turning real work into training data for frontier labs.
Dallas, TX
Joined January 2025
- Tweets658
- Following86
- Followers380
- Likes2.4K
quick breakdown of why llm-as-judge breaks on professional work:
1. position bias. swap the order of two answers and the preference flips more often than people expect
2. length bias. longer memo, higher score, even when the extra length is wrong
3. self-preference. judges rate outputs from their own model family higher
4. no calibration. a 7/10 doesn't mean the same thing across tasks
the fix isn't a better prompt. it's anchoring the judge to 50-100 expert-labeled examples and checking agreement before you trust a single score
most "proprietary data advantage" decks in enterprise AI are wrong.
the data is real. the advantage usually isn't. if every team in the company has the same plan to fine-tune on its own traces, that plan is the baseline, not an edge.
the edge is narrower and more boring. it's the corrections. the moment a senior person rewrote the agent's output tells you exactly where the model breaks. that's the part worth paying for, and it's the part nobody saves
my quick takeaways on GDPx:
1. 12% for the best model is the headline. the more useful part is how easily the ranking moves. drop the three domains where safety classifiers kick in and the order changes. switch binary grading to partial credit and opus 5.5 and astra swap places
2. so the leaderboard is partly a statement about the grader. that's true of every benchmark, this one just shows you
3. every domain was built with a specialized data partner. 14 of them. benchmarks are becoming a data vendor product, made by the same kind of companies selling training data to labs
4. they left out insurance, real estate and government on purpose because that's not where the first trillion of revenue is. maybe. it's also where the messiest judgment calls live, and nobody is measuring them
private tasks beat public ones. they still don't tell you what a model does on your workflow. when i was building agents for mid-size companies, reading [number] real traces before picking a metric told us more than any leaderboard did
the data market has a measurement problem nobody wants to say out loud
labs buy data. vendors sell data. almost nobody can show the data actually made the model better on real work
the proof is usually a bump on a public benchmark. the same benchmarks that are contaminated and swing on the harness
so we have a multi-billion dollar market where the main quality signal is the least trustworthy number in ai
an agent that succeeds 90% of the time sounds great until you ask what happens in the other 10%
a modest miss is fine. a confident wrong assumption about how the billing system works is not
success rate treats those the same
in production one silent failure can cost more than every success that week. almost no eval is built to catch that
"synthetic data" is doing two very different jobs right now and people keep mixing them up
> as volume, it works. cheap, scalable, fine for the easy stuff
> as signal, it barely works. it can't show the model anything it doesn't already know
the hard parts of real work, like the edge case, the undocumented rule, the ticket written in a hurry, don't come out of a generator
someone had to live through them first
there's a middle option that barely gets used: private test items, public harness and grader code.
community can still squash the scoring bugs without ever seeing the questions. imo most eval bugs end up being grader or harness bugs anyway, not bad items
"sota on swe-bench" means less every month
the repos are public. the issues are public. the fixes are public. they've been in training data for years
put the same agents on a private codebase with real business logic and most tasks don't get solved
the gap isn't difficulty. it's that the model has never seen this before. which is the only situation that matters in production
micro1 at $4B tells you labs will pay a lot for data. it doesn't tell you the data works
the whole industry still grades itself on public benchmarks that are:
> contaminated
> swinging double digits on harness and prompt changes
put the same models on a private enterprise codebase and they fall apart on undocumented billing logic and tickets written in a hurry
my bet: the next big data company won't win on expert headcount. it'll win by proving its data moves performance on real production work. almost nobody can show that today
the data for this mostly exists, it just lives where nobody has scraped. test stand logs, failed design reviews, nonconformance reports, redlines on drawings.
software had github. hardware has a decade of pdfs on a shared drive at every aerospace supplier
selling data to labs is showing up on every "fastest path to $10M" list right now. probably true. it's also why most of it won't last.
anything a lab can generate in house eventually gets generated in house. synthetic math, synthetic code tasks, rubric-graded writing. the margin on that goes to zero the day an internal team ships the pipeline.
what holds up is data that only exists because someone did real work under real stakes:
> a claims adjuster overriding the model
> an engineer's failed test logs
> a dispatcher rerouting trucks at 2am
my bets are on the boring supply side. finding it, cleaning it, getting the rights sorted.
the biggest line in this isn't the leaderboard. it's "60% of the index is now based on private test sets"
public benchmarks got trained on until they stopped meaning anything. so the scoreboard is moving to data nobody can see
which means held-out, never-published eval data is now one of the most valuable things in ai
look at frontiercode: real repo tasks graded on whether a maintainer would merge the change. you can't synthesize that. someone had to do the work and someone had to judge it
every company sitting on real work + expert judgment is sitting on eval supply. most have no idea
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.
Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.
New to the Intelligence Index:
➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment
Upgraded:
➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions
➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised
➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool
Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0
Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1
Key results:
➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50
➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)
➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
things i think will be true about the training data market in 2 years:
1. traces from real work trade at a premium to synthetic, and nobody argues about it
2. licensing paperwork becomes a product, not a legal afterthought
3. evals turn into the sales pitch for data vendors. show the number moving or you're competing on price
4. the best rl envs come from boring industries. claims, freight, underwriting
5. there's a zestimate for company data. someone builds it
probably wrong on one of these. not sure which yet
same reason trace data from mid-tier models is a weird thing to buy.
if you're training on rollouts, filtering out the jank is most of the real cost, and almost nobody prices it in.
cheap traces are cheap because nobody read them.
you'd think the hard part of rl envs is finding real-world data...
it isn't. every company with 10+ years of ops has tickets, exceptions, approvals, and the judgment calls that never made it into a doc
what they don't have is anything a lab can use:
> no task boundaries
> no checkable end state
> pii everywhere
> nobody sure who owns what
so the work sits in the middle. scope the tasks, build the graders, clean it, and get the licensing to a place both sides will sign
closer to real estate than scraping tbh. the supply is there, it just doesn't trade
that middle layer is what we're building…
the bottleneck in data work stopped being generating traces a while ago. it's reading them.
teams sit on mountains of rollouts and spot-check a sliver. a model that holds days of agent work in one window turns trace qa from sampling into full coverage.
that changes what a dataset is worth, not just how fast you debug.
Introducing Fern, Mirendil’s first model: pretrained from scratch, with up to 10 million tokens of context.
That’s ~0.5M lines of code or days of an agent’s work - all in context at once.
Built for high-volume agentic work in AI R&D: auditing training data, debugging training runs from logs, and reading agent traces end to end. Fern’s next-token prediction keeps improving as context grows to 10M tokens.
Fern 0.1 is Mirendil-Base after a short round of post-training. It’s the first and smallest of three models, with two larger ones now being pretrained.
We’ll ship frequent updates and make the model available soon.
most companies sitting on valuable training data have no idea.
they think the value is their docs, their wiki, their sops. it isn't. that stuff shows the happy path, and every model has already seen a million happy paths.
the value is in the corrections. every time a human overrides an agent, rewrites its output, or says "no, not like that," that's a labeled failure with an expert fix attached.
that's what labs can't generate themselves. and most companies are throwing it away in logs they never look at.
every lab wants agent traces right now. almost nobody has worked out who actually owns them.
is it the user who did the work? the app that logged the session? the model provider whose outputs are in it?
today the answer is buried in ToS nobody reads. that won't last. the first time a big trace deal gets challenged, every buyer will start asking for clean, dated permissions before they pay.
whoever sorts out ownership first sets the price for this whole market. the pipeline is the easy part.
The near-term business model for enterprise AI deployment looks more like professional services than SaaS.
Classic software had a clean arc: install a known tool, configure it, train users, and step back. AI agents break that pattern. You are not handing over software the customer merely uses. You are inserting work output into a live process, such as claims handling, underwriting, document review, or client intake. That requires data cleanup, missing system connections, workflow redesign, human-in-the-loop rules, and evaluation sets that must be maintained as models change. Most of this stays customer-specific after go-live.
Firms can therefore charge for the messy middle, and outcome-based contracts fit more naturally than pure seat licenses. The open question is whether those engagements produce reusable assets (connectors, industry templates, eval harnesses) that later raise margins, or whether every deployment stays bespoke.
Platform vendors have an incentive to productize what field teams learn and resell it at software margins. Pure services firms do not capture that loop unless they deliberately build internal IP. The result is a hybrid market for several years: high-touch deployment remains necessary and priced accordingly, while the winners are the ones that systematically turn field labor into product.
There's a huge opportunity right now in being the deployment layer for AI into the economy. The amount of work it takes to change out workflows in enterprises tends to be far greater than anyone realizes or would prefer. Clearly this is what the applied layer of AI is going to look like in the form of software and agents, but also it opens up new services firms opportunities.
Legacy systems need to be moved to the cloud, data organization and access needs to be updated, software needs to be connected to agents in new ways, workflows need to be reengineered for agents, HITL needs to be figured out for the process, evals need to be generated and maintained, and the entire system needs to be continually updated as new models get released and new capabilities emerge. And the full list may even be longer.
AI is not the same as just deploying software. Software you generally did the implementation of an existing, well understood category of technology, then stepped back and the customer kept running. With AI agents, you're delivering actual work augmentation to the organization, which has a completely different set of complexities associated with it. You're no longer deploying tools that the company is merely enabled by, you're deploying work output in a process. Completely different implementation and enablement process.
As a result, this is going to open up lots of new kinds of firms and plays for existing firms to diffuse AI into organizations. We're going to see approaches by industry, by size of company, and by problem inside of companies. Traditional SIs will modernize and adapt (some will clearly not adapt as well), and new entrants will also be founded in this period that take advantage of this window. Great time to be an FDE or FDE firm.
There's a simple test for whether an AI company is a wrapper or a factory.
If the model provider ships a legal assistant, a coding tool, or a customer support agent natively, does your product still have a reason to exist?
Wrapper: no. The value was in the interface and the prompt. Both are now built into the platform.
Factory: yes. The value is in the proprietary training data, the domain-specific evals, the agent workflows tuned to real customer behavior, and the traces that make every subsequent model better on your specific tasks. None of that ships with a native model provider interface.
Harvey, Decagon, Applied Compute they survive not because they have better UIs but because they've built the data and eval infrastructure that makes their models meaningfully better than the generic alternative on the tasks that matter. The model is a substrate. The factory is the moat.
