The 80-90% Claim: Claude, Frontier Training, and a Number With No Denominator

We didn't get a paper. We didn't get a blog post. We didn't get a date, a link, an author, or a single line of quoted text.

What we got was two sentences on a crypto outlet, claiming Anthropic projects that Claude will lead 80-90% of frontier model training by 2027. That's it. That's the whole payload. My phone lit up at 3:41 a.m. Auckland time, and by 4:15 the number was in four group chats, two token pitch decks, and one Telegram channel that mostly posts charts of things that don't exist yet.

Here's what bothers me, and I say this as a guy who built a career on being first and occasionally wrong: I have published stories on thinner sourcing than this. In 2017 I ran a script against the Ethereum mainnet to flag whale movements, and when Vitalik's Demo at a San Francisco conference spiked ETH volume, my alert fired 14 minutes before the big outlets moved. I wrote 2,000 words on sharding implications at 5 a.m. with three encrypted developer chats and zero sleep. Speed was the product. I know exactly what this feels like from the inside.

Which is why I can tell you precisely what's wrong with it.

Context: how a research agenda became a percentage

Trace the provenance and it gets thinner fast. No original link. No publication date. No quoted executive. No named reporter doing original work. That's the fingerprint of a content-farm item — assembled from a press release, a podcast clip, and a model's summary of both, then shipped because the keyword "Claude" carries search volume. I've been on the wrong side of that trade. In January 2024 I ran a speculative ETF piece off an insider's tone of voice instead of the actual 19b-4 filings. Click-through doubled. Credibility didn't.

Anthropic's real public position is not a secret. Dario Amodei has spent years arguing that AI systems will increasingly do AI research — that the loop of better model to better research tool to better model is the whole ballgame. Jack Clark has said versions of the same thing with more caveats. This is a stated ambition, a research agenda, a thesis about direction. It is not a shipped capability, and it has never been expressed in public as "80-90% by 2027."

So somewhere between a real statement and a headline, a directional vision got compressed into a percentage with a deadline bolted on. That compression is the story. Not the number.

Now the useful part. What does frontier training actually look like right now? Pretraining runs that cost tens of millions. Data mixture decisions that are part science and part folklore. Architecture choices made by a handful of people who have done it before and watched runs die. And then the part nobody tweets about: loss spike triage at 3 a.m., eval suites that quietly stop measuring anything real, checkpoint rollbacks, and the slow paranoid work of figuring out whether your model got better or your benchmark got worse.

AI already touches chunks of this. Synthetic data generation. Hyperparameter sweeps. Boilerplate training code. Eval generation. Log summarization. Real, measurable, boring gains. That is the honest version of "AI is automating AI research" today: an engineering story about tooling, not an architecture story about takeoff.

Core: the claim has no denominator, and that's the tell

Claude is a model. It is not a training system. Read the sentence again and watch it break. Lead how? As the thing being trained? As the tool inside the pipeline? As the agent making the decisions? Those are three completely different claims with three completely different credibility levels, and the item never picks one.

That ambiguity is not sloppy writing. It's load-bearing. The vagueness is what lets a researcher read "we're building better tooling" and a trader read "recursive self-improvement is here."

The 80-90% — Root: The missing denominator. Eighty to ninety percent of what? Compute share? Wall-clock hours? Discrete task count? Weight of decisions? If it's task count, the claim is already almost true and deeply boring — an agent that files experiment tickets all day racks up volume fast. If it's decision authority, it's a claim about a system that can set its own research objectives, notice when a result is suspicious, and abandon a six-week line of work because the eval was lying. Those two readings sit an order of magnitude apart in difficulty, and only one of them is worth your attention.

Here's the frame I keep coming back to. Automated R&D is a control loop, and the eval harness is the oracle feed. The model isn't the bottleneck. The feedback signal is. If your agent can run a thousand experiments a day but your judgment on those experiments lags by a week, you haven't built a researcher — you've built a very fast machine for optimizing against a stale picture of the world. Latency in the oracle is latency in the entire system. Anyone who has watched a DeFi protocol get liquidated on a price that was true four seconds ago understands this in their bones.

I ran an OpenSea scraper in 2021 that ranked collections by hourly volume growth. It was fast. It was also completely blind to whether a contract was a honeypot, because volume measures attention, not safety. I published off that bot in 45 minutes flat and ate a backlash when a copycat scam slipped into the list. The lesson wasn't "don't publish." The lesson was that a metric without a denominator will always look like a discovery. "80-90%" is that bot, wearing a lab coat.

Strip the number away and look at what Claude can actually do, because that's the only honest input. Code generation and tool use: top tier, no argument. Long-horizon planning across days of autonomous work: second tier at best, and this is the part that matters. Autonomous experiment design: research-stage demos, not production. The gap between "writes the training script" and "decides which training script is worth running" is where the entire claim lives, and it is a gap measured in years.

There's a structural reason for that gap. Software agents got good fast because software has a verifier — the test suite. Generate a patch, run the tests, know in ninety seconds whether you were right. Research has no test suite. The verifier is a human staring at a loss curve at 2 a.m. deciding whether the dip is a bug or a discovery. Until someone builds an automated verifier for research quality, "AI-led training" means "AI-led execution with human-led judgment" — a much smaller claim than the headline sold.

Assume for a second it's half true. Assume 40% of training execution is automated by 2027. The first thing that breaks isn't safety. It's scheduling. Automated research means the experiment queue explodes — hundreds of small runs instead of a dozen big ones — and cluster utilization becomes the metric that decides who ships first. That's a procurement story, a power story, an interconnect story. It's also the one place where the crypto market and the AI labs are staring at the same object from opposite sides: both are now bottlenecked on scarce physical throughput.

Then there's the arithmetic problem nobody flagged. If OpenAI, Google DeepMind, and Anthropic are all chasing automated research — and all three have said so in public, in some form — then "80-90%" is a number that cannot be allocated. Three labs can't each take 85% of the same pie. What the figure actually does is position: it tells investors and recruits that Anthropic can see the finish line. It's a hiring tool and a fundraising tool wearing the costume of a forecast.

Regulators will force the question eventually. Reporting thresholds for high-risk capability models already exist in the US and are hardening in the EU. If a lab claims its model performs most of its own training, that's a capability claim with legal weight — and the disclosure regime is self-reported, which lands us right back on the same problem. A compliance framework that asks companies to grade their own homework isn't oversight. It's paperwork with a logo.

And this is where the crypto side gets ugly. Within hours of the item circulating, the usual basket of AI-agent tokens caught a bid. Not because anything changed in a repository. Because a headline about Claude made a narrative legible to people holding tokens with "agent" in the name. I've watched this movie. In 2021 I built a bot that turned volume into stories and it worked — traffic up 300% during yield farming, 50,000 subscribers in a week off a Bored Ape floor story. It worked because speed is a real edge. It also worked because nobody checked. The same mechanics now apply to AI research claims, except the stakes are considerably higher than a JPEG.

Contrarian: the distribution channel is the actual data point

Nobody is asking the obvious question. Why did this land in crypto media?

Not an AI publication. Not a research newsletter. A crypto outlet, which fed a crypto timeline, which fed a market that prices narratives in minutes. The audience that pays for "AI is about to eat itself" is not the audience that reads eval methodology. It's the audience holding tokens that need a reason to go up. The claim's real product isn't information. It's a narrative with a buyer.

In November 2022 I watched FTX vaporize $8 billion and I couldn't process the balance sheet, so I went to the parties in Dubai and London and read the room instead. The room was fine. The room was dancing. I wrote that the party wasn't over. I was wrong, and I was wrong in a way retail readers loved and institutional readers never forgot. The lesson I took wasn't "ignore sentiment." It was that sentiment tells you what people will believe, technical work tells you what's true, and those are different products sold to different customers. This Anthropic item is a sentiment product. It is engineered to be believed fast.

Which matters because of what got left out. Anthropic built an entire public safety architecture — Responsible Scaling Policy, AI Safety Levels, the whole ladder. The item mentions "robust infrastructure" and not one word about safety. A story about AI automating AI research with zero safety content is like a story about a bank merger with no mention of deposits. Self-reported safety tiers are the AI lab equivalent of project KYC — a self-attestation that looks like a control and functions as a formality. Nobody audits the form. Nobody asks who signed it.

And the competitive read is pure theater. "Whoever automates R&D first wins" is a race narrative engineered to manufacture urgency, not a measured forecast. The party doesn't end because a lab ships a roadmap. It ends when the roadmap gets an independent benchmark.

Takeaway: what to actually watch

Ignore the percentage. Watch three things. First, whether Anthropic publishes a methodology — what counts as a task, who measures it, on what cadence. No methodology, no claim. Second, eval latency: the gap between an experiment finishing and the field knowing whether it worked. That number is the real speed limit on automated research, and nobody publishes it. Third, any third-party measurement of human-versus-AI task share inside frontier training runs — the first genuinely falsifiable data point in this entire conversation.

The question I'd put to Anthropic directly: when Claude proposes the experiment, runs the experiment, and writes the eval that judges the experiment, who signs off on the result? If the answer is another model, say so out loud. If it's a human reviewing a dashboard, say that too.

The number will be forgotten by Friday. The architecture of who checks the checker will not be.