So someone just alerted me to a part of this post which I missed.
“Our eng org is starting to spend as much on tokens as we do on headcount.”
At 1000+ headcount, assuming average salary at $200k, employee costs are already >$200m/yr which is almost as much as the “ARR”.
If the internal token costs are another $200m, that means Harvey will be in the red for about $100m to $200m BEFORE accounting for the costs of tokens to customers who are on fixed pricing.
I imagine they might be accounting for this in R&D instead of operating costs but in reality this is an ongoing operating cost.
And the costs of tokens are just going to keep increasing…
Legal AI superempowers normal individuals with no legal background to fight big institutions in bureaucracies and in courts on a level knowledge/skill playing field, for the first time in human history. As such, it is one of the most inspiring applications of AI.
Kirkland & Ellis, the world's highest-grossing law firm, is setting aside $500M to build its own AI platform rather than rely on tools available to its rivals (Financial Times)
(Visit Techmeme dot com for the link and full context!)
If the models were connected to a firm's knowledge base, results would be different - and they would vary wildly depending on the quality of the law firm corpus.
No Westlaw, firm knowledge, or web, so citation criteria are graded against law in the model's weights (?). This in not reflective of how attorneys actually use AI. Every piece of work should ideally be grounded in a precedent - which is not the case here.
The post-AGI world is a Minecraft server where your friend has built automated factories for everything you could possibly need
That’s usually the point where everyone stops playing because it’s boring
Sometimes people build massive weird art projects at that point, but that gets boring too
The only thing left to do after that is to create new simulated environments (games), where you can compete and feel meaning again—eg. Have duels or a duel ranking system
It’s not hard to imagine how reality could be a fractal of simulations. Once a world becomes “solved”, the only thing to do is to create a new unsolved world to conquer
I'm sorry. People downplaying Kirkland building their own AI platform is either a Harvey/Legora employee/investor or frankly have no idea what they are talking about.
You are going to see MANY more professional services companies (law firms, banks, accountants, consultants, funds) do this exact same thing. DIY Is a very real option now utilizing open source or foundational models. Octus are seeing lots of clients do it themselves - they know their workflows, internal processes and special sauce better than anyone - why delegate an agentic process to a wrapper when you can do it yourself just as good, at a fraction of the cost / not tied to a third party contract.
Will all of them do this? Of course not. As I've always said the wrappers have a true reason to exist (some funds don't want to rely on single models or just don't have the firepower to do it themselves). And that's why they have been so successful. But the barriers to entry in that space is small and hard to defend. Distribution/story telling is all that matters right now and Harvey/Legora/Rogo/Hebbia have done a TREMENDOUS job at both of those.
i just gave each of Codex, Claude and Legora the same task.
had them draft an extension notice letter for a loan agreement.
gave them each the two prompts at the bottom of this post.
Legora: invented the defined term "Extended Maturity Date," which isn't in the loan agreement. pretty bad considering there's two substantive sentences in the notice. using the Word plugin, it then gave me an essentially unformatted document.
Codex: no necessary tweaks but i wouldn't be able to resist polishing Codex's output. for example, it defined the "Loan Agreement" after the parties names instead of after the date of the doc. it spit out a near perfectly formatted document.
Claude: took longer but spit out a near perfect document that i could send on without editing. subjectively, Codex wins in formatting.
P1: "tell me what the term of this loan is and when an extension notice needs to be delivered to the lender. tell me the section and pdf page numbers for your analysis"
P2: "please draft me a notice as a separate document in word. the sample notice is a notice from a lender to a borrower and i only want you to use it in terms of formatting. use the other one more in terms of mirroring the substance of the two sentences or so in the body of the letter."
Using AI to do legal work faster is fine. Using AI to think harder about your clients' problems is the actual opportunity. Most lawyers are stopping at faster.
Legal tech used to be something most lawyers could safely ignore.
You needed an email. You needed PDFs. You needed to survive the firm’s document-management system. But the machinery underneath? That could stay with IT.
In the age of AI, that’s no longer an option. 🧵