Seen, Told, Hoped: Claude Goes to the Lab
Two Anthropic life-science leads talk for 35 minutes about teaching Claude to be a biologist. This film does not question whether they mean it. It asks whether you can check it. Every claim is sorted into three piles: seen (you can check it yourself), told (you take it on trust), and hoped (a plan with no result yet).
Seen, Told, Hoped: Claude Goes to the Lab
Two Anthropic life-science leads talk for 35 minutes about teaching Claude to be a biologist. This film does not question whether they mean it. It asks whether you can check it. Every claim is sorted into three piles: seen (you can check it yourself), told (you take it on trust), and hoped (a plan with no result yet).
What it finds: the connectors and one benchmark are checkable, with a footnote. The three-month-to-one-minute story is one memory with no hit rate. And the two big hopes, Claude running experiments and learning from the lab, are the same brakes Dario Amodei's own essay names: data and the speed of the physical world.
Sources checked: Anthropic, Claude for Life Sciences (20 Oct 2025); Anthropic, AI for Science program and its rules; Dario Amodei, Machines of Loving Grace (2024); the interview transcript.
Your turn: paste a claim an AI company made about science into Claude, sort each sentence into seen, told or hoped, and name the missing number.
Narrated by Liam, in for Bear.
youtube.com/@NikBearBrown
▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
[0:00] Ciao, this is Liam in for Bear. Two life science leads at Anthropic sat down for a long conversation about teaching Claude to be a biologist. One big promise runs through it, 100 years of science in 10. I'm reading it as a skeptic, not whether they mean it, whether you can check it. Here's the whole film in one breath. You arrive hearing that Claude is already a biologist. Correct that, it is still learning to be one. Sort what these two say into three piles, seen, told, hoped, then notice where the piles meet, the lab. Three piles. Seen, you can go and check it yourself, a product page, a
[0:32] Published number. Told, a speaker's report you have to take on trust, a story, a result with no method. Hope, a plan with no result yet. One sentence can sit in any pile, that's what we're sorting. Act one, what you can check. First, the plumbing and it is real. Claude connected to a literature database, a lab notebook, a single-cell analysis platform, a figure library, a shared data repository. Anthropic's own launch page lists them. Seen, unglamorous in the foundation. A model that can't reach your lab's tools can't help in your lab. Seen with an asterisk, on a test of understanding lab protocols, the new model scores .83, the
[1:08] Human baseline .79, its predecessor .74. Anthropic reports all three. The footnote says multiple-choice 10-shot prompting. How the human baseline was set isn't explained. Now read what that test is. Multiple-choice about protocols. It measures picking the right answer, not drafting the protocol, not running it, not noticing it failed, not fixing it. Reading is the first rung of a ladder. The benchmark photographs one rung. Act two, what you take on trust. The story, a founder's essay, is blocked by the sample. The lab spends three months fixing it. Years later he asks Claude, and in one minute it names the fix. A great story. Look at what it can't tell
[1:46] You, how many answers came before that one, and how many would have wasted a week. The fair defense is in the interview. In protocol work, you don't need perfect, you need unstuck. A colleague who says this looks familiar is enough. I agree up to a point. That standard works where a wrong answer costs one experiment. The same speaker calls Claude strong at regulatory submissions. Different price of wrong. Next told claim a major jump in long chains of tool calls, the long bioinformatics pipelines where one broken step ruins the run. No numbers given, not the length of the chain, not how often it finishes. They also say the
[2:19] Coding tool already helps biologists. It's called Claude code, not Claude biology. Another told claim papers are appearing that suggest you may not need specialist biology models. A big general model with the right training might get there. No papers are named. Note the logic. Specialist models are savants on DNA or proteins. The general model would reach them through language. It's a bet about where capability lives. Then the big one, their North Star is an essay by Anthropic's co-founder Dario Amodei, Machines of Loving Grace. In it he writes about giving us the next 50 to 100 years of biological progress in 5 to 10 years. In the interview that becomes
[2:56] 100 years in 10, about 10 times the pace. Keep the essay open. It says something else, too. Act three, what is only hoped. What the essay says next, three brakes on that speed up, data, the speed of the physical world, and intrinsic complexity. Experiments on cells and animals, he writes, are limited by the speed of the physical world. A smarter model doesn't make a cell divide faster. The speed up isn't one number. It's a number with three brakes against it. Now the hopes and listen for the brakes. Hope one, Claude runs the experiments. Talk the idea over, design the plan, draft the protocols, then say, "Run it. I'll
[3:30] Review the data in the morning." The first three exist in conversation today. The last two, run it and review in the morning, are the hope. Hope two, learn from the lab itself. Training on human annotations will saturate, they say, so the data has to come from experiments, a lab in the loop. An experiment throughput keeps climbing, a scaling law of its own. Good idea with a the No throughput number is given. And every turn of that loop is a physical experiment. Put the two hopes next to the essay's breaks, run the experiments, that is the speed of the physical world, learn from lab data, that is the data
[4:02] Break. The road map isn't separate from the skeptics list. It is the skeptics list stated as a plan, which is the most honest thing in the interview. Act four, who can check it? Now, who can check what? The connectors, the partners on their own pages, the benchmark, a footnote in a system card from the same company, the anecdote, one founder's memory, the road map, no result yet. Everything checkable in this conversation is checkable inside one company's walls. That's not an accusation, it's a map of where the checking would have to go. Two things in the conversation are institutions, not claims. A program giving free API credits to researchers
[4:37] With a bold idea, where every application goes through a biosecurity assessment, and Anthropic may keep and study interview transcripts. That's how they say they close the loop. The other is a stance, safety and speed don't pull against each other. The first you can read, the second you'd have to watch for years. They say plainly it's imperfect, that they want to know what isn't working, and that the real test is putting the pieces together. Write standards, so hold them to it. Ask for the numbers that would prove this wrong, the hit rate of that one-minute answer, the chain length and finish rate, how the baseline was set,
[5:07] Experiments per year. Let's recap with Claude. The conversation mixes three kinds of sentence. Seen, the connectors and a benchmark with a footnote told a story that took three months to one minute, a jump with no number. Hoped, labs that run themselves in a hundred years in ten. The hopes are the essay's own break stated as a plan. Ask for four numbers. Your turn, paste this into Claude. Here is a claim an AI company made about science, paste it. Sort every sentence into seen, told, or hoped. For each told sentence, name the number that would turn it into seen. For each hoped sentence, name the physical or data
[5:39] Limit it would have to pass. Then rank the three piles by what you could verify tonight. If everything landed in seen, look again at who reported it. Seen told Hope that Claude goes to the lab at Nick Bear Brown.
