i'm hiring ~1,000 lawyers in the next 90 days to help train frontier AI models. $100–120/hr. 20–40 hrs/week. remote. like reviewing a first year's memo, except the first-year is an agent. interested? reply with your practice area and i'll dm you

a short history of software replacing lawyers: 1973: lexis puts case law on a computer, and the 70s become the fastest decade of lawyer growth on record. 2011: headlines say e-discovery will replace armies of lawyers, and the US goes from 1.2 million of them to 1.37 million. 2025: donotpay, the "world's first robot lawyer," settles with the FTC, whose complaint says it never had lawyers test its work. 2026: AI can draft the contract, check the cites and scan every new rule in every state. if the last 50 years are any guide, a lot of what it makes cheap was never getting done in the first place.

before deciding lawyers are cooked, watch the demo. a horizon scan finds new rules, flags them for review and asks for an owner. a lot of the legal work AI makes cheap was never getting done in the first place.

renting out a movie theater in nyc on nov 4 to watch the new nathan fielder doc on elizabeth holmes. it's for lawyers who want to get paid to break legal AI. paralegals, legal ops and recent law grads welcome too :) reply "break" if you want an invite

radiology already ran this experiment. in 2016 we were told to stop training radiologists, and a decade later there's a shortage, because AI made scans cheaper and someone still has to sign every report

do the math on legal: about $400B a year of work, most of it confidential, so models have barely seen it. code got good on public repos, but law has to be taught one lawyer at a time

where i think this goes: 1. intelligence gets cheap but ownership doesn't, so we could end up with the most productive workers in history owning less of the economy. 2. once everyone can produce an impressive answer, a harvard degree stops proving you're smart and starts proving who let you in. that might make it worth even more. 3. when every argument, testimonial and credential can be faked, you start caring a lot more about who you've actually worked with, because information is endless and trust still takes years. 4. a legal opinion becomes almost free, while deciding what to risk and taking responsibility when it goes wrong stay expensive. professions shift from selling answers to standing behind them. 5. the forms disappear because AI fills them out, but it also gets cheap enough to judge every person, business and transaction all the time. less paperwork, more permission. 6. the most powerful AI company will look less like a software company and more like a school you never graduate from. it teaches your kids, advises your career, weighs in on your relationships and tells you which risks are worth taking. whoever defines "helpful" runs the school. 7. getting your time back won't feel like freedom if nobody needs you. a paycheck gives you money, but being needed gives you a place in the world, and whoever offers people belonging after their jobs disappear could end up more powerful than whoever automated them. 8. the real luxury is opting out: disagreeing with your AI adviser, rejecting an automated decision or choosing an inefficient life without losing access to everything you need. the people selling optimization to everyone will pay extra to escape it. none of this requires AI to become conscious, only useful enough that we stop questioning what it chooses for us. we're going to democratize intelligence and discover how many things were never about intelligence in the first place.

once AI reads your books and writes your opinions, whoever updates the model gets to update you. millions of independent thinkers on the same release cycle.

anthropic's S-1 says it's unclear when an AI agent's actions legally bind whoever deploys it, and its own legal plugins put responsibility on the attorney. so more AI means more work needing a lawyer's sign off, and the billable hour still can't price the part AI does in seconds

harvey and legora sell to lawyers, and legalzoom sells to people trying to avoid one. i'd buy legalzoom, run it on open source models, make every filing free, and charge for the lawyer doing the last mile via api

everyone thinks AI replaces lawyers. by the end of next year, we'll hire more of them than the largest law firm in america

client goals set the direction, firm standards set the floor, lawyer taste sets the style. but the same lawyer redlines differently for every client, so the persona worth modeling is the lawyer-client pair, not the lawyer

the next big buyer of legal work is going to be an agent that knows when it's out of its depth.

all of this is already normal in software, where an agent has a PR up before anyone's even read the ticket. coding agents got there first because engineers do their homework in public and get graded by tests in seconds. lawyers do theirs in private and get graded by opposing counsel years later, so models have read the law but barely watched anyone practice it. biglaw invented RLHF a century ago and called it the cravath system. associate drafts, partner redlines, repeat for eight years, except the reasoning behind every redline stays in the partner's head. getting that reasoning out of partners' heads and into the training loop is the real bottleneck for legal AI, and it's why demand for lawyers to train models is the highest i've seen in 10 years

models learned law from what ends up in court, which is like learning marriage from divorce filings. most legal work is deals that close and advice that keeps people out of court, and almost none of it ever ends up anywhere a model can read

AI isn’t taking lawyers out of legal work, it’s moving them from the start of the workflow to the end of it. agents do the first draft, lawyers make the final call, and the final call was always what clients were paying for.

you forgot the strongest one: already being an approved vendor, because nobody goes through procurement twice

the first question about AI was whether it could produce an answer, and the question that matters now is whether it can finish a job, handle the exceptions, and leave behind work someone is willing to sign.

running a marketplace is hard without a balance sheet because the client needs a lawyer by monday and pays you in 60 days. but that lawyer gets paid every two weeks (lawyers have rent, clients have procurement portals). and if one timesheet gets disputed the whole invoice sits, so you're floating payroll for people who aren't on your payroll. that 60-day gap where you're profitable on paper and broke on payday is the part nobody puts in the deck, but it's also the moat, because lawyers stay wherever the money shows up on time. every marketplace is a bank with a matching feature, and the matching was never the hard part.

an AI data company can have recurring revenue and still start from zero on every order. imagine a model flags a risky contract clause but misses an exception in an amendment. lawyers correct it, the lab gets useful data, and it orders another batch. on a revenue chart, that's a great account, but what got easier for the vendor? a vendor that's learning knows which lawyers catch exceptions like that, can tell a missing document from bad legal reasoning, and has a rubric the next group of lawyers can actually use. a vendor that isn't learning rebuilds recruiting, training, and QC on every order. i'd want to see gross margin on the second order, accounting for how hard the work is, and, more than the number, the reason it moved. a new model may need entirely new examples, and the lab's documents stay the lab's, so the learning has to live in the process: choosing experts, designing tasks, and getting work accepted on the first pass. after the experts get paid and the lab gets its data, what does the vendor know that it didn't before? it's the question i'd ask any expert marketplace, ours included. repeat demand can fund a business for years. i'd want to know whether each order makes the company better at the next one.