teaching, adoption felt slightly lower than this, but it really hasn’t changed as much as I expected over the last couple of years. I think that says something about what drives adoption in the first place.

It’s curiosity and willingness to spend time getting familiar with the tools. Without a structured reason to learn, most students default to whatever is easiest and most familiar. Right now, that means ChatGPT in the browser used casually.

What they actually want help with

The most requested training area, by far, was legal research. 42 out of 53 respondents picked it. Case analysis came second at 25, followed by workflow automation at 21 and general AI understanding at 20.

Students are saying, pretty clearly, that they want AI for legal research specifically. But they’re mostly using ChatGPT, which is a general-purpose tool that wasn’t built with Danish case law in mind.

The question is why. Is it because they’ve tried legal research tools and found them insufficient? Or is it because they haven’t tried them at all and are projecting what they think they’d want? I suspect it’s mostly the latter. The tools exist, but the awareness of the tools is still low based on my experience and the reponses to the survey,

This is obviously the gap the Harvey and Legora law school partnerships are trying to close. But those partnerships are currently being rolled out at Stanford and Oxford, not at Aarhus or Copenhagen. Danish law students are, for now, largely on their own.

Why nobody has cracked this market yet

Firstly, I want to say this: students are a notoriously awkward group to sell to.

On paper, they’re the perfect early adopters. Curious and digitally native. But they’re also mostly broke. Paying €500 per seat for a legal AI platform might be achievable for some trustfund kids, but no one else.

Even my own survey data points the same way: when I asked what students would pay for a one-day AI workshop, plenty of answers landed at “0”. A couple even wrote that they expected it to be included in tuition.

That leaves two paths.

One is the direct-to-student free tier, which is basically what ChatGPT already is. The general-purpose tools have won this battle by default. They’re free, they’re good enough for most student tasks, and there’s zero switching cost.

The other is going through the university, which is what Harvey and Legora are trying to do in the US and UK. That works, but it’s slow.

Faculties have their own procurement cycles, their own agendas, and a healthy fear of getting locked into a vendor that raises prices 20% a year once the students who trained on it start working at law firms. There are also real quality and academic integrity concerns that an enterprise sales pitch doesn’t always address well. Does the tool hallucinate on Danish case law? How do exam rules handle it? Who owns the data students put into it?

They’re going to evaluate carefully, argue about it in committees, and probably move slower than the market expects.

Which brings us back to the students, who in the meantime will just keep using ChatGPT. And get familiar with the platform which means even more usage in the long run.

The employer question

I expected students to care about whether their future employer has an AI strategy. The data says otherwise.

64% said no. Only 21% said they’d factor an employer’s AI tools into their decision. When I interviewed Julius Sejersen for the March edition, he said the opposite: “To put it very simply, I would.” But Julius was building an AI startup alongside his degree. He’s probably not the median respondent here.

For most students, AI is still something that happens in the background. A nice-to-have, not a deciding factor when you’re picking between law firms. This might change once they actually start working and realize that the firm down the hall has tools that save four hours a week on due diligence. For now, it doesn’t really register.

Turns out law students care about... law. Who would have thought.

So are we overestimating student adoption?

Based on this data, I think the honest answer is: yes, a little.

The narrative around legal AI tends to zoom in on the early adopters. The Philip Dalmoses and Julius Sejersens of the world, the students founding AI associations and building legal AI startups. Those stories are real and worth telling. I told them myself in the March edition.

But they’re not representative. The median Danish law student uses ChatGPT a few times a week, has never touched a legal-specific AI tool, has received no formal AI training, and doesn’t particularly care whether their future employer has an AI strategy.

And unsurprisingly it’s not that different from adoption patterns in working law firms, where the gap between what’s theoretically possible with AI and what lawyers actually do on a daily basis remains wide. Anthropic's labor market research puts daily AI adoption in legal practice at around 15%, against a theoretical capability of 80%. Students are just mirroring the profession they're about to enter.

An idea

Students want legal research capabilities but don’t have access to the right tools.

Most students want training but their universities aren’t providing it. They’re curious but not curious enough to invest significant time on their own.

What some AI curious students might actually need is a smaller, more practical community. If 5-10 law students meet regularly to discuss tools, build demos and want to get familiar with the intersection of law and AI, perhaps that will enable them to move to the next level.

Not a course with credits and exams and not a platform partnership, but something more like a study group including mentoring.

I’m considering setting something up myself. It wouldn’t be affiliated with any university, just an open group where students can work with the different tools together, and build solutions.

The demand is there, and based on my conversations there's clearly a segment that has picked up on AI and is ready to move faster than their surroundings allow. This would be for them.

If you’re a law student and that sounds like something you’d show up for, just shoot me a message via Linkedin and I’ll be in touch.

Law students are quietly rewriting legal education with AI

Tue, 13 Jan 2026 09:22:59 GMT

Newsletter #25

Read to the end for a free trip to Prague

A word from the person behind the laptop

At a law faculty in Aarhus, a third-semester student is founding an AI association because his curriculum doesn’t mention technology at all.

In Copenhagen, a young legal associate is building a startup to help students navigate their own university.

And in Prague, a law lecturer has already decided that generative AI is “the new business-as-usual” in her classroom.

Together, they sketch a picture of legal education in transition: AI is everywhere in the profession, but still only patchily present in law school.

I’ve always been drawn to what happens at the universities. Partly because of my role as an external lecturer, but also out of pure interest in “what’s going on” with law students right now.

People (in this case law students and teachers) will shape the future of law more than AI will. So how are they actually working with AI and what do their teachers make of it? For this issue, I spoke to two students, Philip and Julius, and to a teacher who is already redesigning her class around AI, Barbora.

Students taking AI into their own hands

When Philip Haakon Dalmose looked at the job market, one thing felt obvious: whatever else happens in law, AI isn’t going away.

“Employers expect a certain degree of AI literacy in their new hires” says Dalmose, a third-semester law student at Aarhus University. Yet on his campus, “there are currently no courses related to AI or technology available in the curriculum.”

So he did what law schools have not yet done. He started a student association dedicated to AI in law, with the explicit goal of “future-proofing” at least part of their education.

“Employers expect a certain degree of AI literacy in their new hires,”

The association draws a wide range of students. From master’s students who already use tools to first-years who are openly skeptical and “critical of integrating AI into their schoolwork.” The common denominator, Dalmose suspects, is uncertainty surrounding technical qualifications needed and how to apply them in practice.

Over the next year, the association plans a series of practical, case-based events where students can try out legal AI platforms on realistic problems. A hands-on laboratory that doesn’t yet exist in the official curriculum.

Philips association markets itself with a focus on AI, but also legaltech in general

Cheating and AI use isn’t the same

Dalmose’s project is also about creating a space for honesty.

He says that, right now, conversations about AI use “rarely come up at all.” Those who do rely on AI tools often hesitate to admit it. Worried either that they’ll be seen as having discovered a “secret sauce” for grades, or dismissed as lazy.

“I’m sure many fellow students can attest to having met others who have labeled AI use as ‘laziness’ or ‘cheating,’” he says. Some classmates have gone further, making it “their mission to avoid the use of these tools at any cost, even refusing to incorporate them in exams where generative AI is allowed.”

“I’m sure many fellow students can attest to having met others who have labeled AI use as ‘laziness’ or ‘cheating.’”

His impression is that most law students still “rarely use AI in any context or situation, whether it be work, school, or personal assistance.” In that sense, he thinks, existing exam rules, often restrictive or cautious about AI, “correlate with how students work with AI in general.”

That, he adds, is precisely what he hopes will change.

“AI is already deeply integrated into many of our future workspaces,” Dalmose says. The question is whether universities will help students prepare. Else they’ll have to figure it out themselves.

Perhaps universities are not that different from regular jobs in that regard.

Course chaos calls for AI solutions

A few hundred kilometers away in Copenhagen, another law student has responded differently to the same sense of mismatch between what students need and what their institutions provide.

“I’m building an AI student assistant that is able to compile course data, peer reviews, etc. to help students make more informed decisions faster,” says Julius Sejersen, a legal intern at a big Danish law firm who is about to finish his master’s degree at the University of Copenhagen.

The tool is meant to solve a very practical problem: basic orientation, not law. In Denmark, he explains, students must often “navigate 5–10 websites/platforms, social media etc. just to get the information needed to make informed decisions.” His goal is simple: “I would like to change that.”

“I think that we as lawyers in the future will be evaluated heavily on our AI skills, maybe as much as on the skills we’ve traditionally learned.”

Sejersen sees plenty of “healthy” AI use among his peers. Students use tools to tighten their language, sharpen arguments, and create their own study materials. Flashcards, exam questions, and structured notes. “Here, there are real opportunities for optimizing your studies,” he says.

But he is also worried about habits of mind.

“With the current state of AI, especially in the field of law, it is a little dangerous,” he says. The risk, as he sees it, is that students make AI their first instinct instead of their last.

“You quickly get used to asking some kind of chatbot as the first step instead of actually applying the correct methodology,” he warns. “If you get used to AI from day one of law school, and that AI hallucinates (and does so very convincingly), I fear that we will educate less competent lawyers in the future.”

For Sejersen, the answer is better AI, and better training.

“I expect law firms to invest in good tools and to invest heavily in educating their employees,” he says. Being a first mover in the AI space, he believes, will be “critical for future growth.” Asked whether he would choose an employer based on its AI strategy, he doesn’t hesitate: “To put it very simply: I would.”

In fact, he goes even further. “I think that we as lawyers in the future will be evaluated heavily on our AI skills, maybe as much as on the skills we’ve traditionally learned,” he says. For junior lawyers in three to five years, “it is paramount that you have strong AI skills” if you want to be among the best.

Is this what the future classroom looks like?

If students like Dalmose and Sejersen are building from the bottom up, Barbora Obračajová is one of the teachers trying to redesign legal education from the perspective of a teacher.

Barbora - or Baru - teaches Modern Lawyers, a small, data-driven legal design class at the Faculty of Law at Charles University in Prague. For her, generative AI is already “the new business-as-usual.”

To my question about what it’s like to teach AI to law students right now, she responds enthusiastically:

“I love that!” she says. GenAI and vibecoding have dramatically lowered the bar for students to get something concrete on the table. Before, her students were making mockups in Canva and stitching screens together in MarvelApp. Slow, manual work before they could even show what they had in mind.

Barboras structure as described on her website

Now, she says, “we can get prototypes off the ground much quicker, be it clickable vibecoded apps, playing with different information architecture of legal documents, or generating imagery to illustrate the steps in a legal service. Our iteration is faster, discussions are more aligned, and we can focus on the ideas with much less friction.”

Reflections on what the tools can and cannot do also go deeper “when you can play with it in real time alongside others.”

“For me, cheating is the absence of thought process.”

The reception from students is, in her words, “very positive.” She collects feedback throughout the semester, and students are keen to learn with GenAI and to discover new tools and use cases.

But she also sees two recurring challenges.

First, nerves before the first class. Her format is different from a traditional law lecture “we play, riff, and move tables,” as she puts it, so she starts with small, simple challenges before turning up the difficulty.

Second, a distinct kind of AI FOMO. With GenAI moving fast and only ninety minutes a week together, “we only manage to cover a tiny part.” Students often wonder whether they’re “getting the most of the tech.” She recognises the feeling herself and treats “learning to navigate the ambiguity” as part of the skill set.

On her blog, she describes similar patterns: students using AI to explain concepts, restructure long legal texts, or prototype tools and then using class time to debate whether the outputs actually make sense.

Agency will be the main driver

Baru is careful not to overgeneralise. Modern Lawyers is a small elective, not a full curriculum.

“I do not have the data to make a generalized statement, as I teach only a small elective,” she notes when I ask her to compare her classroom to faculties where AI is barely visible. Every semester, she has a couple of students who have never really used AI before “or maybe once or twice” and a few “hyper power users that vibe with Claude Code like a pro.”

“There is definitely a range,” she says. “Not to be fixed, but to be designed for.” Her course is one part of a much larger ecosystem that, in her view, has to make room for every level of proficiency and interest: those who dive deep into AI, and those who barely touch it.

When the discussion turns to cheating versus innovation, she doesn’t start with tools at all she starts with meaning.

“For me, cheating is the absence of thought process,” she says.

Generative AI makes it trivial to produce plausible outputs, and that “puts a new kind of pressure on us as educators to formulate the value of a task.” Instead of trying to block AI, she argues, teachers should design assignments so that students can “ethically flex all of their muscles and focus on the process of making.”

“creating a space for safe and productive exploration is the first and most important purpose of higher education.”

“If the task is uninspiring and hollow or the students do not see the point in spending time on it either because there is none or it has not been well communicated. There will always be ingenious ways to minimize workload to give the space to more impactful things.”

Her own internal benchmark is concrete: “If a student gives me 18+ hours of their life, they should get and feel like they are getting something from it.”

That connects to how she sees her students: not children to be policed, but adults with agency. “My students are adults in charge of their own learning journeys. So if they see the value in doing something properly, I see that they will do so.”

“Ultimately, for me this is a question of meaning and value in giving assignments, not necessarily in policing cheating,” she says and then adds a caveat that reveals a lot about her approach. “But I also teach legal skills, not legal clauses, which is a different sport.”

Where Philip describes a culture in Aarhus where many students hide their AI use, Barbora insists that openness has to be actively designed into the course.

“To me, creating a space for safe and productive exploration is the first and most important purpose of higher education,” she says. University is the time when students can “prototype different careers and directions, and rely on the university safeguards if and when they fail at something.”

That kind of environment, she stresses, “does not just happen, it needs to be proactively sculpted.” If the space is not designed for openness, “students are not going to start trusting and sharing out of the blue.”

So she treats hospitality as part of the pedagogy: making people feel welcome, seen, and comfortable, down to “the tiniest of details, like learning the students’ names, celebrating pivots and mistakes. Including my own.”

One concrete exercise that works well is simply to ask students to use AI to create something outright bad. It deliberately tones down the pressure to be perfect, opens up room for pivots, and signals that the point is to explore and reflect, not to present flawless outputs to the professor.

For the future of law school she’s cautiously optimistic but very aware of the constraints.

“This level of cultural innovation needs to be happening at all levels so it will depend on which uni we are talking about, who are the leaders, ambassadors, and spaces in play,” she says. And the ground keeps moving: “the status quo is changing almost weekly, which makes planning challenging.”

There’s a tension to manage. “On one hand, with the pace of new tech, we need to be flexible and agile. At the same time, we are influencing people, so anything we do and teach should be grounded and informed.” That is “a lot of work and a lot to keep up with alongside your teaching and research practice, or a day job.”

For universities, that means creating “spaces for exploration and experimentation as well as reflection and sharing of best practices” and making those spaces realistic enough that ordinary teachers can actually use them.

Shaping the future

Across their differences, all three voices converge on one point: AI is already reshaping the legal profession faster than law schools are changing. Perhaps even faster than the early adopters can follow along.

Whether universities can catch up, and whether they should move closer to Dalmose’s mandatory AI assignments, Sejersen’s warnings about overreliance, or Baru’s data-driven sandbox, is still an open question.

For now, the future of AI in legal education may be decided by what students and a handful of teachers do in the margins: an association meeting after class, a prototype built as a side project, a small seminar where AI is treated not as a threat, but as homework.

The hidden edge in legal AI: Seeing what the model misses

Mon, 01 Dec 2025 11:16:36 GMT

Newsletter #24

Read to the end for some canine info

A word from the person behind the laptop

You’re a junior at a big law firm. Today’s assignment will decide whether you get staffed on the new litigation team. Your task: legal research. You open the firm’s new AI system.

Expectations are high. You’ve practiced with the tool and you’re ready. You ask for a memo on a procurement dispute.

In seconds, it returns a clean outline… and the same three leading cases everyone cites. It isn’t wrong. It just won’t win you the seat.

You need to prove you can push past the default and surface an argument, but you’re relying on AI for research.

Who do you blame? The vendor, yourself, or the firm? What do you do?

Andrej Karpathy said something on the Dwarkesh Patel podcast that I’ve been thinking about for a while: “reinforcement learning is terrible”

Modern assistants are shaped by reinforcement learning, where long chains of output get judged with thin, end-of-process feedback. That nudges models toward answers that look right, not reasoning that’s actually strong. At scale, that can narrow legal research.

What does that mean for the way AI is currently being used for legal tasks?

That’s the theme of the newsletter today.

Subscribe now

Devil is in the detail

Lawyers are sticklers… mostly.

The “Suits” stereotype is obviously far from the truth, but one thing it gets right is that really good lawyers do share a kind of obsession with details.

Finding the right court decision to change the view of a well established precedent. A footnote that leads you astray and think diffrently about a problem. And so on. And then of course tell all of your lawyer pals about it afterwards.

With the implementation of AI systems in legal research, this comes under pressure.

We’re living throught the biggest shift in legaltech to date. Period. It doesn’t matter if you don’t like or believe in AI. Every firm I talk to is either testing or fully onboarding AI tools and that itself makes it the biggest change.

We’ve gone from simply being able to access legislation digitally across jurisdicitons (already a revolution at the time) to automating fairly complex workflows. And the promise keeps grwoing.

Tbh, I’m a big believer in the potential of AI in the legal sector. As an entrepreneur building AI tools for lawyers, I have to. But we also need to be clear-eyes about the potential challenges that comes with a new way of working.

Because the way reinforcement learning in AI products work has the potential to shape how the next generation of lawyers think and work.

And not always as expected.

If junior lawyers, paralegals and students start using the same handfuld of models for the same types of tasks (which, in my opinion, they will) there’s a real risk their understanding of the law diverges dramatically from the generation before them.

The same way our relationship to legal research is different and digital compared to our parents generation manually digging through legal sources. But this shift feels bigger somehow.

Now lawyers are also outsourcing pieces of their legal research and reasoning to a model that doesn’t actually reason (controversial statement, depends who you ask).

Short note on reinforcement learning

Reinforcement learning is bascially trial and error at scale. The model tries a bunch of possible answers. Most of them are unusable. A few are fine. The ones that happen to get rewarded, either by a human reviewer or an automated scoring system, are the ones the model leans into next time.

It’s a little like training a dog: good dog, bad dog.

This is the actual the description they’re using at Mathworks:

Using reinforcement learning terminology, the goal of learning in this case is to train the dog (agent) to complete a task within an environment, which includes the surroundings of the dog as well as the trainer.
First, the trainer issues a command or cue, which the dog observes (observation). The dog then responds by taking an action.
If the action is close to the desired behavior, the trainer will likely provide a reward, such as a food treat or a toy; otherwise, no reward will be provided.
At the beginning of training, the dog will likely take more random actions like rolling over when the command given is “sit,” as it is trying to associate specific observations with actions and rewards.

Unfortunately, the model can’t tell which parts of its “successful” answer were actually good. It doesn’t isolate the valuable reasoning. It doesn’t critique itself. It simply rewards the entire path that happened to lead to the answer including the detours, the irrelevant bits, and the lucky guesses.

“Sucking supervision through a straw”

In practice, RL “sucks supervision through a straw”. A painfully accurate description by Karpathy.

Basically a long chain of output gets judged by a thin end-of-process signal that up-weights everything in the winning path. Both good steps and lucky detours.

Here’s an example: Reviewers at a procurement practice, like most laywers busy (but well-intentioned of course), tend to reward memos that cite three leading cases. The model internalizes that vibe.

When a junior asks about invoking the “extreme urgency” exception after a strike-related supply shock, the assistant produces a clean outline with those familiar citations and frames necessity narrowly.

It never surfaces an older appellate decision that analogizes a comparable disruption. The weird outlier precedent that could flip the analysis. Nothing is obviously wrong. It’s just predictably narrow because those rewarded patterns dominate. And it’s perhaps a little bit lazy too.

Humans don’t learn (or conduct legal research) like this.

If you went down three wrong paths and one right one, you wouldn’t reward all four equally. Hopefully you’ll reflect. You’ll think. Re-consider. Learn:

“Okay, that part was solid. That part was nonsense. Don’t do that again. Do more of this.”

But current LLMs have no such pause, no internal review meeting. They just upweight everything that happened to appear in the winning trajectory.

My worry is that once that rewarded pattern becomes the default assistant for thousands of juniors, we risk a generation of legal reasoning optimized for familiar vibes rather than discovery of outliers. The model got a gold medal once, so it keeps running the same lap.

That’s not how you win close races… sorry, cases.

Now, you might add that in law, consistency is justice. If AI nudges juniors toward the same settled authorities, we cut variance and avoid rookie mistakes.

But RL-nudged tools standardize on what was rewarded, not necessarily what’s right for your matter. In practice, that promotes majority readings, buries minority holdings that matter, and misses jurisdiction-specific wrinkles.

Intangible change

We’ve focused so much on hallucinations for the past few years that I’m starting to think we’re at Woodstock Festival. Hallucinations are tangible. You can, at least in theory, point at them.

This case doesn’t exist. This quote is made up. You might be too busy or too tired to catch them, but the error lives in the output. It’s checkable.

What worries me more is the thing we don’t see, because it doesn’t show up as a clearly wrong answer.

The bigger risk is the slow, intangible change in how legal research is done when LLMs, trained with the same reinforcement learning patterns, keep reaching the same kinds of conclusions in the same kinds of ways.

Not obviously false, not obviously hallucinated. Just consistently nudging lawyers toward a narrow band of authorities, arguments, and framings that happened to be rewarded during training.

Of course its not only reinforcement learning that causes homogenization. Pretraining data + instruction tuning also push models toward “safe, average” answers. Reinforcement learning then amplifies whatever humans/metrics reward, but the bottom line is the same:

The search space of legal reasoning shrinks and you rarely notice what’s missing, because the model is likely to give you something plausible.

You don’t see the argument it could have made if it had been trained differently. You don’t see the weird outlier that never gets surfaced because it wasn’t in the slice of data that was heavily rewarded. The good lawyers will still of course be able to perform proper research (especially in places where time and money is not an issue), but expectations will change on the time spent on these tasks now that AI systems can support.

But a lot of lawyers don’t have the luxury of no constraints.

So what happens in practice is that “research” slowly gets stale, narrow and more predictable over time.

So what can we do about it?

Last point from the Karpathy interview, something people in ML circles have been talking about for years:

Why don’t we train models on the process, rather than the final answer?

This idea process-based supervision has been sitting in the collective imagination of ML researchers for years. In theory, it solves the whole problem. Humans could review the chain of thought step-by-step, marking which steps were actually good reasoning and which were nonsense. The model wouldn’t have to guess which parts mattered; we’d tell it.

So why hasn’t this saved us yet?

Three reasons:

1. Humans can’t reliably evaluate long chains of machine reasoning

2. The process the model produces is not guaranteed to be the actual process

3. It’s expensive and slow

So here’s my rather average advice:

1. Treat the AI output as a baseline. Never the endpoint

Great for speed, but needs work for the final judgement. Treat it as such.

2. Keep the manual research muscle alive

Even if tools get “really good,” the ability to find the weird case, the uncomfortable exception, or the overlooked angle will always have value.

Nothing I’ve seen suggests this skill is about to become unfashionable.

3. Teach juniors how these models actually learn

When I teach AI to lawyers or law students, we spend quite a lot of time on the technical aspects. Not just “here’s a good prompt.” Also statistics, reinforcement learning, Python etc. Some hate it, and some love it.

I fundamentally believe that having a basic understanding of the tech gives you much better tools to discuss and understand what’s going on under the hood.

My advice is to teach the underlying dynamic:

RL reinforces patterns, not truth.
Confidence is not correlated with correctness.
“Legal analysis" in AI often just means “rewarded analysis.”

Use the assistant for speed. Use your judgment to widen the search. Consistency is the floor, not the ceiling.

The best lawyers will be the ones who notice what the model never shows them.

AI vs lawyers: 3 takeaways from the latest legalbenchmarks report

Thu, 02 Oct 2025 12:17:06 GMT

Hi everyone

I’ll be releasing the rest of my interviews very soon, but I didn’t want to wait to share this.

The LegalBenchmarks team just published their latest research comparing lawyers and AI tools on contract drafting. It’s one of the few places you can find a structured, reliable assessment of how these systems actually perform and the findings are worth paying attention to.

I’ve put together a short video highlighting three insights from the report:

AI is already as good as lawyers in many instances, and the “human-in-the-loop” idea is less relevant for quality than many assume.
Workflow integration is the real differentiator. Accuracy only gets you so far.
Big funding rounds don’t automatically mean better tools.

Watch the video, and if you want to go deeper, I recommend reading the full report on LegalBenchmarks.

The rest of my interviews are coming soon, but this was too important not to share.

Hallucination is not your problem. Sycophancy is.

Thu, 07 Aug 2025 10:14:17 GMT

Newsletter #22

Incoming: More stories from the trenches

A word from the person behind the laptop

As we continue my summer journey exploring the landscape of AI and legatech, I wanted to talk to someone who’s not necessarily trying to “change the world with enterprise B2B SaaS”, but actually trying to approach and assess this rapidly evolving field with a very methodological approach.

Anna Guo is one of those people. Her project legalbenchmarks is moving on the forefront of assessing the usefulness of AI models for legal professionals.

We’ll let her explain herself.

The art of benchmarking legal AI

One of the main concerns for many lawyers is the uncertain nature of large language models. It’s a complex issue: how do you still reap the benefits of AI models that genuinely understand your intent and deliver tailored, high-quality outputs.

Anna Guo is currently addressing this very issue with her community driven project Legalbenchmarks.ai, an initiative aimed at helping legal teams separate marketing hype from reality, and explore how automation and AI will reshape legal roles including what gets replaced, what gets created and what skills the next generation of lawyers will need to thrive.

Anna Guo, Founder of LegalBenchmarks.ai

According to Anna, LegalBenchmarks doesn't aim to become the "definitive" benchmark. Instead, it offers practical evaluations tailored specifically to the everyday realities faced by in-house legal teams. Real-world prompts, directly sourced from in-house lawyers, complete with typos, vague instructions, and realistic imperfections, form the basis of these evaluations.

“In addition to accuracy, we evaluate outputs on usability metrics like helpfulness, clarity, and length," Anna explains. "We also measure platform features, because usability matters just as much as accuracy in daily workflows."

Legalbenchmarks is already getting support from the legal community

Interestingly, LegalBenchmarks also compares specialized legal AI tools with general-purpose models like OpenAI, Anthropic, and Google.

If you read the last version of my newsletter, the in-house counsel Aleksi Hokkanen actually pointed out why this is important: legal professionals will have to be convinced of the additional value of buying a premium priced product if the general-purpose AI tools can achieve almost the same results.

Human-in-the-loop isn’t a silver bullet

When I talk to legal professionals around Europe and the US, lawyers who embrace AI often confidently say something along the lines of “There’s no threat to my job because there will always be a human-in-the-loop”. To some extent I feel like that catch-phrase has become a dummy for legal professionals, but with no continuation of that train of thought.

But balancing real-world legal complexity with the practical constraints of automation isn't straightforward. Simply inserting human oversight into the process doesn't automatically guarantee better outcomes.

Results from the phase 1 part of the study

As Anna explains, “We often assume that adding a human-in-the-loop will automatically improve results. But in practice, it’s not guaranteed. Especially in legal tasks, where complexity is layered with subjectivity and trade-offs, the benefit of human review depends on who the human is, how they’re involved, and what they’re reviewing.”

Navigating this balance effectively is one of LegalBenchmarks' ongoing challenges and areas of exploration. This becomes particularly fascinating because the decision to keep a human in the loop is fundamentally about risk tolerance within a business.

Risky business?

Currently, organizations remain cautious about relying solely on AI for critical legal decisions, even if technically feasible - and understandably so. But this caution doesn’t mean businesses are entirely risk-averse.

As Anna points out, “If the task involves high stakes, say, a multi-billion-dollar deal or a situation where lives are at stake, then 99% accuracy might not be sufficient. In these cases, you’re paying for certainty, judgment, and accountability. However, in many real-world legal workflows, such precision isn't always necessary. Reviewing standard NDAs, routine work orders, or boilerplate agreements often involves low stakes but high volumes. Here, a solution that's just 60–70% accurate might be perfectly acceptable, dramatically improving speed and efficiency.”

In other words, businesses must ask themselves: what level of correctness is acceptable for which task, at what cost, and with what consequences? This isn’t always formally documented or evaluated. It’s often something lawyers understand intuitively through experience. In that sense, it becomes an automation paradox: the assessment of the automation risk itself isn’t ready to be automated.

A lot of focus has been solely on hallucination as a risk for AI work in legal, but Anna points out another risk that lawyers often don’t mention.

Beyond hallucination

“An even bigger and more insidious problem is sycophancy.”

I ask her to elaborate:

“This shows up in many forms: false premise acceptance, misleading assumptions, overly confident agreement with user input. Essentially, the model echoes the user’s bias instead of challenging it. And unlike hallucination, which you can often fact-check, sycophancy is harder to spot, especially when it is hidden in the user’s blind spots.”

This point resonates with me because this is something I’ve experienced myself first hand, but also because a lot of the legal training is based on drafting legal text and having it torn apart by your seniors.

But in maybe one out of ten cases (being nice), you’re actually right, even as a junior. Not because you’re being agreeable, but because your “legal taste” kicked in. An instinct not based on sycophancy, but precisely the ability to challenge authority on their legal assessment when it matters.

Legal “taste”

The term “taste” in a legal context is actually something Anna comes up with. While discussing the challenge of balancing scalability with human judgment, she explains what lawyers uniquely bring to the table:

“On the one hand, legal evaluation depends heavily on human expertise. Lawyers bring context, nuance, and what I call “taste”, a sense of what good looks like in legal writing. But that kind of judgment is incredibly resource-intensive to scale, especially if you want to evaluate a wide range of tasks and tools.”

Legalbenchmarks release the numbers as part of their reporting

Opening Pandoras box Anna addresses the obvious follow up question: using AI to evaluate AI generated outputs. This is of course happening already, but not on a scale that is common in legal businesses. Because in the end who judges the judge?

“I’ve documented this whole journey of building and calibrating “LLM-as-a-Judge” in my newsletter Working Draft. The goal was to be transparent about the trade-offs: where automation helps, where human judgment is irreplaceable, and how difficult it is to replicate legal nuance at scale. It’s still an ongoing experiment. “

What’s next

As we finish the interview I ask Anna about her future plans for legalbenchmarks. The immediate plan is to publish the phase 2 report on contract drafting (insert link) which builds upon the learnings of phase 1.

But as a true community builder and founder Anna's main goal is different - and bigger:

“The core mission is to provide real value to the legal community”

The road leading to there is not yet written in stone, but legalbenchmarks are collecting feedback from the legal community as we speak to learn and improve their methods. Today they have more than 400 lawyers and AI experts contributing across different jurisdictions and practice areas.

So while she doesn't know exactly what legalbenchmarks.ai will become, she knows that it will be evolving with the legal community surrounding it.

She ends our interview with a call to action:

“If you’re curious about where legal AI is headed, whether you're a lawyer, a technologist, or somewhere in between, we’d love for you to join us!”

Feeling the heat? Time to take the temperature of AI in legal

Thu, 17 Jul 2025 08:10:28 GMT

Newsletter #21

Incoming: Stories from the trenches

A word from the person behind the laptop

Welcome to my summer series!

I decided to spend some of my time this summer talking to people who actually operate in the AI legaltech space, and have experience getting their hands dirty in one way or another.

Meet Jerrit Jungman, Aleksi Hokkanen and Kristian Anker that I interviewed for the first chapter in this series

I've spoken to many interesting people with different perspectives that can hopefully give us a glimpse of what is coming in the AI legaltech space.

This first newsletter is about the different perspectives from builders, investors, and practitioners. How much do they have in common and where do they differ?

Hang tight!

Part 1: Builders

Building trust with Kristian Anker, Luma

My journey begins in the startup hub of Matrikel1 in central Copenhagen, home to the innovative legaltech startup Luma. Having followed their journey closely, I believed they were the perfect entry point for a deep dive into the AI legaltech landscape.

Kristian Anker and Oscar Almstrup, the founders of Luma, are childhood friends who've already experienced the thrill of building and selling a successful company together. Sitting down with Kristian, I wanted to understand their unique perspective on AI and legaltech.

Startup lawyer

Kristian's story is unconventional. Originally on track to become a lawyer, he took a detour, joining the army as an officer for a couple of years. Returning to his studies, he discovered entrepreneurship almost by accident, inspired during a visit to Oscar, who was studying in San Francisco.

They quickly built a prototype that took them to a Hong Kong conference, validating their idea and eventually evolving it into LegalHero, a platform connecting gig-based lawyers to clients. After a successful exit to Lexly, Kristian became the CTPO of the acquiring company, before founding Luma with Oscar once again.

When asked what sets Luma apart, Kristian explains: “Our angle is tying together the entire back office, legal, HR, and Finance, from a legal ops perspective. Most companies prefer a finance-centric view, but we've flipped that, simplifying and commoditizing legal operations specifically for SMEs.”

He recalls the exact moment he realised something had changed: “Seeing the first versions of the new LLMs felt like a turning point. Suddenly, we could enable non-legal professionals to manage legal operations intuitively, without needing endless workflows and manual processes. It was as if someone had invented a completely new joystick.”

Having built LegalHero before the GPT era and now Luma post-GPT, Kristian identifies pace and accessibility as the biggest tech shift. "Past technological advances typically benefited large corporations first.

But with ChatGPT, AI went directly to consumers, instantly accessible through the simplest of interfaces: A chat. This consumer-first approach has dramatically accelerated adoption and development speeds."

A hot space

Kristian believes the AI legaltech space has never been hotter: "Earlier AI solutions felt academic, with limited real-world impact. Now, we're seeing applications that genuinely transform workflows.

Companies like Harvey and Legora show the rapid adoption and significant investment pouring into the field, but the real test will be seeing their staying power over time."

When pressed about what makes AI different from past expert systems, Kristian emphasizes its timing and suitability for legal work: “AI models today are text-driven and deeply aligned with legal tasks.

Combined with the rapid growth of AI agents, it's an ideal match. Trust remains central, but AI's direct-to-consumer approach bypassed typical adoption hurdles, making lawyers quicker to adopt and integrate these tools."

Trust issues

However, the challenge of trust remains the hardest objection Kristian encounters: “How can I trust it?”. It’s the one question every founder in this space has to have a good answer for. Not for sales, but because it’s the only question that matters.

As for the trend he's betting on, Kristian doesn't hesitate: “Delivering AI-powered legal services directly to customers rather than just selling systems to lawyers. With proper guardrails and human oversight, AI can replace significant portions of routine as well as core legal work for SMEs"

He concludes confidently: "AI will never replace a lawyer who can sell trust better than AI can, because, ultimately, AI will become just another software tool."

Part 2: The Investors

Inside the VC perspective with Jerrit Jungman

I continue my journey by turning to venture capitalists, who often provide essential fuel for innovation.

Eager to understand the enthusiasm behind recent valuations of promising legaltech startups like Legora and Harvey, I reached out to Jerrit Jungman from People Ventures to gain insight into what investors find compelling in the AI legaltech space.

Jerrit's story is straightforward: growing up in Germany, studying business, and getting hands-on startup experience before crossing over to the investment side. For the past three years, he's been part of People Ventures, a Nordic venture capital firm.

“We support founders across Northern Europe, from idea stage to Series A, typically investing up to €2.5 million,” Jerrit explains. “Legal interests us because it’s filled with broken processes, perfectly positioned as prime use-cases for language models.”

Digitized Scandi landscape

On the hype level in Scandinavia, Jerrit is clear: "We are at the forefront of what’s happening. Multiple companies that are doing breakout things. I believe we have a natural advantage up here, given the superior digitalization compared to other countries"

When discussing valuations like those of Harvey and Legora, Jerrit acknowledges the hype but stresses there's substance behind it: "Without direct access to internal data, it's tough to judge, but both companies must have best-in-class growth rates. Harvey, for example, grew monthly active users (MAU) by 240% year-over-year, with over 70% MAU among paid seats. The potential of this space is definitely also priced in."

However, Jerrit doesn't ignore practical challenges, noting churn as a significant issue: "Churn in legaltech closely ties to change management. Lawyers don't have much time for new systems, so onboarding and training are critical. Both startups and buyers must have clear plans for user education."

Investment focus

I address the fundamental paradox that the legal industry is built on trust and LLMs have a potential trust challenge. He draws a parallel with medicine: "Every transformative technology starts imperfectly. AI in medicine already surpasses human accuracy in diagnostics. Law will inevitably follow, making it potentially unethical not to adopt AI when outcomes demonstrably improve.“

He continues “Our investment thesis revolves around data: high-quality, domain-specific data is key to reducing errors and building reliable AI."

Finally, I prompt Jerrit to finish the thought, “AI will never replace a lawyer who can____"

“…set aside their ego and embrace technological change. You’re not competing with AI, you’re competing with other lawyers and firms who are using it effectively.”

He elaborates “This isn’t unique to the legal profession; it’s true in my field and many others. I see it as a real opportunity to leap ahead of the competition and make a name for yourself by leveraging the best tools available early.”

Part 3: The Practitioner

A realists perspective from Aleksi Hokkanen at Plugsurfing

To find out where we are in the hype cycle I also wanted to talk to someone who works “where the rubber hits the road”.

I reached out to Aleksi Hokkanen, a legal counsel from Plugsurfing in Helsinki, whose thoughtful insights about integrating AI into the legal function in companies caught my attention. Some of his perspectives intrigued me, and I was eager to learn more about his experience with potential vendors, solutions, and his own workflows.

“It’s a bubble. It’s like internet companies in 1999.”

I catch him on the phone in his car while he’s charging somewhere along the highways in Finland.

“Listen, the market is too crowded, it’s overhyped and there’s so much buzz especially from VCs, and I’m actually very positive about the implementation of AI as a tool in the legal toolbox”

The ideal customer is still sceptic

Plugsurfing is a tech-driven company connecting drivers, fleets, and carmakers to one of Europe’s largest EV charging networks.

He’s part of a legal team operating across Europe in the EV charging and software B2B spaces. This is in other words the kind of tech oriented legal team that could be first movers in the AI space.

Aleksi’s pragmatic about AI both its potential and its pitfalls. When asked about measuring return on investment for AI solutions, he’s clear-eyed: “Right now, it’s complicated. Most senior management judges legal efficiency purely through cost. They’re comfortable with the idea that legal teams are overloaded. So, even if AI boosts efficiency, that improvement isn’t always clearly visible on balance sheets.”

Hokkanen explains that while many vendors promise significant enhancements, real-world implementation often hits snags. Plus budget-holders in companies might belong to an older generation who prefer familiar Microsoft tools and are cautious about new technologies. “People are still wrapping their heads around what AI means in practical terms. You’re often met with “Just use co-pilot”” Hokkanen says.

He continues “On the other hand you can’t just plug AI in and expect magic. It’s a tool, and you have to invest time and money ensuring models don’t hallucinate. Trust requires human oversight.”

He’s optimistic about AI’s transformative potential but stresses it won’t replace core legal skills. “AI will never replace a lawyer who can ask thoughtful questions, apply judgment, and truly understand the human context. But lawyers who learn to leverage AI will outperform those who don’t.”

Expensive ChatGPT wrappers?

When it comes to choosing specific AI solutions, Hokkanen mentions platforms like Legora, the Swedish legaltech company, though he finds the pricing models challenging. “Why pay for something that expensive when ChatGPT is already delivering similar outputs? It really has to integrate well with existing workflows to be worth the price tag”.

Aleksi highlights that AI adoption is in its infancy: “Most companies are barely scratching the surface. AI has only been mainstream for a few months. Right now, it’s a nice-to-have, not yet essential.”

He also addresses challenges in integration, noting a disconnect between the sales pitches of AI vendors and actual user experience: “There’s a gap between what’s being sold and what’s truly practical. Contracts can’t just be handed to an AI model and returned ready-to-sign. The same goes for other legal tasks. You still need careful human review.”

He sees current excitement around AI, but his bubble thesis comes from tendencies similar to the dot-com boom and bust: “Most of these legal-tech AI startups won’t survive. But the ones that do, those will genuinely improve how we work. Like internet companies after 1999”

As our call wraps up, he returns to his car charging somewhere in Finland, blending his legal expertise with a dose of realism about the road ahead for AI in legal tech.

Navigating the road ahead

Where does that leave us?

Talking to people with actual experience in this field leaves me with this impression: We can all agree that the train is moving fast right now, but it’s hard to say which destination we’re headed towards.

Is it a looming revolution, just another set in the toolbox or will LLMs ever overcome the fundamental trust challenge? I’m generally optimistic about the future of AI in law, but based on my own experience, there’s still a noticeable gap between the practical realities of legal workflows and how LLMs are implemented today.

With a builder’s mindset, strategic VC support, and careful attention to the trust challenge, we might just bridge that gap sooner than we think.

I’ll continue my summer journey to try and find the answer…

Stay tuned!

Vibecoding your way out of expensive legaltech

Thu, 22 May 2025 06:00:56 GMT

Newsletter #20

Hang around for some info on how to save some dolla dolla billz

A word from the person behind the laptop

Around three months ago legend Andrej Karpathy posted a tweet about “vibecoding”.

I had seen the word used before but this is definitely the breakthrough moment for the term which seems to have turned into a full-blown movement.

It seems like people didn’t really have the words to express what they we’re doing when they we’re building software using only ChatGPT or other solutions like Replit, Cursor or Loveable.

I’ve been an amateur developer for years: building half-functional apps in Python, tinkering with Swift, and generally wasting weekends on side projects that mostly didn’t work very well. But now? I literally vibecoded a Swift application with one prompt that took me 3-months to built back in 2015.

Replit CEO Amjad Masad says 75 % of its users never write a single line of code and Gartner assumes that by 2026, around three-quarters of all new applications are built low-code technologies.

How do you feel about CJEU vibedrafting?

This obviously raises the question: If company X is paying vendor Y mid-six figures a year for a SaaS licence, does that still beat rolling out own internal tools if they’re that easy to build?

Unfortunately, the math isn’t that straightforward.

In this post we’ll dig into why. Hang tight!

“Sorry, we’re building this internally”

Back when my own start‑up was still pitching its very first pilots, we kept hitting the same polite brick wall:

“Love the demo, but why wouldn’t we just use ChatGPT?”

and/or

“Love the demo, but we’ll just spin up something ourselves.”

Vibecoding your way to save time by not dealing with external vendors

Things are moving fairly quickly at the moment and companies are probably just figuring out where AI actually adds value. In that context, onboarding a new shiny AI solution might seem like a stretch (even though our solution is of course the best in the market).

We work in a sector where people can be famously risk-averse (for good reasons). So building something internally with people that understand the internal processes, systems and data might seem like a safe bet on the surface. Especially now that you can build internal software over your lunch break accompaigned by a chatbot.

But here’s where things get messier.

Just because you can build an app internally doesn’t mean you can support it. And if your legal department doesn’t have budget for maintenance, evaluations, security, or any of the other “boring” siblings of vibecoding, then you're not building software. You’re stacking technical debt.

According to Stripe’s Developer Coefficient study:

(..)the average developer spends more than 17 hours a week dealing with maintenance issues, such as debugging and refactoring. In addition, they spend approximately four hours a week on “bad code”, which equates to nearly $85 billion worldwide in opportunity cost lost annually(..)

Truth is that many vibecoders don’t even know what maintenance and refactoring is. That’s 42% of the workweek not spent on shiny new features, but on keeping the lights on.

Stats from Stripe’s Developer Coefficient

If legal departments are already struggling to squeeze value from tools they already bought then getting the budget to vibecode your departments way into the future might even be even more far fetched.

“Good enough” doesn’t scale

Sometimes, the DIY approach makes perfect sense.

I can install a new kitchen using nothing but YouTube tutorials. It won’t look like something out of Architectural Digest, but since it’s just my wife and I looking at it 98% of the time, we can probably live with a crooked drawer for a while. At least until we get around to fixing it.

But that doesn’t mean I’m qualified to build an industrial kitchen for a high-end restaurant in Copenhagen. If I did, it might affect more than just aesthetics. It could impact the quality of the food, the restaurant’s business, and possibly Copenhagen’s culinary reputation (RIP Noma).

Same goes for software.

I can easily build a simple “rental contract checker” using GPT and a bit of logic. It works well enough as a first pass. I’ll still review it, of course, but as a personal productivity tool? Totally fine.

But that doesn’t mean I should roll it out across an enterprise, where hundreds of rental agreements (and maybe entire deal structures) depend on the standards baked into the equivalent of a weekend project. Especially not if I, as a lawyer, have no plans (or budget) to maintain the tool once it’s in production.

Because at that point, we’re no longer talking about a helpful side project. We’re talking about infrastructure. And "good enough" doesn’t scale when real risk is involved.

Imagine if someone wrote contracts like that. Your job as a lawyer would be insufferable, looking through other peoples bullshit contracts. This is probably how many developers feel at the moment.

…and it might not even be worth it

Let’s say you did it. You vibecoded a working prototype. It parses contracts, updates your matter tracker, even pretends to understand your internal folder structure. The team loves it. Legalops is impressed. You’re a genius - and you might even feel like a real developer!

But unless you plan to freeze time at the prototype stage, things get expensive very quickly. So let’s talk a bit about the thing we’re all thinking about: the cost

The fantasy is that building internally is cheaper. “Look how quickly we deployed this tool!” Unfortunately, the fantasy ends somewhere between your third code refactor and your first security audit.

I made a back-of-the-envelope calculation to show that the numbers are actually fairly easy to put together (I’m not a math guy btw, but I did my best):

Danish developer salary would make the calculation even worse for the internal build projection so I used Germany as an example. Still doesn’t hold water though.

This is only year one, but break-even with SaaS might not until year 3 or 4. And that’s assuming nothing breaks, no regulations change (EU, hahahaha), and no developer rage-quits because they were on call during Christmas or people in the department keep talking about how useless developers are now that we can vibecode.

Meanwhile, vendors amortise the boring parts. Ironclad’s 2023 Forrester TEI study shows a 314% ROI over three years after factoring in licence fees, onboarding, and support. I haven’t double checked these numbers, but if that’s the math your DIY project is up against then… good luck.

Then what can we use vibecoding for? It’s magic for proof-of-concepts and throwaway workflows. It’s fun, fast, and often useful. But once you’re maintaining real infrastructure, dealing with 42% of your week spent on upkeep, plus the 60% lifecycle cost that hides in patches, bugs, and mystery downtime the economics flip. So the initial idea of keeping it in-house to save money might fall short.

Unless you're building something that's actual firm IP, something you’d never hand to a vendor on principle, cut the RFP short, pay the invoice, and let someone else answer the 3 a.m. pager.

Start-up of the week(ish)

I’ve got a lot of faith in voice assistants for legal. Not just because I’ve seen Boomer lawyers still rocking dictaphones, but because speaking out loud is actually a solid way to organise your thoughts.

LexisNexis kicked off Legalweek by launching Protégé’s new voice AI assistant in the U.S. Allegedly the first in legal? Not sure about this - but now instead of typing prompts, you can just talk to your research assistant.

Extra toppings

The goalposts for what AGI is are shifting from week to week. This is only 3 years ago:

Why techno optimists belong in Berlin

Thu, 24 Apr 2025 08:30:19 GMT

Newsletter #19

Hang around for some Eurocrat propaganda

A word from the person behind the laptop

While it might seem like the world is falling apart around our ears atm, I wanted to take a minute to talk about a more (dare I say) optimistic outcome of the current situation.

Wake-up call number 1.000 for the European Union seems to be the equivalent of iPhones Radar alarm tune that finally wakes us up from the eternal morning snooze.

While “we” (talking for all lawyers here) are nurtured by the the eco-system of laws, rules and stickler-like behaviour there might be a limit to the amount of regulation a union of 27 nation states can stomach - and how scared you can legitimately (pun intended) be of AI at this stage.

In this newsletter, I’ll explain why the future of AI belongs to EU.

Enjoy!

Apparently investors don’t thrive in the jungle?

It’s only been a few years since the EU rebranded itself as the ultimate tech regulator.

The governmental equivalent of Berghain’s legendary bouncer Sven, deciding which tech giants get in, who needs to change outfits, and who’s not welcome at all.

The impersonation of “the Brussels effect”

With (mostly silent) support from the US, Brussels rolled out legislation like the AI Act, Digital Services Act and Digital Platforms Act - and many, many more pieces of legislation that haven’t received the same level of attention.

All meant to curb untamed development of dangerous tech from both Shenzhen and Silicon Valley. The US innovates, China imitates, and the EU… yeah, you know the rest.

Anyway, I recently joined the OpenAI Forum and took part in a discussion with economist David Autor - and I have to say I agree with his view on the progress of AI deployment.

It’s a slow, but steady process where humans and organisations will have to adapt step-by-step and where the long-term gains are likely to land with the middle class, not just the top 1%.

In my view, this kind of gradual, structural adoption actually plays to the EU’s strengths. Especially if we can swap the current anxiety about AI for a bit of strategic curiosity.

Because right now, the EU has a real chance to become a safe zone for capital. If the US continues to spiral into regulatory whiplash and political volatility every 72 hours, then investors who thrive on predictability and rule of law, not chaos, will start looking for alternatives.

Don’t get me wrong. I admire the US for its ability to reinvent itself and investments will of course continue, but I believe that the underlying volatility (economically and legally) could make the EU more attractive on at least three fronts:

a. Talent may increasingly prefer to stay in - or move to - the EU

b. make it more attractive for investors to invest in EU based companies, and

c. the EU will (finally) be forced to build its own infrastructure. Not just to support AI, but to underpin a broader digital ecosystem (not relying on unreliable partnerships).

If this sounds like a fever dream then just gimme a sec to explain why things are already pointing in this direction:

a. The Niels Bohr effect - but for tech

Being a proud Dane I have to point towards the example of physicist Niels Bohr. Yes, Bohr was a brilliant physicist. But what’s less known is that he had an additional talent which was… attracting other talent. He basically turned Copenhagen into a global centre of quantum research in the 20th century.

Today Silicon Valley has basically vacuumed up European tech talent with a combination of cutting-edge work, cult-like optimism, and compensation packages that redefine the word “income.”

But luckily it’s not just about the paycheck. Talented people want to work on problems that matter. They want to be around other talented people. They want momentum. Silicon Valley knows this and constantly comes back to this idea (for good reasons).

Perhaps Niels Bohr laid out an example for the EU to follow

If you’re at the forefront of software innovation in the EU, UK, Australia etc., you’re most likely in California or trying to get there. I’ll leave out hardware, where the EU actually has a stronger foothold than what it gets credit for (just look at ASML, the Dutch company that makes the most advanced chipmaking equipment in the world).

The EU has known for years that it has a talent problem in this type of tech. And it has tried (mostly half-heartedly in my opinion) to fix it. But the problem with talent is that it accumulates exponentially. Once one tech cluster gains critical mass, the rest tend to get hollowed out. So reversing that trend is hard.

Still, there’s reason to be (cautiously) optimistic. The AI Innovation Package, published in January 2024, includes a commitment to international recruitment of AI talent, and the rollout of targeted education and training programs in priority sectors.

The real test will be whether Europe can build gravitational centres of AI innovation.

Right now, Mistral (France) is probably the one to watch. Yes, it’s mostly funded by US venture capital (lol), but it’s headquartered in France and building open-source models that punch well above the average EU weight.

The Stanford University index actually has 25 of 101 EU-based “notable AI models” on their list - but I guess the real question is how many of these you can mention? Whether companies like Mistral can attract global talent will be a key signal of whether Europe can hold its own.

That said, if you listened to the Hard Fork podcast’s report from the Paris AI conference, it wasn’t exactly a confidence boost. The vibe was… let’s call it “early days.” Not the kind of thing that makes a Stanford student move to Sorbonne. But I respect that these things take time - and frankly, France is doing more than most.

Plus Europe ticks the box in for a very important factor: Affordable, world-class technical education + better food. If the EU can combine that with regulatory clarity, ambitious tech hubs, and just a bit of vision, then yes, it could start reversing the talent drain.

b. Brussels is more predictable than Palo Alto (and that might be a good thing these days)

I highly recommend Nate Silver’s new book, On The Edge, which takes a deep dive into the psychology of risk. One of its most compelling themes is the risk culture of Silicon Valley, which, for better or worse, has become foundational to how tech companies operate (and occasionally implode).

That same risk culture has shaped everything from the way startups raise capital to how the Valley engages with politics. Case in point: Silicon Valley’s increasingly warm embrace of Donald Trump, who some (Marc Andreessen among others) see as a deregulation goldmine.

Maybe that gamble pays off. Maybe it doesn’t. But if you’re an investor looking for long-term clarity, betting on political chaos isn’t exactly a sound strategy.

I’ve been following Silver’s tech stock tracker. An overview of how major Silicon Valley-linked companies are reacting to market shifts and election-year uncertainty. So far, the signs are… mixed.

According to Nate Silver’s Silver Bulletin newsletter, 37 of the 50 companies in the SV50 (the 50 largest companies in Silicon Valley) have declined more than the Dow Jones Industrial Average since Inauguration Day, and two-thirds have fallen more than the S&P 500.

Meanwhile, back in Brussels there’s something investors are starting to appreciate more and more: predictability (or at least MORE predictability than the US)·

The EU might not move fast. But it moves in a straight line. It offers regulatory certainty, rule of law, and a coherent legislative process that doesn’t depend on whoever’s trending on Twitter that week.

Even some of the EU’s big fines - Meta, Apple, take your pick - look less arbitrary in this light. Yes, Mark Zuckerberg recently told Joe Rogan that EU fines are “similar to tariffs,” which hurts my legal pride.

To be clear: A tariff is a cross-border trade barrier/tax. A regulatory fine is the price of failing to follow laws that (in theory) apply to everyone equally. The issue isn’t protectionism: it’s clarity. And clarity seems to finally be improving.

If Europe can stay the course pairing long-term investment with legal transparency it may become the Western Hemisphere safe bet for global capital looking for exposure to AI and digital infrastructure.

In other words: Brussels might be boring. But right now, boring is bullish.

c. Infrastructure is not a joke

Since the Russian invasion of Ukraine, infrastructure has been on everyone’s lips in Europe. All of a sudden, it clicked: the entire continent’s energy system was built on the assumption that geopolitical stability was a given and that a gas pipeline from an increasingly authoritarian regime was a perfectly reasonable way to power your economy…

Nordstream 2 has become a strong symbol of… failed infrastructure

And while (fairly annoying) EU tech compliance people have been yelling into the void for years about our total reliance on US tech infrastructure (cloud, chips, satellite comms, even productivity software) it’s only now, in 2025, that the full consequences of that overdependence are hitting home.

What happens when the US enters a full-blown trade war with China? Something that seems to already be happening in front of our eyes. Will the EU be forced to “choose sides”? What happens when a single US-based cloud provider or AI service changes its API terms, pricing, or access policies overnight?

We might be close to a future where AI services as integrated into our societies as cloud services are today. What happens when entire sectors of your economy transport, health, education can’t function without infrastructure decisions made in Palo Alto or Seattle?

Germany might have been right in printing and faxing documents all along.

According to the Draghi report on competitiveness, the EU must now focus on three key infrastructure pillars:

1. Compute capacity,

2. Trusted cloud infrastructure, and

3. Telecoms and satellite networks.

Satellite comms is increasingly interesting and may prove to be a powerful tool. The current political situation makes EU push even harder for an alternative to Starlink. The French are (yet again) offering the best alternative at the moment. No one in Europe wants wartime communication to depend on the current mood of Elon Musk tbh.

Infrastructure is the foundation for everything else. And will be increasingly important with the deployment of AI… everywhere.

If the EU actually wants to shape the future of AI, it needs to stop renting space in someone else’s basement and start building its own house. Now!

In other news…

OpenAI launches new “agentic” models

OpenAI released o3 and o4-mini, its most capable models yet. They can now independently use all tools in ChatGPT (code, web, file analysis, vision, image generation) to tackle complex, multi-step tasks.

Nvidia hit by $5.5B China export blow

Nvidia shares dropped 7% after the US restricted sales of its H20 AI chip to China. The chip now requires a federal license “indefinitely,” as tech decoupling escalates.

IBM debuts smarter, smaller LLM

IBM unveiled Granite 3.3, its first official speech-to-text model, supporting eight languages. It includes hallucination-mitigation tools and fill-in-the-middle capabilities. Especially useful for code generation.

Start-up of the week

This is not strictly a legaltech tool, but its definitely useful for lawyers. I’ve been using TabTabTab it quite a lot for the last week and I think there are many good use cases. Not gonna try to explain, but just try it out. It takes copy/paste to the next level.

Extra toppings

Nintendo might decide who wins the midterms

2025: Year of the Agents

Thu, 06 Feb 2025 09:31:09 GMT

Newsletter #18

Hang around for some info on how to keep your job.

A word from the person behind the laptop

We’re back!

Its been a while since the last newsletter and not much have happened since… right? Except for the U.S. casually threatening to invade its own allies, AI models quietly reaching PhD-level intelligence, and the occasional asteroid hurtling toward Earth.

Anyway, work has been hectic, so this issue took a little longer than planned. But I finally had time to sit down and sort through my thoughts on the latest AI developments in the legal sector. The big buzzword is agents so that’s what we’ll look into.

Hang tight!

Agents, agents, agents!

Let’s just start with a quick intro to one of the best agent tools on the market right now. Replit has released a coding agent and I’ve been using up all my credits for the last many months creating demos of everything from ongoing liv stock market opportunity tracker to an Pokedex for Aussie animals.

This video shows what its all about:

But what are agents actually? While we’re still figuring out the real business case for traditional AI tools, a this new trend has already taken over - one that moves beyond simple prompt-response interactions.

So let’s first make sure we’re on the same page:

• ChatGPT interface (and its siblings Claude, Gemini, etc.): Reactive AI. You give it a prompt, it responds. End of story.

• Agents: Autonomous AI. They set goals, execute tasks without constant human input, and adapt based on feedback. Less “answering questions,” more “figuring things out and acting on them.”

Now, its easy to see what the business opportunity is here. AI solving entire workflows within a business then the big workforce replacement is truly coming! What techno-optimists (not the kind you meet in Berghain) dream of is something like this:

An AI agent that constantly runs in the background, proactively managing legal risk and streamlining workflows.

It monitors contracts and regulations, flagging potential compliance issues before they become problems. It automates routine tasks like drafting NDAs, reviewing vendor agreements, and tracking key deadlines. It keeps an eye on internal communications, alerting the legal team when an executive casually suggests something that might be non-compliant.

When a complex issue arises, the agent doesn’t just provide research - it synthesises case law, drafts a legal memo, and suggests a strategy, saving hours of work. It also remembers past cases and company policies, ensuring consistency across decisions and quality assurance.

Sounds promising (and maybe a little terrifying), but reality will look very different. Having worked in large corporations and law firms, I see plenty of obstacles to this futuristic vision even though I agree that agents have the potential to be a really valuable tool.

Let’s go through 3 of them here:

1. Unwritten rules

As a lawyer, it’s counterintuitive to think that some rules aren’t written down. But in many cases, that might actually be for the best.

Just not for agents. To be useful, AI agents need clear, structured rules to operate within - something that’s notoriously difficult in large organizations where so much runs on unwritten norms and tacit knowledge.

Lawyers know this struggle all too well. Sure, knowing the law is expected, but real effectiveness comes from mastering the unspoken rules - the politics, the billable hour dance, the subtle (but crucial) decision to cc the right email to the right partner at the right time so it doesn’t disappear into the abyss.

And here’s the problem: these rules are hard to define, let alone codify. Make them too broad, and the agent might start interpreting tasks on its own, leading to spectacularly bad decisions. Make them too strict, and it becomes useless, stuck in a rigid box with no flexibility. Striking the right balance is tough - especially when the “rules” aren’t written down to begin with.

In the latest newsletter from AI policy perspective Seb Krier, who works on the Public Policy Team at Google DeepMind, describes it as “grey knowledge”:

“Employees in most commercial organisations hold critical “grey” or institutional knowledge. This includes insights gained from informal sources, like podcasts, or a nuanced understanding of workplace politics. For instance, knowing how to navigate internal politics at work or recognizing which tasks are worth prioritizing under shifting external circumstances (e.g. political changes) is rarely written down. You won’t find an internal wiki entry that says “Avoid asking this person about X because they’re biased against it” or “The new minister hates automated vehicles, so highlight healthcare topics at the next event.” In human-agent workflows, this kind of contextual knowledge and knowhow gives humans a certain advantage, and presents a significant challenge to overcome before organizations can transition to agent-only companies.”

He also points out that this kind of tacit knowledge can become a form of job security, with employees - consciously or not - withholding or manipulating information to maintain their value. And he’s right. No one has to actively sabotage an AI agent for it to struggle - sometimes, the mere fact that key insights aren’t explicitly documented is enough to keep automation at bay.

That said, he believes agents will eventually overcome this challenge:

“This challenge is solvable - agents could infer and learn quirks over time if management provides enough access to contextual data. For instance, an agent might eventually learn, ‘This is the quirk to remember when submitting a finance request.’ However, this process will be slow and uneven, especially for roles requiring physical or social interactions.”

And that’s the key takeaway: it’s not a question of if, but when. AI agents can adapt to unwritten rules, but learning institutional quirks takes time - and in many cases, longer than businesses are willing to wait.

2. The human bottleneck

This brings me to my next point: the ever-annoying “human in the loop.” Oh, if only AI could just handle everything - no approvals, no oversight, no fragile human egos slowing things down. Wouldn’t life be easier?

As Tyler Cowen has pointed out that as AI advances, the real constraint on productivity isn’t always the technology itself - sometimes it’s humans. I recommend listening to some of Tylers points in this episode of the Dwarkesh Podcast where he also talks about why he writes books for AI to read(!?).

The point he’s making is that AI can analyze data, generate reports, and even make recommendations in seconds, but if those outputs require a human decision at every step, the speed advantage disappears. The AI moves at lightning speed; the human takes a week to “review and revert.” A system artificially bottlenecked by our limited attention spans, risk aversion, and deeply ingrained need to schedule follow-up meetings.

This is especially true in legal and corporate settings, where human approval processes aren’t just slow - they’re structurally necessary. Regulations demand human oversight. Liability concerns mean no one wants to be the person who let the AI sign off on a multimillion-dollar contract without review.

The question is how useful AI is if its constantly limited by so-called human bottlenecks. The fact is that the more we keep AI on a short leash (even for good reasons), the less useful it becomes.

If every AI-generated contract, memo, or decision requires human approval, we don’t get true efficiency gains - we just create slightly more polished drafts for the same slow-moving process. At some point, businesses will have to decide how to handle this.

For low-risk, repetitive tasks, keeping humans in the loop might be more costly than beneficial, making automation a no-brainer. But for high-stakes decisions - where a mistake could mean regulatory penalties or major liability - companies may decide that slower, human-driven workflows are still the safer bet.

At some point, though, the cost of hesitation will outweigh the risk of automation. Businesses that move too cautiously will watch their more agile competitors reap the productivity gains first. The challenge isn’t just removing the human bottleneck - it’s knowing when it’s too expensive to keep it in place.

3. Infrastructure

Right now, plenty of companies are selling AI-powered workflow automation, promising seamless integration with Gmail, Slack, and other APIs. There’s a big challenge for these companies in my opinion (at least when it comes to enterprise infrastructure).

Most of these solutions take a shallow approach to enterprise infrastructure. Large organizations don’t just run on neatly packaged SaaS tools - they rely on a tangled mess of in-house software, cloud platforms (Azure, GCP), legacy databases, and, inevitably, that one critical document saved on some guy’s laptop.

This. Takes. Time.

There’s no quick fix. You can’t just bolt an agent onto a system and expect it to work - you need to rethink entire workflows. What needs to be automated? Where does the agent pull context from? How do you keep the dataflow up to date? How do different systems communicate? The setup process itself can be so time-consuming that it only makes sense if the long-term efficiency gains outweigh the upfront investment.

The Nebraska meme is old, but still holds water

But infrastructure is just one piece of the puzzle. AI agents don’t just assist; they observe and learn. A well-integrated legal agent, for example, wouldn’t just respond to prompts - it would track case law updates, anticipate relevant filings, and draft memos in a lawyer’s preferred style. Over time, it would become an extension of their thinking, catching mistakes before they happen and automating repetitive legal work. Again the data input would have to be consistent and updated to provide the right output.

I do think there’s a real opportunity here. A startup that targets a single, well-defined process - one that’s common across many companies - and builds a scalable agent workflow could carve out a strong niche. The key would be standardization: using the same APIs, tapping into the same legal sources, and integrating seamlessly with existing enterprise software. The result wouldn’t be a fully autonomous AI employee, but a highly effective, limited-scope agent that actually works.

Of course, this would still require significant effort - both in development and in navigating enterprise adoption. And pricing? That’s another challenge entirely. But if AI agents are going to move beyond flashy demos and into real business impact, this kind of focused, infrastructure-aware development is where progress is likely to happen first.

Where does that leave us?

As a person who has been working with AI adoption in the legal sector for at last many years I can say with a high degree of certainty that the current predictions are right, but the timeline is wrong.

Most companies will figure out how to standardize workflows, optimize data flows, and strategically remove human roadblocks will unlock serious productivity gains, but it will take longer because of the 3 reasons mentioned above (and several other tbh). And the first big winners probably won’t be legacy corporations trying to retrofit agents into outdated systems but leaner, AI-first startups that build around automation from day one - probably around one single process.

So, while the big workforce replacement isn’t happening overnight, the foundations are being laid. And if history is any guide, by the time most organizations realize what’s coming, it’ll already be here.

In other news…

California wants AI chatbot safeguards

California’s SB 243 would require chatbot developers to prevent “addictive engagement,” remind kids that AI isn’t human, warn parents about risks, and report on chatbot use linked to suicidal ideation. The bill follows lawsuits alleging AI chatbots contributed to teen suicides.

AI-Powered Reports for $200/Month

OpenAI’s new Deep Research tool generates research reports by analyzing hundreds of sources. It’s not real-time chat - reports take 5–30 minutes to compile, with citations.

While some praise its potential, critics warn of hallucinations, misinformation risks, and academic spam. Also missing from OpenAI’s promo? The huge energy cost - each report would obviously consume far more power than a typical ChatGPT query.

Italy Bans DeepSeek

Another Deep newstory. Italy has banned DeepSeek, OpenAI’s latest research-focused model, citing privacy and security concerns. Given its past ChatGPT ban, this isn’t shocking. Expect more fragmented AI regulation across Europe, making compliance a growing headache.

Start-up of the week

Here’s an interesting example of an AI agent in action: Zania AI, a compliance-focused tool that monitors regulations, automates audits, and flags risks before they escalate. Sounds great, but for me the big question is whether it’s too broadly defined to work effectively in complex, high-stakes environments? We’ll soon see. I’m excited to see what solution they come up with.

Extra toppings

With any luck, AI agents will finally kill off the never-ending parade of SaaS dashboards built solely to make manual tasks slightly less painful.

Will this be the AI podcast apocalypse?

Mon, 07 Oct 2024 09:02:58 GMT

Newsletter #17

Hang around for some Big 4 “cost efficiency opportunities”

A word from the person behind the laptop

Happy Monday!

A lot of news since last newsletter: OpenAI released a new model, Mira Murati leaves OpenAI and OpenAI received a new crazy valuation. I promise the content of this newsletter is not OpenAI related though, but perhaps its not a good sign that a lot of this is not product related stuff (while Claude Sonnet 3.5 is quietly becoming a personal favorite for many).

Today’s theme: we’ll be diving into the Deloitte report on how GenAI is being used in the legal industry.

And like everybody else I’ve been playing around Notebook LM from Google which means I’ve actually created a podcast that runs through the main conclusions if you dont have time to read through it yourself.

Note: If there was a problem with the hyperscale of podcast production before because you’d only have to get a mic, then this problem is about to explode because of NotebookLM.

Amazing tool though!

Check out the result:

The future is (probably) already here

I thought I’d never have to slave myself through another consultant report from a Big 4 firm after I quit, but this report from Deloitte gives some pretty good insights into what lawyers themselves expect from GenAI at this point.

We’re at the point where everyone acknowledges GenAI will reshape legal work, but the product hype isn’t exactly matching up. GenAI companies oversell, and lawyers are left unimpressed when the products don’t deliver. Meanwhile, a whole army of legal professionals is already using ChatGPT and Claude to handle basic tasks, often under the radar.

So what do the lawyers think themselves?

What does the report say?

Firstly, the report was written based on interviews with not only attorneys and court room lawyers, but also other people from corporate law: Deloitte basically conducted an online survey from February to March 2024, targeting senior legal leaders (e.g., GCs, Heads of Legal) to understand Generative AI's impact on corporate legal departments.

Find the full report here.

The survey gathered responses from 43 leaders across industries like Financial Services, Technology, Life Sciences, and Energy, representing legal departments of various sizes (from under 25 to over 200 employees) with participants from the US, Europe, and the UK. So a lot is covered here. So far so good!

Benefits

Let’s start with the benefits - and surprise, surprise: the Big 4 consultancy says it represents a so-called “cost efficiency opportunity”.

GenAI (like all other technology) represents a cost efficiency opportunity.

I don’t want to spend too much time on the cost efficiency because that’s a story we’ve all heard by now either as “AI will steal your job” or “this process can be optimised”. Less time spent on each task = cost efficiency.

Efficiency aside (and this is where I think it gets interesting) the report focuses on GenAI’s ability to improve the day-to-day experience for both lawyers and clients alike. For lawyers, this means less time on routine, mind-numbing tasks and more on strategic, meaningful work - which, Deloitte found, could even boost morale, with 60% of respondents expecting an uptick in job satisfaction.

I mean, they might end up disappointed, but it’s interesting that people aren’t necessarily worried about losing their jobs (and in my opinion probably shouldn’t be); instead, they’re excited at the prospect of focusing on more engaging tasks.

Next up, GenAI also has the potential to reshape how legal departments engage with the broader business. Imagine AI-powered chatbots delivering quick, clear legal guidance or an AI-driven "legal front door" that streamlines access to essential information. This can perhaps eliminate legal’s reputation from bottleneck to a more “valued” sparring partner (to put it politely).

From Deloittes report

Last but not least, GenAI might push lawyers to focus more on data. If you can identify patterns in litigation and investigations, it could give you a competitive edge. The report notes that:

“Better access to, and insight from, legal and business data will enable more data-led decision making and more proactive risk management by legal teams.”

But knowing law firms, data hygiene probably hasn’t exactly been a top priority over the years, so there’s likely a big cleanup needed to make the most of the data they’ve accumulated - and that could take years if the expertise doesn’t exist within the organisation.

Now let’s talk about risks

There has been a lot of focus on risk. Perhaps also a bit too much in my opinion, but the report also mentions what some of the risks factors for legal are.

The biggest barriers to Generative AI adoption, according to the survey, are the usual suspects: financial and resource constraints, plus the challenge of making a strong business case - especially with the tech still being so new. The report states that:

A slim majority of our respondents report no budget having yet been made available for Generative AI initiatives, restricting the ability of legal teams to meet their objectives and timelines.

This is really weird in an area with so much capital, but I guess the reason for not priorising would be “why fix something that is already working” so competetion will hopefully be the real driver of investments here.

Then there’s a big “skills” when it comes to getting Generative AI off the ground. It’s not just about adopting the tech, but also about making sure legal teams know how to actually use it. That means for example training staff on how to prompt AI effectively, understand its capabilities, and tackle the ethical questions that come with it. It might also involve bringing in fresh talent with AI and data science expertise.

I’ve been teaching AI to lawyers, and it’s interesting how many are more focused on the regulatory aspects than the actual application of the technology. Of course, that makes sense - it’s what we’re trained for - but it’s still curious that when something has the potential to completely transform how we work, the primary concern is how it’s regulated.

Which brings me to the next point: regulatory and compliance risks

I’m honestly surprised that only 38% mentioned this as a barrier. In my conversations with legal professionals, it feels like the biggest concern, especially when it involves external vendors. I’ve even heard from startups in the legal AI space that they’ve stopped mentioning AI entirely when pitching to clients - too many get spooked by the regulatory and confidentiality issues. Instead, they just focus on showing the functionality and skip the “AI” label altogether.

Last but not least

The report is definitely worth a read - or you can listen to the “homemade” podcast I shared earlier. But keep in mind, a lot has changed since the data was collected, and things are moving fast in this space. The benefits and risks might look a little different today.

One piece of advice that still holds true: experiment with these tools yourself. It’s the only way to really understand the pros and cons firsthand.

In other news…

“Dollar, Dollar Bills, Y’all”

OpenAI's valuation just hit a staggering $157 billion, thanks to its latest funding round. Investors are betting big on the company’s continued dominance in AI, despite questions around profitability. OpenAI will also be moving away from their non-profit structure.

AI regulation in California

California Governor Gavin Newsom officially vetoed the widely supported SB 1047 which I wrote more extensively about in another newletter. Newsom expressed concerns over its implementation and potential impact, putting a stop to legislation that many believed would bring significant changes to state policy.

Start-up of the week

This week’s pick is a little different: Casetext isn’t exactly a startup anymore. Founded over a decade ago, Jake Heller, the founder, is sharing insights on the opportunities for vertical SaaS in the legal space and how they made it work. A lot has changed in the 10+ years since they launched and the last couple of years everything has changed.

Extra toppings

Is privacy dead?

Crikey! Aussies reckon AI adds more work, not less

Tue, 10 Sep 2024 08:05:56 GMT

Newsletter #16

Stay around for some news from the land down under

A word from the person behind the laptop

No secret - I love Australia (and my Aussie wife didn’t even have to hold a gun to my head to make me say that).

The people, the nature, the sports (even if I haven't the faintest idea what's happening), and the strong sense of mateship and go-getter attitude. Sure, the meat pies might never compete with Danish pastry, but there's something about the Aussies' no-nonsense, down-to-earth approach to life that I find great.

Anyways, the story of this week is the Australian Senate Committee’s findings on testing out AI to support with government tasks. This is AI without the lens of VC dollars, start-up hype, or hyperscale dreams. This is government officials rolling up their sleeves and hoping LLMs can make their daily grind a bit easier - and this is where it gets interesting.

Because the result shows that it actually does the opposite.

Aussie AI might look different in the future

As a staunch advocate of the idea that “AI change is already here”, it’s refreshing to see someone cutting through the noise and summarise the results of the LLM with a quote like this about their experience:

“It wasn't misleading, but it was bland. It really didn't capture what the submissions were saying, while the human was able to extract nuances and substance. I think that is not a bad summary.”

Short, sweet, and brutally honest. If the ASIC was to use an AI setup like the one in the test it would actually add more work for the employees. Here’s why.

Aussie, Aussie, Aussie: AI, AI, AI

Earlier this year, Amazon put AI to the test for Australia’s corporate regulator, the Securities and Investments Commission (ASIC), using submissions made to a Parliamentary inquiry.

Between January 15 and February 16, 2024, ASIC ran a proof of concept to see how well LLMs could summarise those submissions. What on the surface seems like fairly simple task - but actually requires a lot of attention to detail.

Teaming up with Amazon Web Services Australia (AWS), ASIC set up an proof of concept (PoC) AI project. The PoC wasn’t about making AI a regulatory or operational tool just yet - it was more of a trial run to see what AI could do. The findings were all reported and are publicly available here.

ASIC is Australia’s integrated corporate, markets, financial services and consumer credit regulator

A fair conclusion would be that there’s still work to be done. The LLM-generated summaries lagged behind those created by humans across all criteria that had been determined as part of the project.

It might just be this use case and/or the implementation, but it seems like a pretty methodologically well done analysis, where there was both openness to cut out some boring work by using AI, and the right support to actually implement this in practice.

In my opinion the timeline was perhaps a bit too optimistic measured against the level of perfection needed from a tool like this. Government offices (should) have a low tolerance for errors, and the trial was only running for a month. But to understand how it was done we should look at how it was actually done in real life.

Here’s how ASIC put the LLM to the test

So let’s have a look at how the project unfolded: forget everything about traditional AI benchmarking that you’ve seen plastered all over Twitter for the last few years. For the ASIC

”the objectives of the PoC were to explore and trial Gen AI technologies, to focus on measuring the quality of the generated output rather than performance of the models”.

And honestly, that’s a smart move because unfortunately very few people care where your AI ranks on some leaderboard or which model you’re using as long as it gets the job done. It’s all about results, not the tech behind it. Solid start!

The workflow used for the project

Phase 1: Model Selection

The first challenge was picking the right AI model for the job. ASIC tested three different models: Llama2-70B (from Meta), Mistral-7B, and MistralLite. Each model was tasked with summarising public submissions made to the Parliamentary Joint Committee inquiry on the consulting industry.

Some examples of prompts from the project are: “What does this document say about ASIC?” and “How should conflicts of interest in audit firms be regulated?”

On the surface these are pretty simple, but it turns out, the LLM isn’t great at picking up on those subtle nuances the employees of ASIC would find obvious. ASIC gave each model the same prompts and compared their outputs to human-created summaries.

To ensure fairness, the models were blind-tested - referred to only as Model A, Model B, and Model C. The humans reviewing the summaries didn’t know which AI was behind which output, so there was no bias in the judging. After several rounds of tests, Llama2-70B emerged as the best of the bunch - not because it was flawless, but because it was less flawed than the others. So, ASIC took Llama2-70B forward for the next phase.

Metas Llama2-70B was the model chosen for the PoC

Phase 2: Optimisation

Only a week was set aside for the optimisation phase. The adjustment and fine-tuning of the LLM to see how the adjustment of the prompts to the change of technical settings like "Top-K" (how many potential next words the LLM should consider) and "Temperature" (which controls how random the LLM outputs are) affected the output. What could get the LLM to produce more accurate, coherent summaries.

Like always the challenges that ASIC met during this phase were perhaps not the ones they expected (like it often is with AI PoC’s btw). One big challenge for example was to get the LLM to list page numbers for where ASIC was mentioned in the submissions. Turns out, LLMs don’t naturally track page numbers, so the team had to get creative to solve this problem.

They chose to chunk the documents by page and trained the LLM to associate specific content with metadata - basically tagging each piece with a page reference.

Temperature explained by Colt Steele

The problem is, this kind of optimisation takes time, and one month isn’t nearly enough to tackle challenges like this. In fact, they ended up having to leave this part out in the end.

Phase 3: Final Assessment

With the AI optimised (or as optimised as it could be within the timeframe), ASIC entered the final testing phase.

The AI-generated summaries were blind-tested against human-written ones - ASIC staff weren’t told which summaries came from an LLM and which were from their colleagues. Each summary was scored on criteria like coherency, consistency, and how well it addressed the specific prompts. In the end, the human summaries scored 81%, while the LLM came in at a disappointing 47%.

So, what went wrong? The LLM struggled with nuance and failed to highlight key points or missed them entirely. It also tended to include irrelevant information, or worse, hallucinate details that didn’t exist in the original submission. One of the major issues was the models inability to capture complex, context-heavy arguments, which left the summaries feeling a bit flat and bland.

Here are some quotes from the ASIC assessors in the report:

Sounds familiar?

That’s because it’s the same challenges almost every organisation is facing with LLMs right now - these things take time to get right.

I’ve got to hand it to the Australian administration for being so open about experimenting with AI, especially in a space where it clearly has potential. But I think a lot of organizations still underestimate the amount of effort it takes to actually produce good results. It might even require your organisation to change an entire workflow.

ASIC’s proof of concept drives that point home. With just a one-month implementation window, you’re bound to hit limitations. But it’s not just about time - it’s about creating the right circumstances. Experimenting means testing, failing, and trying again.

Hopefully, this is just the start, because the potential is there.

In other news…

AI and copyright: again, again, again…

A new study commissioned by Germany’s Copyright Initiative has concluded that training AI models on copyrighted material constitutes copyright infringement under European law. According to the report, infringement occurs at multiple stages, from the collection of copyrighted works to their replication in AI output. This adds fuel to the growing legal debate on ‘fair use’ in AI training, with more experts questioning the legality of current practices.

‘Duel’ might get a driverless sequel

Aurora Innovation, the Pittsburgh-based autonomous trucking company, is gearing up to expand its operations. The company announced plans to extend its current Fort Worth-to-El Paso route all the way to Phoenix by 2025. This expansion aslo marks a change in the legal landscape. I don’t see solution currently, but perhaps the incremental roll might actually be the solution to make it driverless trucks on the roads legal.

Start-up of the week

This one isn’t strictly an AI startup for legal folks (we’ve seen enough document software for now), but more of a personal recommendation based on a tool I’ve been dabbling with. Playground is an image generation platform that simplifies the process by providing an intuitive framework and enhancing the quality of output by narrowing your input. It’s surprisingly effective. Give it a spin here.

Extra toppings

Talking of image generation: AI is tricking us into doing the boring work…

Schwarzwald off the rails

Wed, 21 Aug 2024 08:17:52 GMT

Newsletter #15

An introduction to German geography? Keep on reading

A word from the person behind the laptop

I can’t help but wonder if the Brothers Grimm, with their wild imagination, ever envisioned that the dark forests and twisted tales they conjured would one day inspire the name of a generative AI company.

Is this image AI generated? Yes/no?

An AI company, mind you, that’s now cranking out unhinged images of copyrighted cartoon characters wielding semi-automatic rifles and promoting drug and alcohol abuse - all backed by a billionaire who dabbles in spaceship construction… Probably not.

Anyway, that’s the theme of the newsletter this week.

Enjoy.

In The Hall of the Mountain King

I have a friend who once told me that after a big night out, he’d always play ”In the Hall of the Mountain King“on his ride home from a night out, speeding up on his bike to amp up the drama. It’s like creating your own personal epic, where your drunken legs propel you through the city, making you the narrator of a late-night fairy tale come to life.

Using the Stable Diffusion model of Black Forest Labs gives me that same vibe. No wonder the dramatic Schwarzwald in Southern Germany has inspired countless fairy tales, and it should be no surprise that a GenAI company would tap into that magic when naming their operation. Their latest Stable Diffusion model creates results that make the images that your boomer friends and family post on Facebook look like the sketch of a 4-year old. So clearly there’s a new AI powerhouse in play here.

The test result of BFLs latest Stable Diffusion Model look better than the other actors in the market

But Black Forest Labs isn’t a new player in the AI world. They’re led by Robin Rombach, Patrick Esser, and Andreas Blattmann - the original brains behind the Stable Diffusion models that flipped open-source image generation on its head. Their work on latent diffusion models laid the groundwork for Stable Diffusion and influenced key design elements in heavy hitters like DALL-E 2, DALL-E 3, and Sora - and they just raised a Seed round of $31 million from Andreessen Horowitz.

So, why am I rambling about the Schwarzwald, Stable Diffusion models, and my friend’s drunken bike rides home? Because this has the potential to crank up the legal drama in the AI image generation game by more than just a notch.

Elon entered the chat

First things first: How did Black Forest Labs all of sudden get such massive attention? Well, their Stable Diffusion model is now powering the image generation engine behind Twitter 2.0 (also known as X). The AI model, named Grok (don’t ask), is Twitter’s take on the many LLMs out there, but now they’re dabbling in image generation too. And to pull this off, they teamed up with Black Forest Labs.

Grok is reportedly modeled after Douglas Adams' humorous science fiction novel The Hitchhiker's Guide to the Galaxy.

What’s worth noting here is that this model has far fewer safeguards compared to other Stable Diffusion models on the market. This lack of guardrails has unleashed a wave of wild and downright disturbing images - everything from school shootings to blatant copyright infringements. OSINT Analyst Oliver Alexander has been going through these questionable outputs on Twitter, and trust me, they’re NSFW.

As he rightly points out, this is an open-source tool that anyone can use however they like.

But there’s a big difference when it comes to an individual generating images and a company making the technology available for millions of users worldwide. Just look at what happened when Google released Geminis image generating function.

So what’s the legal drama?

Let’s break down the legal drama. First off, the question everyone’s asking: Can you really put such powerful image generating technology in the hands of a bunch of Twitters crypto bros? Seems like a Monkey JPEG disaster waiting to happen.

This is a free Monkey JPEG. Sell it and make a profit.

But jokes aside, the real legal mess here is twofold: IP/Copyright and safeguards. However, they’re really two sides of the same coin, as proper safeguards can help prevent IP and copyright infringements in the first place - but it's often discussed separately because the legislation addressing IP issues are already being addressed in court.

As you already know, AI needs a ton of training data to produce high-quality results, and a lot of that data is copyrighted. This raises big questions about the legality of collecting and using this data and what happens when these AI outputs start looking a lot like (or straight up copy) someone else’s work. Grok is already spitting out images that have put the lawyers of Disney, Warner Bros, and Nintendo into a hyperloop (pun intended) of infringement cases.

The IP and copyright issues are obvious, and the landscape is changing. Just to repeat myself from an earlier verison of this newsletter:

“Notable examples include “Authors Guild v. Google” and “Perfect 10, Inc. v. Amazon.com, Inc.”, where the outcomes were surprisingly favourable to the tech giants. Some legal experts argue that these cases could have easily resulted in less favourable outcomes, impacting the trajectory of digital copyright law significantly. For instance, in “Authors Guild v. Google”, the court ruled in favour of Google’s book-scanning project, determining that it qualified as fair use.“

This also means that the approach of Ex-Google CEO Eric Schmidt that says successful AI startups can steal IP and hire lawyers to “clean up the mess” might not work as well as it used to. Let’s not forget that this is a guy who says you still need to show up at the office to be productive even though research shows the opposite. Strong 00’s vibe.

The Verge brought the story with the Eric Schmidt quotes that have now been taken down.

In my opinion what’s interesting here is the collision between the so-called "ultimate free speech platform" and the AI safeguards.

This same conflict plays out across many areas of society, which is why the absolutist approach might not play out as nicely as it sounds. Twitter 2.0 will have to limit content on the platform and the Grok output, which flies directly in the face of the worldview Musk keeps promoting. Not that there’s anything wrong in promoting free speech on SoMe platforms btw, but the odds are 100 % that you’ll end up sounding like a hypocrite when you remove/prioritise content based on your own beliefs.

As Jacob Mchangama, CEO of the Future of Free Speech Project at Vanderbilt University, pointed out when Elon bought Twitter in 2022:

“In short, there is a compelling case to be made for why free speech should be strengthened, not weakened on social media. But the skeptics are unlikely to be persuaded by a conception of free speech based on partisan grievances and trolling. Instead, Musk should focus on demonstrating how the benefits of a more robust commitment to free speech on Twitter will outweigh the harms.”

If Grok turns into a partisan AI mouthpiece pushing Elon’s worldview, we’re not just stuck where we were before - we’re actually worse off. Now, that partisan view is AI-fueled, raising the stakes and complicating the landscape even further.

The real question is whether Twitter and Black Forest Labs can craft a Stable Diffusion model that both respects the legal framework and still champions free speech. My guess? Probably not.

In other news…

Finally some law enforcement on deepfakes

San Francisco City Attorney David Chiu has filed a lawsuit against the owners of 16 top nonconsensual deepfake porn sites on behalf of the People of California. The suit claims these companies violated state and federal laws against deepfake pornography.

New privacy regulation making (brain) waves

Colorado just made history as the first U.S. state to pass a privacy act that includes neural data under the category of “sensitive data” protected by law. This move sets a new precedent in data privacy, recognizing the importance of safeguarding the most personal and intimate data - our thoughts.

Start-up of the week

No real website yet, but YC-backed startup Promptarmor is already making waves by doing something refreshingly simple: explaining how their solution works. While most startups drown you in buzzwords and vague promises, Promptarmor offers a clear breakdown of their approach to safeguarding AI. No one currently knows what the need for a solution like this will be, but my guess is that they’re on to something.

Extra toppings

Talking of copyright infringements…

IT’S A BUBBLE!

Thu, 15 Aug 2024 09:25:05 GMT

Newsletter #14

Hang around for an introduction to the Japanese stock market.

A word from the person behind the laptop

So apparently a 5 % drop in the stock market now qualifies as either a deep crisis or the start of a recession. I can’t figure out if it's the fault of Kamala Harris, the Japanese Central Bank or AI, but my Elon-powered-Twitter feed suggests that it’s probably a cocktail of it all.

How people sometimes connect these dots will forever be a mystery to me, but I guess that’s how Twitter 2.0 keeps me intrigued.

“Run you fools”

Anway, in this post-vacay scribble, we’ll have a look at AI bubble tendencies and what it might mean for the use of AI in legal. Hint: There’s a difference between big tech and legal firms.

Disfruten!

Even broken analysts are right twice a day

It’s easy to make fun of AI bros. The sub-category of crypto bros spending time telling me I’m falling behind if I’m not using AI. Exactly what “falling behind” means is not entirely clear to me though, but whatever. It definitely doesn’t sound good so please spare me.

AI bros in their natural habitat: Twitter.

AI bros are the canary in the coal mine for bubble tendencies, just like Monkey JPEGs were for crypto. The more online AI courses you see advertised by a guy in a snapback and calf tattoos, the closer we are to the bubble bursting.

About a week ago, the stock market dropped, with tech stocks losing over $800 billion in value - ouch. This sparked conversations about the massive investments big tech companies have made in AI and the relatively little value it has delivered so far. It raised questions about Nvidia’s valuation, the return of profits to investors, and cost efficiency. As I mentioned in last week’s newsletter:“The burn rate is through the roof, and generative AI demands more energy than the power grid can handle. Training these models is equally untenable, thanks to ongoing legal woes (courtesy of some alleged theft) and the sheer volume of training data required.”

As Reuters reports it all started with a big market drop in Japan. What the relation is to AI is not entirely clear to me, but right now everything is apparently related to AI.

Analysts love to cry wolf every time the S&P 500 takes a hit, often blaming an AI bubble - and they will eventually be right in my opinion. Perhaps in a more dotcom kinda bubble rather than a 2007 crash, but at one point the AI companies will take a hit. I just don’t think now is the time.

Vincent Hendricks, the philosophy professor behind the Centre for Bubble Studies (yes, that’s a real thing), says a bubble happens when prices are driven by speculation, not actual value. It’s when people buy in just to sell higher to the next guy, not because it’s worth it. Eventually, reality checks in, and the bubble pops.

Now, look at AI. The hype has all the signs of a classic bubble. Investors are pouring money into AI, not because it’s proven, but because everyone thinks it’s the next big thing. It’s all FOMO, not solid returns. This frenzy is inflating AI’s value beyond reason. If these technologies don’t deliver or the market gets too crowded, the bubble’s going to burst - just like Hendricks predicts when speculation overtakes substance.

But perhaps that’s the idea?

They are both good at making money, but legal firms are not big tech

As Casey Newton from Platformer explains: “Amazon didn’t turn a profit for the first nine years of its life. Uber reported its first full-year profit this year - 15 years after it was founded. The end of zero-interest rates has made it much more difficult for tech companies to operate this way. But particularly for the richest public companies, making long-term investments and ignoring investors’ complaints has long been the norm”.

In other words these companies are used to pouring money into long-term bets, brushing off investor concerns, and sticking to their guns because they know that winning the race is what counts.

Now, let’s shift gears to legal firms. Unlike Big Tech, law firms aren't in the business of sinking billions into unproven technologies and waiting a decade to see a return. They’re good at making money - really good actually. But that also means they’re far more cautious about where they put it. When it comes to AI, law firms are dipping their toes in, but they're not diving headfirst. In a recent report by Deloitte on the future of legal work 76 % of legal firms have no Gen AI adoption at this point - but still 87 % expect GenAI to get adopted within the next 2-3 years.

These numbers prove that there are still huge expectations on what to achieve with AI, but implementation is hard/slow/[insert your own reason]. Perhaps some lawyers are introduced to the use of GenAI by some businesses that promote primitive GPT-wrapper-tools, or perhaps their image is tainted by half-baked projects like Microsoft Copilot for legal, where the results so far might make you wonder if no lawyer at all would have been a better option.

What is this? Seriously.

No surprise that legal professionals might feel underwhelmed by GenAI if their first experience comes from such a shallow take on the legal profession. In a field full of stubborn sticklers, slapping your contract onto a GPT-4 powered tool isn’t going to impress anyone - and it sure isn’t going to magically boost your drafting skills. If anything, it might just turn you into a lazier, less competent lawyer. I’m not saying better tools don’t exist - they probably do. But companies, especially ones with Microsoft-level resources, need to realise that this isn’t a one-size-fits-all situation.

So, are you telling us not to invest in AI?

Actually, no. Quite the opposite - especially time! Dive in headfirst, use it as much as possible, find the flaws and the good bits, and figure out what tool works best for your specific legal needs. Like Foo Fighers would say: "One of These Things Is Not Like the Other". It all depends on your role, practice area, language, jurisdiction, and about a hundred other factors. The tools are all very different, and there are still plenty of areas in legal where the market is not very saturated. Good tools will come.

Right now, I’d be more inclined to go for an AI product that comes with some hefty consultancy fees (never thought I'd say that), because you need someone who can tailor the tools to your needs - and keep them updated. In other words: it’s not going to be cheap.

And if a website offers you a one-size-fits-all solution, run. Even the Balrog is less terrifying.

In other news…

What bubble?

AI startups pulled in about a quarter of the $23 billion in total venture funding for July, raking in $5.8 billion. Only healthcare and biotech companies outdid them, with $6 billion. But let’s not ignore the slight decline - $6.3 billion in June and $12.8 billion in May. Even so, AI funding more than doubled in Q2 to $24 billion, marking the highest since ChatGPT launched in 2022. Seems like the hype train isn’t slowing down just yet.

Another day, another AI related lawsuit

A major class-action copyright lawsuit against AI image generators Stability, Runway, Midjourney, and DeviantArt is moving forward to discovery. The defendants tried to get the case dismissed, but U.S. District Judge William Orrick allowed the Copyright Act claims to proceed. This is a big deal in the growing field of copyright lawsuits against AI companies. If this case goes to trial and the AI companies lose, it could open the floodgates and seriously disrupt their business models.

Start-up of the week

This week’s pick is Bryter, an AI platform promising to streamline legal workflows without a line of code. It sounds great on paper - automating processes, managing compliance, and smoothing out client interactions with just a few clicks. But I also feel like I’ve heard it all before. I know that this is possible with the current technology, but I’m curious to see the trade-off between individual company customisation and automation. Anyway, I like the idea!

Extra toppings

Waymo has been set loose!

Subscribe now

Suns out, Bots out

Wed, 07 Aug 2024 08:02:43 GMT

Newsletter #13

Hang in there for some online Harvey drama

A word from the person behind the laptop

Like a proper Continental I’ve spent my July on a boat in Greece, guzzling Retsina and pretending Greek salad is a balanced meal, as long as it’s got a rooftop of feta.

Relatable video

With AI news flooding in like my inbox after a week off, I’ve decided to grace you with a rundown of the biggest AI updates from the summer. Yes, you’re welcome.

Next week, we’ll return to our regularly scheduled programming. Until then, endure!

1# Harvey drama on Twitter

Like before in this newsletter we've discussed the actual value of an AI product like Harvey. An AI legal tech product with bunch of hard-to-find-customers. Despite their numerous partnerships and truckloads of cash, trying to find a genuine customer review of Harvey is like hunting for Mew in the first couple of Pokémon games.

This summer, they raked in another $100 million, prompting even non-legal folks to raise their eyebrows. With a valuation like that, you'd think there'd be at least one review floating around somewhere. Even some anonymous profile talking shit on reddit. But no.

Anyway. New drama kicked off with a tweet predicting Harvey's demise as "roadkill." The author later backtracked after Harvey's fundraising success (because money apparently equals credibility).

Yet, quite a few people ran with the first statement - and surprisingly, many agreed. The real question is: How long can Harvey keep riding this gravy train if the product is simply not good enough? The error rate in the LLMs powering these products (GPT-4 as far as I know) might simply be too high for the lawyers to accept the results.

From my experience these products take a whole lot of customisation which takes time and resources. It you add that to the cost of using the models then the profitability decreases.

Anyway, some people say that it has to do with “land grab” and being the first to market. Makes perfect sense if you’re into that sort of thing. The real entertainment will be watching if big law firms end up saddled with hallucinating software, forcing lawyers to “delve” into the messy aftermath.

Time will tell.

If you - or anyone you know - have experience with Harvey and want to talk, please reach out to me. I’m curious to hear more.

2# A slow summer for Sam

Talking about the models powering Harvey, OpenAI has had a pretty sluggish summer. They did a delayed release of the voice assistant and a prototype of SearchGPT, which I’ll dive into in a future newsletter, but the revolution is not exactly around the corner with these two updates.

But besides that, OpenAI seems to be accumulating new problems. I’ve mentioned the slew of pending legal cases, but now there’s a potential profitability issue on the horizon.

The burn rate is through the roof, and generative AI demands more energy than the power grid can handle. Training these models is equally untenable, thanks to ongoing legal woes (courtesy of some alleged theft) and the sheer volume of training data required.

This is all laid out and well-argued in Ed Zitron's newsletter.

On top of that, Microsoft—the sugardaddy to end all sugardaddies—is starting to consider OpenAI a competitor rather than a friend. This spells bad news for OpenAI, while probably not having much of an effect on Microsoft, which has access to the IP powering OpenAI’s models.

Personally I still believe there must be some way to make this work. As I always say: This product is already useful as it is. But the way to profitability and a sustainable business model might be different from what we see today,.

3# Who’s the big bad regulator now?

There are some potential new constraints on AI development - and it's NOT in the EU (written while enjoying the sun on a random Greek island).

California State Senator Scott Wiener is behind SB 1047, a bill mandating that developers of large AI models must test and certify their creations to ensure they won't cause significant harm. This legislation has received endorsements from Geoffrey Hinton and Yoshua Bengio, two of the most prominent and frequently cited researchers in the AI field.

Vox has the story on SB 1047

Critics worry this heightened liability could discourage companies like Meta (who just released Llama 3.1) from releasing open-weight models in the future. That would be a particular blow to startups, many of which use open-weight models because they cannot afford to train their own models from scratch.

SB 1047 wouldn’t apply to the Llama models Meta released - not even the model with 405 billion parameters. Meta says that training this model required 3.8x10^25 floating point operations (FLOPs). That’s less than half the 10^26 FLOP threshold that triggers potential liability under SB 1047.

But given the exponential growth of large language models in recent years, it’s easy to imagine Meta’s next generation of language models exceeding the 10^26 FLOP limit. If SB 1047 is the law in California at that point, Meta could face a new set of legal requirements.

With smaller models improving daily, even they could soon face the same scrutiny. While I’m skeptical that the new California bill will be a cure-all, it might at least inject some much-needed accountability into the mix - and that’s not necessarily a bad thing.

See you next week!

The robots will probably take over while I'm on vacation

Thu, 04 Jul 2024 08:56:48 GMT

A word from the person behind the laptop

Summer is here, which means I'll be slowing down the newsletter machine for the next month or so. I'll be sending out newsletters every now and then until I ramp up again in August.

This newsletter will provide you with a list of extra AI/law toppings you can enjoy during the summer.

Enjoy the sun! Boom!

Extra toppings (summer edition)

Hallucinations still plague legal AI: at least an LSD trip ends eventually

In a shocking twist, AI tools for legal research still haven’t figured out how to stop hallucinating. Legal professionals must continue verifying that key propositions are accurately supported by citations—who could have guessed? The study reveals an urgent need for rigorous, transparent benchmarking of legal AI tools. Unlike other AI domains with established benchmarks, legal AI vendors prefer keeping their systems as opaque as possible, offering little to no details about their models or benchmarking results.

Some firms, like Paul Weiss, have taken matters into their own hands, conducting secret internal evaluations. This disparity leaves smaller players guessing about AI risks and functionality, further fueling skepticism about AI’s potential to improve access to justice.

Despite the hype, these AI tools aren’t ready to replace your legal intern just yet. They might be handy for initial research, but without hard evidence of reliability, claims of hallucination-free legal AI systems are, at best, a legal fiction for now.

Read the study here.

Does naive RAG mean that you’re naive? Probably.

Many companies are testing out RAG and setting up RAG pipelines as we speak. If you’re unsure what RAG means, check out my article on the subject for the University of Copenhagen’s LegalTech blog.

But the different types of RAG significantly impact the results you get, which is why I recommend this article on Medium. It thoroughly explains the differences between Naive RAG, Advanced RAG, and Modular RAG.

Your clients don’t want you to use AI - yet

Legal companies are under immense pressure to implement AI to improve legal work and streamline the connected processes around the work. However, a recent report from Thomson Reuters reveals that this demand isn’t coming from clients.

Thomson Reuters’ 2024 report on generative AI in professional services is a real page-turner. Generative AI, once the stuff of sci-fi, is now here to revolutionize the legal, tax, accounting, risk, and government sectors. Since late 2022, platforms like ChatGPT and GPT-4 have shown just how disruptive they can be, churning out high-quality content faster than ever.

While not everyone has jumped on the GenAI bandwagon yet, over half of professionals think they should be using it daily. This report dives into their perceptions, current usage, and the future of work in this brave new AI-driven world.

See you soon

And that’s a wrap for this week! Remember, while the robots might be plotting to take over your job, they’re not quite there yet.

I’m heading off on a much-needed vacation but stay curious and critical while I’m gone - and don’t let the bots dazzle you too much. See you soon!

You’ll be spending more than 10 hours using AI after this

Thu, 13 Jun 2024 08:10:01 GMT

Newsletter #11

Scroll, scroll, scroll if you want to see why AI is a terrible bar exam student

Subscribe now

A word from the person behind the laptop

Apple passed the crypto test.

During the peak of the crypto/blockchain frenzy, Apple stayed conspicuously silent about integrating crypto features into your iPhone, apart from the usual encryption for protecting data - which, ironically, no one seemed to care about meanwhile losing millions on monkey JPEGs.

Now, Apple is introducing Apple Intelligence (AI… lol) to your iPhone, marking a significant investment in AI. It's likely because they see it as beneficial for users and a lucrative opportunity. This also means closer collaboration with OpenAI, which seems to have ruffled the feathers of a certain sink-carrying tech mogul.

This week, we’ll look into Apple’s recent announcements, privacy/security updates, and more.

Hang tight!

Wait… D-Day Anniversary was a week ago

Seeing Americans parachuting from planes at the start of June immediately brings one thing to my mind: D-Day. Maybe it’s because I’m a massive fan of the HBO series Band of Brothers, but that was all I could think of during Apple’s WWDC kickoff this year.

I’m not gonna blame you if your reference wasn’t the same as mine. I’ve clearly watched too much BoB.

While Apple’s conference may not carry the historic weight of D-Day, it was delivered with a similar sense of high-stakes drama. Expectations soared, especially with the highly anticipated integration of AI into Apple’s ecosystem.

But honestly… the result fell a bit flat for me. The conference, though polished and impactful for many, didn’t quite meet the overblown expectations surrounding Apple’s foray into AI. The combination of Apple and AI seemed promising, but AI is still too risky for big tech to fully embrace.

As Sarah Perez of TechCrunch aptly put it:

“(..)with iOS 18, which Apple showed off at WWDC 2024, the company is taking a more cautious approach: Instead of trying to overwhelm users with too many AI features to count, the Cupertino tech giant is carefully rolling out AI where it believes it could be useful. That means the tech won’t be included where it could be much of a threat to the carefully crafted consumer experience of using an Apple device.”

This is probably a smart choice. Apple has branded itself as the protector of privacy and the teacher’s pet when it comes to responsible tech development (whether this holds true is an ongoing discussion). This seems to have worked out for them. But how does the new AI initiative collide with this image?

The crypto test

Unlike the speculative frenzy that has characterised the crypto world - often a solution in search of a problem - Apple’s cautious approach to AI seems motivated by a desire to protect its credibility.

They are integrating AI only where it benefits the user, maintaining their reputation for user-friendly tech implementation.

The presentation passed my personal crypto test as well:

Step One: If you need to highlight the underlying technology just to make it seem smart, it’s likely a fail.

Step Two: If the same objective can be achieved using other technology with a proven track record, there’s probably no reason to include this particular tech in your stack.

Step Three: Bubble tendencies.

I participated in a blockchain summer school and several hackathons to assess the practicality of blockchain technology. The conclusion in many cases seemed to be the same: we might as well build the same product using technology with a proven track record.

I’ve spoken to people at companies initially built on blockchain who have since stopped advertising their blockchain base because now it evokes images of monkey JPEGs, Sam Bankman-Fried, and dubious crypto festivals.

Journalist Ryan Broderick has covered crypto festivals

As I mentioned last week, LLMs are different. They might check some boxes for certain use cases, but that’s not the point. The underlying technology is already proving to be useful. While tech companies keep chanting “AI, AI, AI,” we can hope (or fear) that the hype will subside as AI becomes naturally integrated into everyday applications.

This is where Apple comes into the picture.

Apple will make you an AI expert

Ethan Mollick’s “10-hour rule” famously claims that you’ll need at least 10 hours with AI before you find it genuinely useful. I’m with Ethan on this one. But thanks to Apple, even those reluctant to clock in their AI hours won’t have to try that hard anymore.

Why? Because Apple is weaving AI so deeply into its systems, you’ll be using it without even realising it.

Mike Gorman nails this in his article, “Apple Intelligence is Right On Time” Apple isn’t flooding the market with glitzy AI gimmicks. Instead, they’re methodically integrating AI into their ecosystem, enhancing user experience without overshadowing their core values. Here are some of the main points:

Strategic integration: Apple’s AI enhances features (like personalised health insights and intelligent photo management) in iOS 18, running subtly in the background.

Hardware synergy: Apple Intelligence leverages the latest hardware (iPhone 15 Pro, M1 Macs) to boost AI performance without compromising user experience.

Ecosystem enrichment: AI enhances Apple’s ecosystem, improving usability and streamlining interactions, with features like Siri’s advanced app controls.

Privacy-centric design: Last but not least. Apple is focused on on-device processing for tasks like Siri commands and keyboard suggestions, keeping user data secure.

While these features created quite a buzz online, the real highlight (besides finally bringing the calculator to the iPad) is how Apple manages to process AI while staying compliant with privacy standards. This is a newsletter for lawyers after all, so let’s dig into the details.

Brace yourself for the next Mira Murati Meme

When asked in a Bloomberg interview whether OpenAI used YouTube videos to train their video model Sora, Mira Murati couldn’t provide an answer. Unsurprisingly, this moment quickly turned into a meme that’s still making the rounds online.

I wonder how many journalists out there are gearing up to ask Mira Murati the same question, but in the context of OpenAI’s newly announced cooperation with Apple.

But here’s why Apple’s AI initiative is different when it comes to privacy and security, and might change how we view the use of data in AI applications and hardware:

1. On-device AI processing

Most AI tasks (e.g., Siri commands, photo enhancements, email support) are handled directly on the device, thanks to integrated machine learning models. This keeps data local, ensuring privacy and security.

Apple isn’t chasing frontier AI models like Google or Microsoft; instead, they’ve built smaller, efficient models for on-device use and a medium-sized model in the cloud for extra heavy lifting. The power of this on-device model can be compared to Phi-3.5-mini.

2. Selective requests to OpenAI

While Apple doesn’t have a cutting-edge model like GPT-4, they leverage OpenAI’s expertise for questions their models can’t tackle. For complex tasks like advanced text generation and content analysis, Apple selectively sends anonymized and encrypted data to OpenAI, ensuring user privacy by asking the user if they're willing to use ChatGPT.

3. Encryption and anonymization

Apple encrypts and anonymizes data before sending it to OpenAI, safeguarding user privacy. This is how they describe it:

When Apple Intelligence needs to draw on Private Cloud Compute, it constructs a request — consisting of the prompt, plus the desired model and inferencing parameters — that will serve as input to the cloud model. The PCC client on the user’s device then encrypts this request directly to the public keys of the PCC nodes that it has first confirmed are valid and cryptographically certified. This provides end-to-end encryption from the user’s device to the validated PCC nodes, ensuring the request cannot be accessed in transit by anything outside those highly protected PCC nodes. Supporting data center services, such as load balancers and privacy gateways, run outside of this trust boundary and do not have the keys required to decrypt the user’s request, thus contributing to our enforceable guarantees.

As someone on Twitter put it: this is a compromise that makes no one happy. People that want to use ChatGPT will have to click an extra button, and the privacy folks think this compromises their existing privacy settings on the Apple devices.

We’ll see.

But Elon Musk defs wasn’t happy with the new partnership:

In other news…

Crikey! Google’s SurfPerch is having a yarn with corals

Google Research and DeepMind have just launched SurfPerch, an AI tool that’s practically an aquatic Steve Irwin. This beauty automatically processes thousands of hours of underwater audio, helping scientists get the lowdown on how coral reefs are holding up. Turns out, even the Great Barrier Reef needs a good chat now and then. Thanks to SurfPerch, marine biology just got a whole lot more conversational!

Another legal team with a 100 % AI adoption rate incoming?

PwC has partnered with OpenAI to add some AI magic across their consulting services - I guess this is especially relevant for legal. This collaboration is set to enhance everything from document review to report generation. Following a successful run with AstraZeneca’s legal team hitting a 100% adoption rate, it looks like this kinda cooperation is the way of testing out whether AI brings value in a corporate setting.

Startup of the Week

Meet the Belgian startup LegalFly. They secured €2 million back in 2023 to revolutionise your legal workflow with AI. Yes, they’re packing in the buzzwords, but here’s the twist: it’s actually quite smart, using an agent-based approach to automate complex tasks.

I haven’t tried it myself yet, but it sounds promising. And btw: companies branding themselves as “The Most Secure” on the landing page says something about the main AI concerns clients have atm.

Extra toppings

Here’s someone arguing that ChatGPT was actually pretty bad a the bar exam.

“All in all, it’s just another brick in the AI wall”

Thu, 06 Jun 2024 08:23:43 GMT

Newsletter #10

Read further to see the inside of a Concorde

A word from the person behind the laptop

Rebranding linear algebra as “artificial intelligence” may be the most successful marketing campaign of all time - but its also undeniable that AI brings real value.

How do I know?

Firstly, I use it myself all the time - simply because of the fact that it’s useful. Secondly, you currently start seeing results from all branches - legal included - where this new human/computer cooperation is invaluable.

However, this doesn’t necessarily mean we’ll experience another wave like the release of ChatGPT. Making LLMs publicly available changed the entire narrative about AI's usefulness. The road ahead will likely be more stable, but there’s still a lot of work to be done. Today, we’ll look into why this is.

Oh, and we’ll also chat about Nvidia superchips, AGI, and the usual startup stories.

Cheers!

“All in all, it’s just another brick in the AI wall”

I’ve been watching the ongoing discussion about the so-called plateau of improvement for LLMs, especially on Reddit. I decided to examine these arguments for and against.

First I wanna say that I actually don’t think it matters so much at the moment because there’s so much potential in the models as they are right now. Even with their current capabilities, there will be many more use cases for personal, business, and government use. For lawyers, the potential might be even greater than in other sectors - but the pace might be slowing down.

Large language models not learning anymore?

Let's start with this tweet by Ethan Mollick:

“If AI really does plateau at the 60-80th percentile of human ability (no sign it will/won’t), the impacts may be stabilizing. Whatever you’re best at (often what you enjoy most), you’re likely to be better than an AI, but whatever you’re not good at, AI can help fill in the gaps.”

Ethan Mollick is essentially saying that AI still holds significant potential as a tool to enhance human performance, particularly in areas where humans are not experts. This perspective is crucial for understanding the progress we’ve made, even if the performance and improvement of LLMs might be stalling.

So “stalling development” is not the same as we won’t see AI getting integrated into more aspects of our life because there’s still so much potential.

Background on the plateau of improvement

So what does stalling development actually mean?

Gary Marcus, in his article "Deep Learning Is Hitting a Wall" argues that despite deep learning's successes, it's facing significant limitations. The article is around two years old and a lot has happened since, but the general logic remains: current AI systems, particularly deep learning models, struggle with tasks requiring true understanding and reasoning, often making critical errors.

“Indeed, we may already be running into scaling limits in deep learning, perhaps already approaching a point of diminishing returns. In the last several months, research from DeepMind and elsewhere on models even larger than GPT-3 have shown that scaling starts to falter on some measures, such as toxicity, truthfulness, reasoning, and common sense. A 2022 paper from Google concludes that making GPT-3-like models bigger makes them more fluent, but no more trustworthy.”

Even though this prediction didn’t turn out to be true, the conclusion remains relevant. Trustworthiness is still very much a challenge, regardless of fine-tuning or advanced RAG pipelines. Marcus contends that solely scaling up data and models won't overcome these issues.

He advocates for a hybrid approach that integrates symbolic reasoning with deep learning to achieve more reliable and comprehensive AI systems. We won’t have time to dig into that today, but I recommend his article.

The argument against the plateau

Despite the debate, I believe the current capabilities of LLMs are already game-changing. As I mentioned earlier their potential applications in various fields, especially in the legal sector, are vast as is.

While it’s true that there are challenges and limitations, ongoing improvements show that we are far from reaching the end of AI advancements. Let’s dive into the specifics of why this is:

• Compute power: GPUs, TPUs, and NPUs are becoming more efficient and powerful, driving the performance of AI models.

• Context size: Context size is increasing rapidly

• Tokenization: Tokenization techniques are improving, as seen in GPT-4o, making models more effective across various languages.

• Multimodal models: Models are becoming multimodal, performing better on tasks that integrate text, image, and other data types.

• Data quality and quantity: Data quality is constantly improving, and the availability of multimodal and synthetic data is increasing.

• Efficient training: Pre-training and post-training processes are becoming more efficient and automated.

• RAG: Retrieval-Augmented Generation (RAG) techniques are improving, enhancing the reliability of AI outputs. More on this later.

• Cost reduction: The costs associated with developing and deploying AI are decreasing.

The question for legal teams is whether the AI advancements above outweigh the lack of trustworthiness in responses from a potential legal AI. Or is it worth waiting to see if scaling LLMs continue to improve their performance? Especially because a recent study from Stanford University found that LLMs and RAG tools are unreliable in 17 to 33 % of cases.

I’ll spend the next week looking into the paper above, but knowing that LLMs can hallucinate doesn’t negate their usefulness in my opinion. The current benefits are substantial, and waiting for a perfect model could mean missing out on significant advantages these tools offer right now. Imperfections exist, but their potential to enhance productivity and efficiency is already transformative.

No autopilot just yet

Now let’s talk about commercial airplanes for a second.

The progression from not-flying to flying was obviously significant, but I often hear the argument that progress has been “small” since the development of the jet engine. That’s a misconception. The changes have just been less obvious.

There have been substantial changes to commercial airplanes even in recent times, but most of them aren’t immediately noticeable from the outside. For example, the use of digital controls (fly-by-wire) has replaced mechanical controls in the cockpit, the aircraft body composition has become stronger and lighter, and engines are now quieter and more fuel-efficient, with the Airbus 350 as a well known example.

Comparing the development of LLMs to the aviation industry might sound odd, but it’s quite fitting. Just as aviation experienced a plateau in terms of jet propulsion advancements, AI, specifically LLMs, seems to be undergoing a similar shift in the pace of development. However, it's important to note that just because the fundamental nature of flight hasn’t changed dramatically, it doesn’t mean we've hit a wall.

The Concorde was regarded as the pinnacle of mass aviation, though it wasn’t known for its comfort. Fortunately, the journey from London to New York took only 3.5 hours.

Aviation was already incredibly valuable in its early days. Air travel revolutionised global connectivity, and hopefully LLMs will also transform industries, including law. The key is to recognize the plateau not as a dead end but as a call to innovate differently.

LLMs have seen rapid advancements in recent years, with models becoming larger, more powerful, and capable of more tasks. However, we’re now reaching a point where simply adding more data and parameters might not yield the same leaps in performance.

So, while LLMs might seem to be hitting a wall, it’s likely that we can break through with the right tools and innovation.

In other news…

"Heja Sverige!” Microsoft bets big on Sweden

Someone once told me that Denmark and Sweden are the two countries that historically have been in the most wars with each other. Today it’s mostly banter, but Scandinavian rivalry is beautiful. Unfortunately, it seems like Microsoft put their money on the Swedish horse.

Microsoft is pumping $3.2 billion into expanding its cloud and AI infrastructure in Sweden over the next two years. They’re rolling out 20,000 advanced chips across three data center sites in Sweden, using Nvidia, AMD, and their own in-house AI chips.

At least the Nvidia CEO has a Danish nickname

Nvidia CEO Jensen Huang announced that their next-generation AI chip platform, "Rubin," will be available in 2026. This new platform is set to push the boundaries of AI performance and efficiency. People are excited when Nvidia is involved!

Start-up story of the week

This week, we're looking at Definely, the legal tech startup trying to make document drafting less of a chore - like everyone else (sorry). Founded by a bunch of legal experts and tech enthusiasts, their platform uses AI to streamline the whole process. The idea is to cut down on the mind-numbing parts so lawyers can actually focus on the law. Keep an eye on Definely.

Extra toppings

The logic of LLM scaling is applied to other areas of life:

Privacy, safety and security are sexy now

Thu, 30 May 2024 11:56:17 GMT

Newsletter #9

Hang on for a trip to an Inca citadel!

A word from the person behind the laptop

I never imagined I'd be able to post a picture of Machu Picchu in a newsletter about privacy and security being sexy. But hey, welcome to 2024, where anything is possible, and here I am, embracing it.

This week, we dive into the takeaways from Microsoft's Build conference and their AI ambitions. We’ll also chat about OpenAI’s new (and former) safety team, plus some other juicy tidbits from the AI world.

Hang in there!

A monopoly wrapped in a friendly user interface?

Let’s not kid ourselves: Microsoft has a rather convenient structural and legal business advantage.

With its software embedded in the infrastructure of most enterprise companies, convincing customers to buy a little more from their existing supplier is a cakewalk compared to enticing them into a completely new ecosystem. It makes it much easier to document compliance and avoid contractual negotiations if Microsoft's standard contractual terms apply.

This applies just as much to new and shiny AI applications as it does to the latest version of Excel - boring, but effective.

Rapping about “logs, SCIM, SSO” doesn’t make it fun

Let’s start with this mesmerising video that makes the series ‘Silicon Valley’ seem like a documentary:

Much like the legendary Microsoft Windows 95 launch where the leadership awkwardly danced to the Rolling Stones, Canva's Hamilton-inspired rap video may well earn its place in the annals of cringeworthy corporate moments.

What caught my attention was their struggle to incorporate rhymes about enterprise infrastructure, security, and privacy. None of them flow exactly like Biggie which - to be honest - is probably the least of their problems.

But why write rhymes about integrations and tech setups at all?

Because this is where the money is, highlighting how crucial these aspects are for customers.

Developing an AI application and ensuring it complies with legal standards through a third-party wrapper? Good luck convincing the compliance team.

Onboarding a new AI-focused application in an enterprise setting can feel like Hiram Bingham’s first trip to Machu Picchu: battling through the jungle, getting battered by branches, and bitten by unknown creatures—until you finally reach the sacred Incan city and uncover its mysterious treasures.

It was a struggle to get to Machu Picchu back in the day even though the locals were well aware of its existence.

Microsoft, however, is in a different position. They can essentially land their helicopter right on top of the mountain, providing both the shortcut and the proper equipment for those who prefer the procurement jungle route.

I might have gotten a bit too carried away with the Machu Picchu analogy. This post is about a tech conference: back to Microsoft Build!

Microsoft’s grand strategy: trust

From their recent presentations, it’s clear that Microsoft plans to use its existing market position (which, by the way, makes total sense). The corporate entanglement in Microsoft’s products has the potential to deepen even further with some of the initiatives presented at the Microsoft Build conference. Headline could be AI, AI and AI - and then some other minor things.

The 3 takeaways from Microsoft Build

So, three takeaways from Microsoft Build:

1. AI Everywhere: Microsoft’s vision involves AI not just as an add-on but as an integral part of the Windows experience. Dedicated AI processors in PCs are their next big bet, making sure that every text box you encounter is an opportunity for their AI to make an appearance.

The most famous example was the new “Recall” function that I also mentioned in the newsletter last week. It basically takes screenshots every x amount of minutes so you can search in your history based on visual factors. Potentially good for micromanaging bosses. Not so good for their employees.

2. A New Era of Chips: Qualcomm, not Intel, will power the next generation of ARM-based laptops. Running generative AI locally requires more computing power, making strong battery life essential. This could make Windows on ARM much more compelling.

For those of us who still remember the Wintel days, this shift away from Intel is remarkable - and probably a necessary business move in the AI era.

3. Still a lot of uncertainty: The future of tech is uncertain, with endless possibilities in generative AI, but Microsoft is determined to make sure it’s a feature they can’t be ignored for. The company is throwing everything at the wall—from agents to multimodal models—and hoping something sticks. If you can’t innovate your way to the top, you might as well outlast the competition.

Here are some additional highlights:

Where is this going?

I must admit that Microsoft has already come a long way since the first implementation of OpenAI into the Azure environment. The tools are user-friendly and easy to use, and while the use cases are not 100 % yet, this will probably come with time - we just dont know exactly where yet. To quote Benedict Evans in his latest newsletter:

“(..) there’s this ‘AI’ thing, and Satya Nadella is determined to grab onto it with both hands and ride it back into the arena. No-one in tech really knows for sure what generative AI will look like in a year or two years, and no-one is sure what kinds of software and experiences will really make sense. Agents? Wrappers? Multimodal? Voice? Microsoft is spraying mud at the wall with the best of them. Incumbents always try to make the new thing a feature, and Google and Microsoft are trying to add an LLM to every text box on the internet.”

In essence, Microsoft’s strategy seems to be about leveraging its existing dominance and customer trust to pivot into AI seamlessly. It’s a classic move: why build something new when you can just integrate it into what everyone is already using?

Enterprise companies already trust Microsoft, so they are likely to embrace these new AI integrations without much hesitation. Trust, after all, is the key factor here—Microsoft isn’t just selling new tech; they’re building on a foundation of reliability, established relationships and existing contracts.

As AI continues to evolve, this trust will be crucial in navigating the complex landscape of enterprise adoption and innovation.

In other news…

OpenAI’s new (and old) safety & security team

OpenAI has announced the formation of a “Safety & Security” team led by board directors Bret Taylor, Adam D’Angelo, Nicole Seligman, and CEO Sam Altman. This committee will evaluate OpenAI’s safety processes over the next 90 days, coinciding with the training of their next frontier model. This move follows the departure of key safety researchers and criticism that safety has taken a backseat to product development.

Survey reveals limited awareness of AI tools

Do you feel like you haven’t fully taken advantage of AI yet? You’re not alone. A Reuters Institute survey of 12,000 people across several countries found that only 53% of U.S. respondents have heard of ChatGPT, despite it being the most recognized AI tool. Usage is even lower, with just 7% using it daily. Awareness of Google’s Gemini and Microsoft’s Copilot is even more limited, with over 85% of Americans never having heard of or used these tools.

Start-up story of the week

Not really a “start-up” tbh, but I like the idea. Using AI, Transcript Genius allows attorneys to semantically search across vast libraries of transcripts and dynamically interact with individual ones to unearth relevant information.

Extra toppings…

I had a lot of fun looking through people’s results from Google’s AI overview this week

“No” was apparently Lost in Translation

Thu, 23 May 2024 08:31:07 GMT

Newsletter #8

Scroll through for a real-life Black Mirror episode.

Subscribe now

A word from the person behind the laptop

I try to find different things to cover each week which means not covering OpenAI every week.

Luckily there’s usually a lot happening in the space of AI and law so it's usually not that difficult - but recently it seems like the legal challenges against OpenAI are stacking up with the same pace as they’re rolling out new features.

That means we’ll cover OpenAI (again) this week, but this time with a focus on… (drumroll) Scarlett Johansson. This legal case has the potential to show us how courts will tackle major issues like the use of biometrics, artistic likeness, and consent.

We will also talk about billing, Black Mirror and more copyright.

See ya on the flip side.

Bob Harris wouldn’t let something like this happen to Charlotte

I’ve always been obsessed with the movie ‘Lost in Translation’. My millennial hipster heart is a sucker for Sofia Coppola's grainy shots of Tokyo’s neon-lit streets. The movie transports me back to a simpler time when the biggest problem was translating between Japanese and American languages and cultures, and screen time was naturally limited by the size of a cathode-ray tube.

The movie was also Scarlett Johansson's breakthrough on the big screen. Since then, her husky voice has become iconic. This distinctive quality made her a perfect fit for the voice acting role in "Her," where she (or her voice) played an AI assistant.

Since the launch of ChatGPT, tech bros have been obsessed with the idea of a personal assistant inspired by the movie "Her." With the new capabilities of GPT-4o, this might suddenly be possible. This prompted (no pun intended) OpenAI's Sam Altman to reach out to Scarlett Johansson with an offer to be the voice of GPT-4o’s character, Sky.

When she declined, they seemingly proceeded without her consent. What could have been a great tech gimmick blurring the line between fiction and reality ended up being yet another example of “move fast and break things” even though many people thought tech had moved on from the 00’s.

To coincide with the launch of the demo, Altman posted a tweet celebrating the new AI assistant, which many noted sounded suspiciously like Johansson's voice.

Here’s the infamous tweet by Sam Altman considered by many as a reference to the movie of the same title

A comparison video circulated online, highlighting the eerie similarity between Scarlett's voice and the AI-generated one—though, in my humble opinion, there are notable differences as well.

I’m not one to judge whether it’s actually Scarlett Johansson’s voice or not, but there’s no question that the events leading up to the demo release put the case in a different perspective.

Listen to a comparison here:

Scarlett’s lawyer must be rich

This isn’t the first time Johansson has taken legal action against major corporations. She previously sued Disney for releasing "Black Widow" on Disney+ simultaneously with its theatrical release, breaching her contract which guaranteed an exclusive theatrical run. Disney responded aggressively, disclosing her $20 million upfront payment for the movie, a move seen as a tactic to undermine her case.

“In a time where we are all grappling with deepfakes and AI reproductions,” Johansson stated in a response to the GPT-4o demo release, “I believe these are questions that deserve absolute clarity.”

The new lawsuit against OpenAI not only seeks to protect her personal and professional rights but also addresses broader concerns about the ethical use of AI and digital likenesses. These issues are hot topics that everyone seems to be discussing, but there are currently as many definitions as there are Marvel movies. No exaggeration.

The case is important though, because it could help set significant precedents for how biometric data and personal attributes are used in technology, and whether and when explicit consent is required.

Let’s go through 3 reasons that she might have a good case:

Judges might not treat technology companies as favourable as before

As mentioned in the tech podcast “Hardfork” recently, judges might historically have given tech companies like Google what could be considered a favourable treatment in copyright cases during the 2000s and 2010s.

Notable examples include “Authors Guild v. Google” and “Perfect 10, Inc. v. Amazon.com, Inc.”, where the outcomes were surprisingly favourable to the tech giants. Some legal experts argue that these cases could have easily resulted in less favourable outcomes, impacting the trajectory of digital copyright law significantly. For instance, in “Authors Guild v. Google”, the court ruled in favour of Google’s book-scanning project, determining that it qualified as fair use.

If the courts are not swayed by OpenAI as they were by the young, fresh, and "Don't be evil" Google, the outcome could be different - especially given the changing public perception of tech companies. This shift in attitude might significantly influence the Johansson vs. OpenAI case.

Scarlett or not: the reference is clear

Even if the AI voice isn’t an exact match for Scarlett Johansson, the reference to her voice might be clear enough for people to think it's her. This raises obvious questions about the boundaries of using a likeness - auditory, visual, or written - in technology, especially in AI where imitation is a core functionality. Can AI replicate or imitate someone’s unique attributes without explicit consent?

This issue becomes even more pressing when the voice in question is a significant part of an individual's brand and can potentially mislead consumers.

A surge in AI copyright cases

The Johansson vs. OpenAI case is unfolding in the context of a broader wave of copyright disputes in the AI domain. As AI technologies become more sophisticated, the legal system is increasingly challenged to address the nuances of these new capabilities. Recent high-profile cases, such as “New York Times v. OpenAI” and “Authors guild v. OpenAI”, illustrates the growing complexity of AI-related copyright issues.

In “New York Times v. OpenAI”, the newspaper sued OpenAI over the unauthorised use of its articles to train an AI language model. The court might find that even though the AI-generated content was not an exact replica, it was similar enough to mislead readers and dilute the value of the original works. This case underscores the importance of protecting original content from unauthorised use, especially in the age of AI.

Let’s see what the future holds

OpenAI has put “Sky” on hold for now, and claims that a different voice actor is the voice behind “Sky”. This probably won’t stop Johansson’s law suit though.

Her case against OpenAI taps into this broader legal trend, emphasising the need for clear guidelines and protections in the rapidly evolving intersection of AI and law.

As courts continue to navigate these uncharted waters of bits and bytes, the outcomes of such cases will be crucial in shaping the future of digital copyright law and the rights of individuals in the age of AI.

“Does it get easier?” Charlotte asks in ‘Lost in Translation.’

My guess is it won’t.

In other news…

Welcome to Black Mirror Season 71

Microsoft announced Copilot+ PCs at its Build conference, a new category of laptops designed for AI integration. A new feature “Recall” sounds like a joke - but unfortunately isn’t. The feature basically allows your laptop to constantly take screenshots to make user activity searchable via an on-device AI model. Now privacy professionals will have something to write about on Linkedin for the next many months. Oh dear…

And back to copyright…

Music generator Suno raises $125 million at a $500 million valuation. Despite the impressive funding, there are growing concerns about the lack of transparency regarding Suno’s training data. As the debate over AI-generated content and copyright intensifies, Suno's undisclosed data sources could spark legal and ethical questions, potentially impacting its future in the competitive AI music industry.

Start-up story of the week

It’s hard to find start-ups that are not working within the space of document reviews. Anyway, this one isn’t. At least not a 100%. Brightflag now uses invoice summaries for clients to enhance billing transparency. Will this work? I guess we’ll soon find out.

Learn something new

Have you cloned your own voice yet? Your voice might not be as distinctive as Scarlett Johansson's, but it's still incredibly easy to clone. I recently tried ElevenLabs, and while it’s not perfect, it’s pretty damn good - like most AI products these days.