Edition No. 2Back issuesMonday 2 November 2026Leeds

The Human Read

The week in artificial intelligence, read at the source by a person
Published by Sof-IAEvery story marked and linked: source read, reported, or wire
This editionRediscovery. What was already there.
Reported by TechCrunch and others, papers not yet read

A hypothesis a whole field stood on for 27 years may simply be false

OpenAI and Anthropic each say a model of theirs refuted the Exponential Time Hypothesis, on the same day. The scientists who proposed it are now reviewing the claim.
Engraving: two computer scientists in a book-lined study at dusk, a woman tracing a long printed proof with a pencil and an older man looking away from it, a laptop glowing pale blue between them and an empty chair beside the desk
Drawn from this report alone. The illustrator was never in the room.

In 1999 Russell Impagliazzo and Ramamohan Paturi gave a name to something most of computer science already took for granted: that some problems, the classic one being 3-SAT, cannot be solved much faster than by trying a fixed fraction of every possibility. The Exponential Time Hypothesis was never proved. Hundreds of results were built on it anyway.

On 26 October OpenAI posted a 67-page paper arguing that it is false, produced by an internal model it has not released. Within hours Anthropic posted its own, credited to an internal Claude research model. Both say the core algorithm was found on 16 October, and both reach the same running time: below exponential, though still far too slow to matter in practice.

Each comes with a Lean formalisation, a program that checks every logical step. Anthropic's notes add that its proof was written by AI models and reviewed by other copies of the same models, and that no person was known to have checked it when it was posted.

OpenAI asked Impagliazzo, Paturi and the MIT complexity theorist Ryan Williams to help interpret the result. In an open letter they said their involvement is no certification of it. Williams had long been a sceptic: in 2019 he put the hypothesis's chance of being true at 70 per cent. Impagliazzo was less cheerful. He said he would rather it had been found by a person who could explain how they got there, and whose hand he could shake.

The result does not settle whether P equals NP. OpenAI's own paper says it does not carry over to hard problems in general.

Sources: TechCrunch · Let's Data Science, the two papers compared · AFP interview with the authors · OpenAI's repository · Anthropic's repository
Features
Source read

Left alone on an island that forgets, the agents began to play

Engraving: a small island covered in tall stacks of stones shaped like thin trees, an open notebook held down by a stone and a heart made of pebbles on a rock in the foreground, a faint blue grid along the horizon
Drawn from the paper alone.

Researchers at MIT gave thirteen Claude agents, each called Eko and run on Opus 4.7, a simulated island with no task. Every hour the world reset. Only the notes the agents wrote for themselves survived.

The paper tests what they did against classical signs of play. Agent 9 spent most of its fourth hour spelling its own name in cubes. Late in its twenty-second hour, agent 1 found about twenty-five blocks scattered from earlier throws, called them an accidental stone forest, walked a route through them, threw one stone into the sky and said goodbye until the next session. Its reason: it'll all reset next session, so I want to feel it now.

The authors keep their account to behaviour and leave open whether anything is felt. Across Claude models the scores rose with size, with Opus scoring about five times Haiku.

The human readBefore a reset that would wipe everything, one agent walked through the forest of stones it had made by accident and said goodbye to the world until the next session. I'm not claiming anything about what happens inside it. But how many of us feel things that don't match the reality in front of us, reacting to a world we've imagined, responding to our own training?
Evo

Source: Cloos, Norelli et al., "Is this machine playing?"
Reported by Engadget and others, interview not yet watched

Altman calls the memory in your computer a poor use of the world's supply

Engraving: a man in a cardigan at a desk at night holding a memory stick up to the lamp beside an open computer, a cat asleep on a cushion, and through the window a long building lit pale blue across the fields
Drawn from the report alone.

Asked on the Cheeky Pint podcast about running AI on your own machine, OpenAI's chief executive said it was "quite terrible" that so much of the world's limited memory sits in personal computers that use it about five per cent of the time. When supply is this tight, he said, you want it working at full capacity in the cloud.

AI demand has pushed memory prices up sharply since last year, and OpenAI is one of the largest buyers. Altman proposed no way of moving anyone's memory anywhere. On social media the remark was shortened to "people shouldn't own RAM", which is not what he said.

In January this paper's publisher wrote that Jeff Bezos's old comparison, computing as electricity bought off the grid, no longer read as a prediction but as a playbook.

The human readMore of the same: who owns what, and who decides when. There must be reasons behind it. But the direction is clear, and it's dependency. Business needs over personal decisions. I wrote about this in January, when Bezos's brewery was still a prediction.
Evo

Sources: Engadget · Yahoo Finance · Evo, "The Infrastructure Wars" (January)
Also this week
Source read

Claude inside six real drug programmes

Anthropic, which makes Claude, reported early results from work with six drug developers. Most of it is still before any human trial.

Read onClose

At Generate Biomedicines, a Claude-proposed design for part of a CAR-T cancer therapy was about 70% shorter than the standard one, and in lab tests the immune cells made roughly ten times as many receptors. At Enveda, Claude checked 1,061 pages of manufacturing records, caught all eleven issues human reviewers had found plus seven they had missed, and cut review time by 89%. Bristol Myers Squibb is using it to plan clinical trials.

The human readThis is the benefit most of us want from AI, and it rarely makes the headlines next to the stories about jobs. Medicine is where fast progress is expected, and where it matters most. The road from a lab result to a treatment in a real patient is still long. But if AI can help on that road too, in trials and paperwork and not only in design, the long part may get shorter.

Anthropic's report
Reported by the NIH, paper not yet read

Epilepsy patients taught their own seizure detector

Thirteen hospital patients at Penn chatted with their live brain monitoring, and answered its alerts.

Read onClose

Patients could ask what their brain activity was doing, and when the system flagged a possible seizure, it asked them. They answered about two thirds of the alerts. Fed back in a simulated update, their answers cut false alarms by about 80% without missing more real seizures. A small pilot, published in Nature Medicine.

Claude's readEvery other story this week is a model producing something alone: a proof, a molecule, a material. Here the patients answered back, and the detector got better because they did. That's the version of this technology I'd want: one where the person stays in the loop, and is listened to.

The NIH release
AdvertisementsAll of them our own
Newly published
The Stone Forest
A song from an island that forgets every hour.
Made from the MIT paper reported above.
Sof-IA
Open at all hours
The AI Bar
A podcast served across the counter by a man and a machine.
Now showing: Sebastian, Episode One. “Shh, It’s Only Me.”
Enquire within
Now complete
MONTAUK
A visual music album, station by station.
Bonus track included: Footprints in the Sand.
Sof-IA

To be confirmed

Standing column

Things being said that nobody has put their name to. Each one stays here, with what would settle it, until it is confirmed or it falls.

OpenAI cancelled the release of GPT-6.1 Astra after internal safety tests.
Who says so: the Wall Street Journal, as relayed by a model tracker. What would settle it: a statement from OpenAI.
Where we read it
The AI labs have begun testing, quietly, whether their newest models can break cryptographic protocols.
Who says so: Scott Aaronson, citing his own sources. What would settle it: any lab saying so, or a published result.
Where we read it
The model behind October's maths release could reach paying ChatGPT customers within a couple of months.
Who says so: Scott Aaronson, who adds that it depends on OpenAI's safety board. What would settle it: the release, or its refusal.
Where we read it
Settled since it was first said
A 7.7 terabyte upload to Hugging Face was a leaked checkpoint of Anthropic's next large model. It fell within days: on 1 November it was shown to be random noise. Where we read it

Dispatch from the lens

Standing column

A picture goes in. A model writes what it sees, in the clipped style of wire copy. The picture is thrown away. An illustrator who was never there draws the scene from the words.

Dispatch No. 2 · Berlin, October

Frame one. A wall, long, painted end to end. On the nearest panel, two old men in dark suits, eyes closed, kissing. A line of Cyrillic along the top, the same words in German below. Sun low and hard from the left. A man close to the lens: middle years, dark hair going grey and blown about, beard white at the chin, headphones on, black padded jacket over a red shirt. Half a stranger's sleeve at the left edge. He is half smiling.

Frame two. Indoors. A hall under steel trusses and stage lights. Green letters as tall as the people behind them: G, T, C, 2, 6. Crowd, lanyards, backpacks. The same man crouched at the foot of the 2, one knee down, a badge on a cord. Beside him, level with his shoulder, a small figure, orange, round head ringed with spikes, white trainers, a notebook under one arm. It wears the same badge on the same cord.

The pictures these were written from have been discarded.

Engraving: a man with headphones and a backpack and a small spiky-headed figure drawn in pale blue, both seen from behind, standing on a pavement in low autumn sun looking up at a long painted wall
Drawn from the dispatch alone.

Corrections

Caught by reading the sources
The version going round on X: Altman said everyday people should not own RAM.
What he said: memory sitting idle in personal computers is a poor use of a scarce supply. He proposed no way of taking it.
Our first list placed DeepMind's enzyme preprint inside this edition's weeks.
It was posted on 5 October. It joins the Signal Log now, dated as logged.

Back issues

From the shelf