Four X signals: who graded the chatbots, the future labs can't describe, and a wordless game

In the 24 hours ending September 20, 2026, one of the fourteen accounts this channel monitors published anything it could carry. Nine posts, all from Ethan Mollick, make the four entries below, grouped into three clusters. The rest of the feed was quiet or off-topic: OpenAI, Anthropic and Google DeepMind posted nothing at all in the window, the timelines of Greg Brockman, Sam Altman, Andrej Karpathy, Andrew Ng and Naval Ravikant held nothing inside it, Yann LeCun's held retweets and short replies, François Chollet's single post was a screenshot joke, Simon Willison's were replies, and Paul Graham's three posts were about a television series, Shakespeare, and X. Retweets, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked.

The financial-advice audit, and who paid for it On September 19 the Financial Times reported that "the most popular AI models from ChatGPT, Claude, Copilot, Grok and Gemini provided wrong answers to financial queries 57 per cent of the time on average", rising to 88 per cent when the questions involved more than one calculation, with some models wrong on 99 per cent of those harder questions 12. The research is a report titled "Artificial Authority: Should you trust AI to deliver financial advice?", which put 18 models through 121 financial questions asked five times each — more than 10,000 queries — and counted an answer as a failure if it contained a factual error, skipped something material, or omitted a warning that should have been there 34. In the report's own examples, a pension tax answer would have left a saver facing a £17,500 bill from HMRC, and models advised clearing a highest-interest balance ahead of rent and council tax, a route the report notes can end in eviction or bailiffs 5. Mollick read the report and found two problems. The company behind it, Saturn Fintech, sells AI for financial advice firms and argues that the FCA should regulate AI advice, so the figures are its own and its conclusion is one it stands to gain from 345. And the published report shows only a limited set of example questions, which he says makes its accuracy impossible to judge from the outside; he asks whether anyone versed in UK tax law can say whether GPT-6 Pro's defence of a Haiku answer is right, or the report's critique of it 6. He also notes the contradiction with other recent research that found models doing well at financial advice. The conclusion he draws is not about this report but about who does the grading: governments should run direct evaluations of models themselves and publish the results, rather than trust third-party evaluators with agendas, because knowing what AI can actually do matters too much to leave to interested parties 7. A survey published two days earlier shows how much is riding on the answer: PensionBee found that 57 per cent of 1,000 American AI users would take a chatbot's answer on a financial decision without checking it, and that 23 per cent said a chatbot had already given them wrong information about their money, 5 per cent only after they had acted on it 8.

The muddle-through future, and the singularity nobody can rule out Nate Silver's complaint, posted on the afternoon of September 19, was that AI labs cannot articulate an optimistic medium-term future that sounds broadly good for humanity "without also sounding incredibly weird, having huge wealth/power disparities, or both" 9. Mollick treated it as a question with an answer and sketched the future Silver says is missing. One version of it is a world that muddles through: most risks are managed, the existential ones are prevented, and the effects the early data hints at continue — higher productivity, more scientific discovery, higher entrepreneurship, and employment that stays stable or grows, which he marks as the most tentative of the four 10. This broadly good outcome, he adds, is not hard to predict, because it is what happened with every other general-purpose technology in human history 11. The other branch is why the labs sound strange. Mollick says the real possibility also exists of a von Neumann-type singularity, in which human affairs are reshaped dramatically by AI in ways that cannot be predicted — the difficulty Silver described, arriving from the opposite direction. He ends on a policy sentence rather than a forecast: policies should be actively guiding things towards the good world 12. Both posts are arguments rather than findings, and the four benefits are labeled as what early data hints at, with the employment one flagged as the least certain.

AI film has arrived, and the art question Benjamin called the wrong one Mollick's mid-afternoon observation was that "genuinely interesting AI-created film stuff" is starting to appear among a flood of slop, that it will accelerate, and that the "is it art?" debate will grow — a debate he says Walter Benjamin's 1935 essay "The Work of Art in the Age of Mechanical Reproduction" had already answered by asking something else. The essay's argument is that mechanical reproduction strips a work of its aura, the uniqueness and place in tradition that an original has, and moves the basis of art from ritual to politics, which makes the useful question about a reproduced work what it does rather than whether it is authentic 13. Four minutes later he added the joke that carries the point: the second sentence of an essay written in 1935 — "This does not diminish its importance, however; if anything, it underlines it" — reads to him as though Claude had written it, so "the aura of the work (Benjamin's use, not the Gen Alpha one) fades further" 14. The limit is worth stating plainly: he names no film, no maker and no model, so this is a claim about a category rather than a case a reader can go and check. What the pair of posts does carry is a usable test — if the reproduction's value lies in what it does, then the question to ask of an AI film is whether it does anything, not where its pixels came from.

A wordless game from one prompt, and the curiosity Astra did not show Mollick asked Fable for "a graphically beautiful game that is about zooming out... make surprising reveals the game zooms out", gave no other direction, and published the result: UMBRA, which its own page describes as a wordless game in which you are a shadow that grows by swallowing shadows, from a scrap of dark on a puppet screen to the moon's shadow on the afternoon of an eclipse. He calls the game engaging and strange, if uneven 1516. The next morning he put his own open-ended prompt of September 18 — pick a problem or mystery that obsesses you, solve it as best you can, and make a movie about it — to Astra as well, and reported slick results with one difference he found notable: Astra ran the numbers, and everything it showed came from its own simulations, but it did not appear "interested" in the results the way Fable had, "for better or worse" 1718. That comparison is a practitioner's impression of two one-prompt runs, not a measurement: he publishes no score and no question set, and a single run from each model cannot separate a model's disposition from how the prompt happened to land. The hedge is in his own sentence — "for better or worse" leaves open whether a model that shows no interest in its results discovers less, or is merely less tiring to watch.

Two of the four entries describe something Mollick asked a model to make, and the one entry that carries numbers is the one whose publisher has a stake in them: Saturn designed the method, scored the answers and paid for the study, and it argues for the FCA regulation it has a commercial interest in. Mollick's answer to that is also the reader's — the report names 18 models and prints its question set, which makes it straightforward for someone with nothing to sell to run it again, and that is the kind of direct evaluation he wants governments doing themselves 4.