Week 41 Dispatch - Le Chonk, Dying Declarations, and the Forgiving Ghost

Hearsay week. OpenAI dropped 722 math proofs from a model nobody outside the building can run, after the mathematicians it consulted asked it not to. Plus the dead man whose forgiveness got a sentence thrown out, two Western open models sold as the answer to China, and a 90% price cut that counts 30% more tokens.

10/05/26 – 10/11/26

The Open

On Tuesday OpenAI posted 722 math papers written by a model nobody outside the company can use. Some come with a proof a computer can check. Most don't. The mathematicians OpenAI recruited to advise it had asked for one thing a week earlier: don't do this with a model we can't question. OpenAI thanked them in the blog post and did it anyway.

Same week, an appeals court in Arizona threw out a prison sentence because the judge had watched a video of the dead victim forgiving his killer. The video was AI. The script was the victim's sister's. The judge said he loved it. The court's problem wasn't the technology, it was that the defense never got to ask the witness a question.

That's the whole week. Two Western labs launched "open" models with benchmark tables you can read and weights you can't download. Three fired safety researchers and their former employer told two stories that share no facts. Everything asked to be believed by someone with no way to check, and the only people who refused out loud were three judges in Phoenix.

Noise in. Here's what cleared.

Find your signal.
— BG

Before the Jump

Berlin, 1904. A horse named Clever Hans tapped out sums for crowds, and a thirteen-member commission found no trickery. Then a grad student named Oskar Pfungst asked Hans a question the questioner didn't know the answer to. Hans went from 89% right to 6%. He'd been reading the human. Worth a minute while 722 proofs arrive from a horse nobody is allowed to question.

Sonic Companion

Simon & Garfunkel, "The Boxer" (Spotify). Simon wrote it while the critics were working him over, and the second verse is this issue in one line: people hear what they came to hear and throw out the rest. A judge heard forgiveness in a script the sister wrote, and a lab heard "advances the field" from a committee that had just told it to stop.

Word of the Week

Hearsay: the actual evidence rule, why courts exclude it (nobody to question, not that it’s false), and the exceptions as the usable part. Closes on “which exception does each launch think it qualifies for,” which is the question a reader can carry into the week.

The Roundup

// Money. Claude Haiku 5.5 shipped Wednesday at $0.10 in, $0.50 out per million tokens, 90% below Haiku 4.5. Two asterisks: the 90% applies under 100,000 prompt tokens, and the new tokenizer counts about 30% more of them, so last month's 90K prompt now weighs 117K and lands in the tier that costs five times more. The price is a claim. The invoice is the proof.→ 90% off, 30% more tokens

// Platforms. Reflection's Beam (Monday) and Mistral's Le Chonk (Tuesday), both sold as the West's answer to Chinese open weights. Three years ago "open" was an American word and China was catching up. This week both Western launches were framed as catching China, a16z's Martin Casado puts 80% of open-tool developers on Chinese models, and Meta took its frontier closed. The sides switched and nobody announced it. Both models are previews, and the first weights you can actually download arrive from Mistral on October 27.→ DeepSeek of the West, weights pending

// Culture. The appeals panel's actual holding: conviction upheld, sentence vacated as fundamentally unfair, resentencing with the AI portion excluded. It doesn't ban AI victim material. It says a judge can't be moved by a witness the defense can't question, which is a rule that would sink most of what got announced this week.→ the witness nobody could cross-examine

// Talent. The three safety researchers OpenAI fired last week published their side Thursday: within their mandates, leadership in the loop, no leak, and if past conduct is now a firing offense nobody knows where the line is. OpenAI says "pattern of misconduct." Same week, David Robinson, who drafted the Preparedness Framework, quit and called the culture broken. There is no forum where either account gets tested, so the auditors the labs promised Washington will get the version the lab writes.→ shifting lines

// AI. The math drop, sourced: 722 manuscripts, 372 families, roughly 4,000 problems attempted, three hours of ChatGPT Pro compute per result, Lean for about 42% of top-line results by the README's own count, and a concession that the unformalized ones "could have issues." The IAS group's September 29 guidance asked for repositories no lab controls and a public tally of what the model tried and failed. Neither happened. By Wednesday the Association for Human Mathematics called it a show of power and told mathematicians to stop working with OpenAI. The Lean files are real and anyone can run them, which is more than most of this week offered. The PDFs without Lean are Clever Hans, right exactly as often as the only questioner in the room says they are.→ show the proof, keep the horse