ThomAub retweeted
Most AI agents are trained in fake offices. We build digital twins of real ones.
What’s a digital twin? A copy of a real company, with the same tools, workflows and code, where AI agents learn to do real work.
Same people, same structure. All identities are anonymized.
What is a Digital Twin?
It is the base of every RL environment we build at Ooak!
An anonymized replica of a real company’s data environment: its docs, emails, tools and workflows.
Agents trained on real work beat agents trained on toy tasks.
Fun pod! We are also licensing company workflow data!
1. ooakdata.com/partner
2. Call with Chris
3. Done
a 50-person company’s slack and email data is selling for $300-400K right now, and the labs are calling brokers on fridays saying they don’t care what it costs.
here’s everything from our conversation with data broker ryan lock (sponsored by Ahrefs):
ryan sold his solar company ($35M revenue, 1,000 reps) last year and stumbled into data brokering when someone told his co-founder he could license his company data. the first offer gave him two days to accept or it exploded. his co-founder negotiated it up 4x. that gap is the whole business: the labs are desperate, business owners have no idea what they’re holding, and nobody’s advising the sell side. he launched in july, went viral once, and was doing 15 meetings a day with business owners. team of 13 now. every single lab reached out to him inbound through X.
the model is simple: buy data assets from companies, package them into training sets for frontier and mid-tier labs. what they want is operational data: email, slack, drive, accounting, shopify, klaviyo. raw text, not photos or video. the part nobody understands: it’s not volume, it’s context. a customer email that gets pinged into slack that gets resolved in a ticket is a complete story a model can learn from. switch from slack to discord two years ago and lose the chain? the data’s worth a fraction. ryan’s own solar data was nearly worthless because everyone just picked up the phone. his co-founder’s remote slack-native team got the full ask.
the numbers: average deal is $300-400K for a 50-100 person company, deals go up to $2-4M for the big ones. quick valuation proxies are number of software tools, years of full-context data, and headcount, because a hundred people generate idea-discussion-approval-feedback loops that five people in one office never write down. the de-identification alone costs $10-50K per terabyte, done by a third party who never lets ryan touch the raw files. and don’t try to scrub it yourself: tony becomes “employee six” everywhere consistently, or the whole set is destroyed.
ryan’s analogy for margins: raw data is oil at $100 a barrel, the labs turn it into jet fuel at $2,000, and the broker takes the risk that the barrel has no engine inside. the spirit airlines bankruptcy sold its data to google for $10M, which he thinks is low for the largest public deal. the data companies are all talked about as very profitable, but the real edge is procurement on demand: he forecasts what the labs will want next and ranks his pipeline by acquisition priority. current requests: circuit board CAD files, architecture firms, heavy construction, and video game 3D development.
the gotchas before you sell: never sign a perpetual exclusivity. perpetual use is standard (nobody can track files down the chain), but exclusivity should be termed, usually two years, or you’re locked out of a future revenue stream forever. and if you’re selling your business in the next year or two, a $500K data deal can become a due diligence landmine on a $20M exit. tom’s referrals keep saying exactly that. micro one pays $50-75K referral fees because they’re playing a different game. ryan pays 6% and tells brokers to charge the seller for running the process instead.
where it goes next: ryan thinks the buyer pool expands past the frontiers to enterprises training their own models. a 20-person HVAC company only knows what it knows, but add 150 years of industry data and you’ve got a vertical model, the harvey-for-law playbook in every trade. the video side is wilder: labs pay $10-20 an hour for people to wear headcams doing manual work, and there are state-sponsored warehouses in china and every low-cost country where workers rotate through sewing and dishwashing stations for robot training. jacky’s fraud prediction: 500 fake chinese companies with active linkedin pages.
ryan’s answer: if the simulation is that good, the data’s probably valuable anyway.
bonus: jacky’s own slack community, 500 people and 20K messages a month, has already gotten offers. “no amount of money, well, that’s a lie, I’d take my pants off for seven figures.” the over/under for ryan’s next 12 months was set at $300M gross, and micro one just passed a billion annualized. tom asked to become a master affiliate and jacky immediately called it master and slave. tony’s response: “the white guy’s still on top. we went back a hundred years.” the industry is a gold rush and the shovels are already being sold.
NGMI with @itstonyyu and @tomwang24, guest @RyanLockAI.
watch/listen ↓
CEO version of my
1. ooakdata.com/partner
2. Call with Chris
3. Done
Companies are sitting on years of operational data. AI labs need exactly that to keep improving.
At @Ooakdata we turn your workflows into a new revenue line for your business: earn up to $1M while keeping your IP.
How it works:
1- Get a live estimate of what your data is worth on ooakdata.com/partner
2- Fill a short form about your company
3- Our team reviews your request and gets back to you in 24h.
ThomAub retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
If you want in on the data broker thing, this is literally all it is
1. ooakdata.com/partner
2. Call with Chris
3. Done
If you want in on the data broker thing, this is literally all it is
> sign up as a micro1 referral partner and a Polyshares partner so you have your own tracked links
> find a business with 30+ employees and years of records sitting in their systems
> check the Wayback Machine for 3+ years of company history before you email anyone
> send the owner your link
> let the buyer evaluate the company and price the deal
> get paid up to $50K per completed micro1 referral, or a share of deal value on Polyshares
We submitted 46 data deals in 3 weeks doing exactly this and 13 are moving to next steps
The article has the whole process, the email I send owners and the earnings math, so read it next if you want to capitalize on this. Links in the comments
Handling a few TB of data with Ray, PubSub, CH by day launching millions of rollouts by might come talk to me!
Three months, one batch, a lot of late nights in SF.
We came to @ycombinator with @ooakdata to bring company workflows to AI training. Turns out a lot of companies are ready for it.
Now the real work starts, and we're hiring in SF and Paris.
👇 link in comments
ThomAub retweeted
Last week we ran out of table. Now we're running out of office...
New Ooak Data home in Paris coming soon.
ThomAub retweeted
Someone rewrote Kafka in Rust and embedded it in Postgres using pgrx 🤯
Kafgres is a Rust-based Kafka broker that runs within Postgres. Kafka clients connect to Postgres on port 9092 and literally can't tell the difference. 🧵
ThomAub retweeted
You dropped this king 👑
When several people talk at once, a transcript can get messy fast.
Our new Nemotron 3 Diarization model tracks who spoke when, even when voices overlap. It handles up to eight speakers, has 100M parameters, and is now available on @huggingface 🤗
ThomAub retweeted
When several people talk at once, a transcript can get messy fast.
Our new Nemotron 3 Diarization model tracks who spoke when, even when voices overlap. It handles up to eight speakers, has 100M parameters, and is now available on @huggingface 🤗
verify the Claude Agent SDK using Lean 🤯
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.
TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.
I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.
Is formal verification the future of coding (or at least, bug finding)?
ThomAub retweeted
How auditing RL Environments (ie. looking at the data) will help us explain model behavior
"models are benchmark shaped" bc GRPO style RL is essentially
- synthetic data generation by a model
- induced by the environments + harness design
- behavior (trajectories) reinforced by verifier design
the "data" is RL environments & tasks, a model generates a series of tokens in its exploration of that environment trying to solve the task
we don't really know what it will do, but we reward behavior that aligns with the verifier
Encouraged behaviors:
important: the only behavior that even has an opportunity to get rewarded is trajectories our environment "induces" or "encourages". something that isn't produced, can't be rewarded
if our Harness (via a tool description) or Task design encourages lots of parallel search calls over a massive document corpus to find a few facts, the model will start copying that behavior because it got rewarded for it
in the future, when some other problem is roughly shaped like that search Task, the model will mimic similar behaviors whether or not that strategy is optimal
models do things that their data encouraged them to do via reward
Tracing Behavior to Environments:
every behavior can be fuzzily traced by the accumulation of different environment design and reward combinations the model saw over training
i think we all explicitly know this but we don't have the interpretability tooling to understand induced behaviors at a fine-grained level over a massive training run
this is why I'm very exicting that ppl are swarming around Trace analysis over Evals/Environments
I bet a lot of "weird" or "bad" behavior comes from some quirks in environment design that inadvertantly rewarded that behavior
the only thing that even has an opportunity to get rewarded
Vestigial Behavior
RL is in part inherently exploratory. Models need to do a bunch of stuff in rollouts to figure out what works, the verifier will reward what works
But the verifier won't explicitly penalize any behavior that has a neutral effect but that co-occurs in a successful rollout
this is how we get behavior that used to be helpful in some tasks, doesn't hurt current tasks, and thus persists over time
similar to how humans have vestigial organs like the appendix
i bet we can collectively figure out a ton about model behavior by mechanistically working backwards from the data and I'm bullish humans and agents working together will make a lot of progress on this in the next year
ThomAub retweeted
Running long-horizon, agentic data projects is the best preparation for building robust enterprise AI apps bc it forces you to build thoughtful expert <> agent interactions and optimize LLM usage. Data companies have a massive advantage in serving the enterprise.
We now spend more on LLMs than on headcount. That does not make human operators, and especially experts, less valuable. It actually makes them more productive and more valuable. Purely synthetic data without experts embedded in the authoring process doesn't work, and data companies that don't dig into their own data will ship slop. They won't produce the nuance, realism, or complexity the task demands, and the labs struggle to catch all these issues because they rely on agentic graders that have the same blind spots.
But the token spend is significant, and it goes to three places: (1) hybrid synthetic + expert environment / task authoring, (2) rollouts to validate task complexity, and (3) QC on the data. You cannot produce a high-quality long-horizon agentic task without LLMs across all three buckets. We also use agents to automate internal processes, but that is a small fraction of the bill. Any data company telling you the bulk of its LLM spend is internal automation is lying.
Producing this data at high quality and reasonable cost comes down to two problems.
(1) the expert-agent interaction paradigm - how the expert creates, reviews, interprets, and audits the LLM's outputs (synthetic input files, prompts, rubrics, and so on). Solving this means knowing precisely where agents fail or underperform, designing the interface through which experts validate those outputs, and building the tools that let experts steer the agent going forward.
(2) optimizing every QC check for the right tradeoff between cost, quality, and latency. That means constructing complex and diverse eval sets, configuring an agentic system with the right models (often chained prompts, guardrail LLM calls, and deterministic checks), running the eval, then iterating on the harness, prompts, chaining, and model choice until the mix is optimal.
Both are now core skills of our SPLs.
Enterprise deployments follow the exact same pattern. Understand the workflow and where models fail or underperform, design the user-agent interaction to mitigate those failures, build evals that test the agentic system across the workflow, and optimize the system against those evals. Running data pipelines forces us through this loop hundreds of times a month across a wide range of project types, and we have to build product and infrastructure to do it fast and repeatably.
You cannot build an optimal enterprise deployment without these muscles, and the largest data companies have already done it hundreds of times.
ThomAub retweeted
releasing many high quality open-source RL environments is the most impactful thing anyone can do to push the open-source frontier right now
the equivalent of sharing high quality pretraining data but in the new RLVR paradigm
