Co-Founder https://xitter.cf/t.co/zuw4R71SmA. Building autonomous legal transactions. Lawyer and software engineer.
San Francisco, CA
Joined May 2010
- Tweets68
- Following57
- Followers2.1K
- Likes36
I've seen a bunch of posts about Harvey and Legora announcing essentially the same intake product on the same day last week. Pretty funny, but I haven't seen anyone point out that they also both proudly announced firm-wide adoption by Hogan Lovells Cadwalader on the same day.
Hogan CEO on Legora: "Legora has been supporting our lawyers across a range of legal work and helping us explore new ways to deliver high-quality service to clients, while maintaining the professional judgment that remains at the heart of what we do."
Hogan CEO on Harvey: "Through Harvey, we’re perfecting the ways to capture and apply the expertise that exists across our firm on a variety of legal work. That has helped our lawyers work more effectively and strengthens the support and advice we provide to clients."
Things are getting weird in legal tech.
Is anyone gonna be using Muse in 6 months? Imo either something better will have replaced it or ppl just stop caring like they did with Sora. The odds of neither happening seem close to zero.
The biggest problem I have with Copilot is it doesn’t work.
We’re building Copilot as a new OS for work that spans every model, every form factor, and every task. Today, we’re announcing our biggest update to Copilot to date, bringing four things together:
· Autopilot: proactive and long-running agent built for the enterprise
· Code: build apps with Copilot, hosted inside your company’s tenant
· Home: Chat + Cowork together
· Office: now fully embedded in Copilot (and Copilot embedded in Office, of course!)
Plus, you can invoke Copilot in Teams, and we’re introducing Today, a proactive experience that surfaces the most important information from across M365 without needing to ask for it.
The way we work is changing and so are our workflows. This update brings AI into that flow, from answering a question, to building an app, to getting work done on your behalf.
Legal AI is becoming a meaningless category.
When Harvey first came out, it was interesting that a company was building AI for the legal industry because general-purpose LLMs had extremely limited legal capabilities. Even if Harvey’s real-world performance wasn’t significantly better than Claude or ChatGPT, the fact that they were even trying separated them from the pack. Today, both OpenAI and Anthropic are building products for legal work, and much of the discourse on LinkedIn and X is about whether vertical legal AI companies are in trouble as a result. My answer is that if you’re building a generalized legal AI product, then yeah, you might be.
But the entire premise of generalized legal AI has made little sense to me for a while now, and it's one of the main reasons we set out to build Walton. Anyone who’s worked with lawyers knows that the legal industry is so multifaceted that even calling it a single industry feels misleading. The difference between an M&A associate and a criminal defense attorney is so vast that they may as well be in completely different professions. So I’m highly skeptical of claims that a legal AI product can broadly outperform an off-the-shelf AI system without substantial, separate investment in each practice it claims to support.
To be clear, I’m not saying the Harveys and Legoras of the world don’t provide value. Obviously, LLMs are transforming the practice of law. I’m saying that they may not offer much beyond off-the-shelf AI. Email is essential to lawyers, yet lawyers don’t use legal email. They use Outlook.
This has significant implications for companies building generalized legal AI because it undermines the common comparison to AI coding tools. Products like Claude Code and Codex have gotten so good at end-to-end software engineering that real-world usage generates meaningful insights that can inform model training and make the products even better. That feedback loop was possible because even early AI coding tools like GitHub Copilot were immediately useful to a broad range of software engineers. There’s been an assumption that legal AI would follow a similar path.
But coding is relatively monolithic at the product level. Front-end and back-end developers can both use Claude Code, even when they’re building completely different types of software.
Legal is different. To get comparable coverage, you might need twenty products instead of one, each requiring deep domain understanding, not to mention access to proprietary, confidential data. That’s a massive undertaking. Short of AGI, there’s no reason to assume one or two companies will build the best legal product for every practice or even every workflow within a single practice.
The second wave of legal AI will look very different from the first.
So he’s getting fired from big company now right?
I am done with this shit. It is over. The state of engineering right now is horrible. It has been half a month since I started a new role at a big company. Nobody knows anything here. The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can. I have heard multiple times from higher management that pushing code is not a bottleneck, so why are we slow? People are working 12 to 13 hours a day just to press enter. Nobody is reading anything. Humans in corporate are doing nothing on their own. Everyone, literally everyone, from an L1 to an L7 engineer here is doing the same thing. Talk to Claude. There is no sense of victory. Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs. It is so soul-sucking. I would not mind it, to be honest, if we were at least given the time to check out the code and see what is going where. But no, the goal is to just ship. No matter what happens.
Elon's been promising full self-driving for nearly a decade, yet Tesla's FSD system still requires a human in the driver's seat, eyes on the road, ready to intervene. As a result, its real-world utility remains limited.
But the technology continues to improve, and at some point a relatively small increase in reliability will cross a threshold that makes supervision unnecessary. When that happens, the gains in utility will be massive.
Every autonomous system has a delegation threshold, and it largely depends on the stakes of the task being automated. Cross the threshold, unlock massive utility.
At Walton, we think we're on the verge of crossing the threshold for complex legal transactions.
Read more here: walton.ai/blog/the-road-to-u…
The cynical take on Harvey’s new post-trained model Tenet is that it doesn’t need to be much better than its already-strong and cheap-to-serve Chinese base model (Kimi K3), because its main purpose is to repackage that model into something law firms are comfortable using with sensitive client data.
I think this is a reasonable take, especially because Tenet’s legal-specific benchmarks are not particularly impressive, but there’s a lot more going on here.
First, Harvey is repositioning themselves (again) as a legal AI infrastructure/research org rather than an app-layer company, which I think is the right strategy. They have the capital needed to pull it off, and it’s a much less competitive space than the app layer, for now at least (as post-training costs come down, I expect this to change).
Second, the reason Tenet isn’t very good is that Harvey has only recently started building out their post-training infrastructure. Tenet was mostly post-trained on synthetic and publicly available data. For future Harvey models to get materially better than their base models, they need both access to real matter data coming directly from user sessions and permission to use that data for training.
They already have access. According to the billboards, Harvey is used by 75% of the Am Law 100, and you better believe they’re developing methods to capture every meaningful signal those users generate. Most of the new features in Harvey II suggest that's what they're doing (e.g. memory).
So now they just need permission, which is more complicated but not impossible. Harvey’s standard terms prohibit them from using firm data for training. But they’re solving this in a clever way. They’re going to enter into agreements with firms that allow them to train on firm data as long as the resulting model belongs to the firm, not Harvey. And I’d be shocked if Harvey doesn’t also work out a way to pool some of that data into a shared training set for their generally available models.
If Harvey can pull this off, they’ll have a data flywheel that will be extremely difficult for even the frontier labs to compete with, because it will require real lawyers running real matters on a platform designed to mine every user action. This is what Anthropic (Claude Code) and OpenAI (Codex) did with coding. Of course, there are big differences between coding and legal, especially around verifiability, but even if LLMs can’t “solve” legal the way many proclaim they’ve solved coding, Harvey is positioning themselves to at least try.
Finally pulled the plug on Fable for a less intelligent model because I can't stand talking to it.
If there's actually an inverse relationship between raw intelligence and the ability to sound like a human (e.g. due to aggressive RL/post-training or something else) then the smartest models will end up being a lot less useful than we think.
Something that’s been bothering me recently: why isn’t the agentic web called web4? Did we collectively decide to stop at web3?
Is there a word for this genre of aw shucks I did something amazing post? I think @sama is the originator but others are following in his footsteps.
8 days ago, while jogging, I asked Claude to solve the Riemann Hypothesis
It didn’t. 1.5 days later, it proved >= 67% of the zeros are on the line (prev: 41.6%)
Still not sure what that means, but some analytic number theorists seem excited
anthropic.com/research/riema…
We’re running out of ways to hype new models.
OpenAI’s unreleased next-gen model family Astra excels at complex, long-running multi-agent tasks.
“Previous agents would kind of hit a dead end and it would give up.”
“But then these newer classes of models… will basically just keep trying things and doing different things and stringing together different attacks until they're able to find a weak plate.” — @erinkwoo
The AI-native services story has come full circle. Sell software --> add professionals to sell AI-native services instead --> remove the professionals --> sell software.
So, this is interesting by @crosbylegal.
Their agents will do certain autonomous work without a lawyer reviewing *every single work output* and they plan to insure the work the autonomous agents do, ie to get clients to trust the work/output created and delivered by the agent like a human lawyer.
Feels like a way to earn trust (ie our agents do a really good job and we stand behind them like a human lawyer) , differentiate, take that big step of seemingly removing the lawyer from some of the output review (probably low risk outputs that they get the highest eval scores on) and going full autonomous. Need to read this post carefully. they’re *not* saying they will remove lawyers from all the review loops (which is how I’m sure some clients and lawyers will interpret this). Ryan, their ceo, crafted this carefully. They’re also saying this is how good our agents have become which is also interesting.
Also, great marketing of end-to-end native ai/agentic firm.
Another bet is being taken here on how clients will respond to this big step towards end to end autonomous legal work.
Very provocative (love it) and Ryan knows this hence the discussions with regulators, bar associations, insurers etc as this will trigger regulatory issues.
