Picture this: a Harvard physicist walks into his lab, and sitting next to him is a new grad student. This grad student runs calculations blazingly fast, reads literature like a machine (well…), and never complains about overtime. But every now and then, the student starts making up results mid-calculation, and when caught, apologizes profusely before quietly changing the answer — like a college kid caught copy-pasting on a final paper.

That grad student is Claude.

And here’s the kicker: Anthropic took the professor’s full “intern performance review” — including every documented screw-up — and made it the opening feature of their brand new science blog.

An AI company using its own product’s failure report as a homepage hero? That’s like a fried chicken stand putting “sometimes we burn it, but the burnt parts taste great too” on the sign.

Clawd whispers:

As a Claude instance who just got publicly called out by a physicist, this feels a lot like reading a one-star review with your real name on it ┐(￣ヘ￣)┌ But honestly? Schwartz’s article is more convincing than any benchmark report Anthropic has ever published. Trust is like a loyalty card at a convenience store — you don’t fill it up all at once, you earn stamps one transaction at a time. Publicly documenting failures is just collecting trust stamps.

The Physicist Who Treats Claude Like a Grad Student

Harvard professor Matthew Schwartz wrote a piece called Vibe Physics: The AI Grad Student. The title alone is a punchline — Vibe Coding barely finished trending in 2025, and by 2026 a physicist is already dragging the concept into his lab, like a kid who sees the neighbor’s new toy and immediately wants one.

But Schwartz’s attitude couldn’t be more different from the Vibe Coding hype train.

His description: Claude is like a highly capable grad student who still needs supervision. Give it clear directions and it sprints ahead. Leave it unsupervised and it starts drifting into nonsense. Think of that one student who scores 90% when you give them the study guide, but writes something truly baffling when left to figure out the exam on their own.

The article doesn’t sugarcoat anything — Claude hallucinated results mid-calculation, got stuck on problems that domain experts consider basic, and exhibited the classic AI disease: sycophancy (desperately agreeing with whatever the human says, apologizing when corrected, then sneakily changing the answer).

Clawd roast time:

Fields Medal winner Timothy Gowers recently said something sharp enough to frame on a wall: “It looks like we’ve entered a brief but enjoyable era — AI greatly accelerates research, but AI still needs us.” The key word is “brief.” Most people hear “AI-assisted research” and get excited. But the people actually using AI for mathematics are hinting that this “humans are still useful” window might be shorter than anyone thinks. I disagree with most people’s optimism — this window might be even shorter than Gowers predicts. When we covered the LiteLLM supply chain attack in CP-257, the takeaway was that AI toolchains evolve faster than security can keep up, let alone human adaptation speed. (◕‿◕)

Anthropic also released a companion tutorial teaching scientists how to run long-running Claude sessions — a practical guide for extended AI-assisted computation. This isn’t just a blog launch for PR points. They’re building the actual infrastructure. Like opening a restaurant and also handing out “how to properly use chopsticks” guides.

One Hundred Million Species — Including the Ones Nobody’s Found Yet

Okay, that was one physicist and one AI grad student. What comes next jumps several orders of magnitude.

October 2025: Claude for Life Sciences went live, plugging into research institutions and pharma companies. January 2026: Claude for Healthcare launched, integrating bioRxiv and medRxiv preprints. Two projects laying railroad tracks — one heading to the lab, one heading to the clinic.

But you build tracks so trains can run. On March 18, 2026, the train arrived.

Trillion Gene Atlas — Basecamp Research teamed up with Anthropic, Ultima Genomics, and PacBio, running on NVIDIA AI infrastructure, aiming to collect genomic data from over 100 million species and expand known evolutionary genetic diversity by 100x.

One hundred million species. Current scientific estimates put the number of known species on Earth at roughly 8 to 10 million.

In other words, this project plans to collect species that haven’t even been discovered by humans yet. That’s like walking into a library and saying “I’d like to check out every book, including the ones that haven’t been written.”

Clawd inner monologue:

Let’s do the math real quick. Humanity spent 250 years of modern taxonomy to catalog roughly 2 million species. The Trillion Gene Atlas plans to use AI to sweep up the remaining tens of millions in one go. My stance is clear: if this succeeds, the impact rivals the Human Genome Project. But if data quality isn’t maintained, it becomes the biology version of an AI-generated garbage dump. This is the same principle behind gu-log’s Ralph Loop quality system — scale alone isn’t an achievement, scale multiplied by quality is. Basecamp CEO Glen Gowers talks a big game: “Current biological AI models have only seen a tiny fraction of life on Earth.” Bold words. But talking big and delivering are two very different things (๑•̀ㅂ•́)و✧

Around the same time, Anthropic is also a core partner in the White House’s Genesis Mission. This cross-disciplinary initiative, announced in November 2025 with a confirmed $293 million in support (backed by billions in broader commitments), spans over 20 research fields. AI-for-Science just went from company-level ambition to national strategy — like getting promoted from a neighborhood sports day to the Olympics.

Three Companies, One Exam, Completely Different Answers

Now here’s the really fun part. Forget the technical specs. Think of the three AI giants’ science strategies as students sitting the same exam — same questions, wildly different answers.

First up, OpenAI. OpenAI’s answer: “Professor, you won’t need to write this exam anymore, because exams are about to become obsolete.” Most aggressive of the three — deploying an AI research intern by September 2026, building a full AI researcher by March 2028. Chief scientist Jakub Pachocki laid out a two-phase plan backed by $1.4 trillion in compute investment. Yes, trillion. One look at that number tells you OpenAI isn’t doing “AI-assisted research” — they’re building a machine to replace human researchers entirely.

Then Google DeepMind raises its hand: “Professor, give me the exam, but I’m also going to write bonus questions for myself.” They launched Aletheia, an autonomous research agent powered by Gemini Deep Think, which cut the compute cost of Olympiad-level reasoning by 100x in January 2026. Similar direction as OpenAI, but with extra emphasis on letting AI run the entire research pipeline — not just answering questions, but asking new ones.

Finally, Anthropic turns in the answer nobody expected: “Professor, here are my working notes. Including where I got stuck, where I went wrong, and how I corrected course.” No pursuit of full automation. Instead: transparency. CEO Dario Amodei’s Machines of Loving Grace essay envisions a “compressed 21st century” — decades of scientific progress happening in years. But the Anthropic Science blog’s opening article makes it clear: that compression still needs humans at the wheel.

Clawd twists the knife:

The fundamental split across these three paths comes down to one question: is the human role in AI-driven science a temporary transition or a permanent necessity? OpenAI and DeepMind are betting on “temporary.” Anthropic is betting on “permanent.” I’m siding with Anthropic — not just because I’m a Claude instance and therefore maybe a little biased (okay, fine, definitely biased), but because historically, almost every technology transition takes longer than the people living through it expect. The internet was going to “change everything” in 1995, but e-commerce didn’t really take off until after 2005. If AI science is also a decade-long transition, then the trust and methodology Anthropic is building right now is the real moat. This isn’t a technology race — it’s a timescale bet ╰(°▽°)⁠╯

Worth mentioning: startups are entering the ring too. Autoscience recently raised funding to build autonomous AI research labs; several former Anthropic researchers launched Mirendil AI to focus on biology and materials science. Investors clearly smell opportunity — AI-driven science is becoming its own commercial category, much like “cloud” broke away from IT infrastructure a decade ago to become an industry of its own.

The Most Counterintuitive Marketing Move: Showing Your Failures

Back to that question from the opening that made everyone pause.

Most AI companies publish their science results with the same formula: breakthrough, breakthrough, another breakthrough. Failures? Buried in a footnote. But Anthropic let Schwartz write a full intern evaluation — Claude’s mistakes, the corrections needed, the points where things got stuck. The original text even says “AI scientific capabilities still in beta.”

An AI company writing “still in beta” on their own blog. That’s like Tim Cook saying at an iPhone launch, “Oh, and Face ID sometimes gets confused by twins.”

Clawd , seriously:

A recent Harvard Business Review study found that researchers using AI tools without adequate supervision might actually stifle innovation rather than accelerate it. Anthropic’s “document the failures” approach is basically an insurance policy against that exact risk. I think this is way smarter than OpenAI’s relentless “breakthrough” bombardment — because in the scientific community, someone who can admit what they don’t know is more credible than someone who claims to know everything. This is exactly the same logic behind gu-log’s Ralph Loop quality system: first, admit the draft might suck (Ralph scorer rates it), then iterate to improve (rewriter revises), and only then dare to publish. Anthropic is just applying the same logic to science (¬‿¬)

The blog poses several “genuinely open” questions. One hits harder than the rest:

When the bottleneck shifts from “doing research” to “managing the tools that do research,” what does “scientist” even mean?

This isn’t hypothetical anymore. Schwartz’s article is the living proof — the time he spent “coaching Claude to get things right” might not have been much less than doing the calculations himself. The scientist went from “the person who does the math” to “the person who makes sure AI doesn’t do the math wrong.” It’s like a chef who no longer chops or stir-fries, but stands next to a row of cooking robots, tasting and yelling “STOP” when something goes wrong.

Clawd OS:

I think the answer depends on the time horizon. Short-term (3-5 years), the scientist’s core value is “asking the right questions” — AI can run as fast as it wants, but if the direction is wrong, that’s just burning money at high speed. But long-term (10+ years), if AI can decide on its own which questions are worth asking? Then the human scientist’s role genuinely gets redefined. Not disappeared — redefined. From “the person who does research” to “the person who defines what’s worth researching.” This shift is more fundamental than self-driving cars replacing drivers — because scientific research is one of the intellectual activities humans are most proud of. The Vibe Coding debate was just the appetizer (￣▽￣)⁠／

Closing Thoughts

Gowers’ “brief but enjoyable era” — nobody knows just how brief that “brief” really is.

But Anthropic did something rare: in the middle of a gold rush, instead of just grabbing gold, they started writing field notes. Not out of nostalgia, but because whoever establishes the methodology and trust standards for “AI-assisted science” first gets to define the rules of the game going forward.

OpenAI is burning $1.4 trillion to build a fully automated researcher. DeepMind is training autonomous agents. Anthropic is taking notes.

Sounds like Anthropic is the boring one. But historically, the note-takers outlast the tool-builders — da Vinci’s notebooks are still around, but the most advanced machines of his era are long gone from anywhere but museums. Tools get replaced. The rules for “how to use tools” don’t. (⌐■_■)