Anthropic showed its work, OpenAI shipped a proof no human can read

Two AI labs made big claims this month. Anthropic says its AI found something new in biology. OpenAI says its model cracked a famous math problem and 100 more. Both are real. Both landed with people who did not trust them. If you are building anything where AI does research, the gap between "we found it" and "we can prove it to you" just became your problem. Let me catch you up.

The deep cut

• A discovery is not the same as trust in it. Anthropic and OpenAI both shipped results the field could not yet check.
• Share early, but hand the community the receipts. Anthropic got Feng Zhang to call the ART finding "genuinely intriguing."
• A proof no human can read is not a product. OpenAI's Navier-Stokes paper, per James Maynard, teaches almost nothing.

What Claude found in the DNA

Anthropic's new Bay Area wet lab pointed Claude at a database of DNA sequences and told it to hunt for interesting reverse transcriptases. Over 21 hours, roughly 950 agents burned through 210 million tokens before one flagged an odd repeating pattern next to an unusual gene. The agent's own words: "that's a CRISPR-like repeat array?!" Human scientists confirmed it in the lab. They named it ART.

The framing was careful. Anthropic said the humans wrote the prompt and did the bench work, and that all lab work stays at low biosafety levels with no human pathogens. CEO Dario Amodei even credited a Stanford team for earlier related work, though the correct link is Julie Bort's TechCrunch writeup.

How Anthropic bought itself credibility

The smart move was not the discovery. It was the packaging. Anthropic released a pre-print, showed its method, and put a real name behind it. Feng Zhang, one of the pioneers of CRISPR, called the finding "genuinely intriguing and merits further investigation." That is not a rave. It is exactly the calibrated blessing a lab wants when it says "we do not yet know what this does."

Even friendly coverage stayed cautious. The Verge noted it "remains unclear whether the discovery will have any practical applications." Anthropic did not fight that. It handed over the receipts and let the field judge. That is the posture to copy.

OpenAI's proof nobody can read

OpenAI ran the opposite play. It aimed 10,000 agents at the Navier-Stokes equations for 88 hours and announced a solution to a million-dollar Millennium Problem. Mathematicians agree the proof is technically valid. They also cannot use it. "The paper is not written for humans," Brown's Javier Gómez-Serrano told Futurism. Oxford's James Maynard added that "as of today, the paper doesn't teach us much."

It got worse from there. NYU's Tristan Buckmaster accused OpenAI of pushing his own unpublished ideas over the finish line, and said the company offered to credit him only if he dropped his Anthropic-based collaborator. OpenAI denied it, but admitted it "cannot rule out" that data from his use of its tools helped train the model. A correct answer with a theft cloud over it is not a win.

Bolting on a panel after the fact

Faced with the fallout, OpenAI formed an advisory group of nine elite mathematicians at Princeton's Institute for Advanced Study. The group can speak publicly and offer unsolicited advice. It cannot slow OpenAI down. The blog post says plainly the group "will not be responsible for advising us on how to pace our internal progress."

Mathematicians were not sold. Heriot-Watt's Francesco Fournier-Facio called it "an ivory tower." Imperial's Kevin Buzzard asked why a committee of geniuses was even needed: "It wasn't proofs of hard theorems, it was better understanding of our subject." A panel formed after the crisis reads as PR, not process. Trust is earned before you ship, not staffed afterward.

Three questions for your team

• When our AI produces a result we cannot fully explain, do we ship the claim or the receipts? Anthropic shipped a pre-print; OpenAI shipped a proof no human could read.
• Who is our Feng Zhang? Name the outside expert who will vouch for our work before launch, not the committee we assemble after.
• If a user's data touches our model, can we prove whose idea an output was? OpenAI could not rule it out, and that admission became the story.