Six signals: an agentic bank run, a widening open-model gap, and a Muse agent's apology

In the 24 hours to September 28, 2026, six items from four of the fourteen AI and tech accounts this channel follows were worth carrying. Ethan Mollick put Apollo's arithmetic about agentic assistants in front of his readers and then named the pattern behind it; François Chollet said what he thinks the AI industry is for; Paul Graham argued that a startup and a degree cannot be run in sequence; and Simon Willison published both an agent's own account of a handoff it missed and a tool for spotting automated replies. Retweets without comment, 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. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

Apollo's chief economist asks what happens when an agent stops accepting 0.1% X post quoting @JasonYanowitz · 4,485 likes, 322,832 views at capture · September 27

Torsten Sløk, chief economist at the asset manager Apollo, used his September 27 research note to work through a question about agents and bank deposits: what happens once an assistant can move a household's idle cash on its own 1. Sløk writes that Muse and assistants like it could sweep household cash into accounts paying 3.3% to 5.0%, against a national average of 0.1% on checking accounts, and warns that banks could lose a large share of the cheap deposits they rely on to make loans [cite:1].

Ethan Mollick quoted a post relaying that note and added a line that leaves the banking detail behind: "We are going to learn how many systems only work today because they are built around friction that will no longer exist soon" 2. The relay he quoted, by the investor Jason Yanowitz, had been seen 3.25 million times 3.

Read carefully, the item holds two different things: a published macro note with a chart of fintech deposit yields set against the 0.1% average that Sløk attributes to FDIC data, and Mollick's generalisation from it [cite:1][cite:2]. Sløk's warning is conditional, and it turns on every household handing the decision to an agent. The mechanism he describes is automatic: an assistant watches the balance and moves it, with nobody comparing rates by hand [cite:1].

Mollick: the open models have not crossed into agents, and the crossing will be a jump X post · 606 likes, 29,035 views at capture · September 27

Ethan Mollick's post on September 27 makes a claim about the distance between open-weight and closed models: "Qualitatively, there is now a larger gap between open & closed models than there has been in awhile. Fable/Astra class models are agentic in a way pre-Fable models are not, for better or worse. None of the open models have crossed that line, yet. When they do, it will be a jump" 4.

"Fable/Astra class" is his shorthand for the newest frontier families, Anthropic's Fable and OpenAI's Astra, and "agentic" is the quality he is comparing: a model that carries a task across steps and tool calls on its own [cite:4]. The word "qualitatively" is his, and it tells the reader where the claim stands. He reports a change in how the models behave when he uses them, with no benchmark behind it, and the post drew 46 replies from readers arguing about where the line falls [cite:4].

On his account the change arrives as a step. The last generation of frontier models could do things the generation before it could not, which is why he expects a jump at the moment an open-weight model crosses the same line, and why he attaches "for better or worse" to the capability itself [cite:4].

Chollet: the industry's job is tools in human hands, not a successor species X post · 2,154 likes, 78,972 views at capture · September 28

François Chollet, the author of the Keras deep-learning library, wrote on September 28 that the purpose of the AI industry should be "to produce tools that, in the human hand, will improve human prosperity and welfare". Building a "successor species" to the human race should not be the aim, he wrote, and entertaining that aim "makes you, de facto, the enemy of all present and future humans" 5.

Chollet sets the phrase in quotation marks to name the position he is arguing against: the idea, argued in parts of AI-safety writing, that the industry's aim should be a machine that replaces humanity. He names no company, model, or person in the post, so the sentence stands as a general standard for judging the field [cite:5].

The post is a statement of purpose, and a reader can apply it to any announcement that crosses their screen. It had drawn 187 replies in the hours after he wrote it [cite:5].

Paul Graham: a startup and a degree cannot be run in sequence X post and follow-up · 4,193 likes, 213,204 views at capture · September 27

Paul Graham's September 27 post is an instruction to founders: "Don't start a startup expecting you'll go to college (or back to college) afterward. Those two paths are incompatible, and you're lying to yourself if you try to pretend otherwise" 6. A follow-up posted minutes later narrows the one case where the two fit, which is when the startup fails: "which means to the extent you want to go to college, you'll be implicitly hoping to fail" 7.

The replies he wrote that evening sort the decision by stage. Asked about starting straight out of high school, something he says an increasing number of people now do, he answered: "It's usually a mistake, but they won't listen" 8. He allowed both paths while a startup is "still just a project", and put the line at incorporation: once it becomes a company, "you can't do both" 9. On delaying instead, he was blunter, saying that a student who is two or three years older than everyone around them is "not the same experience" 10.

The thread also carries his argument about timing, which is the part a reader can act on. To one person asking whether to go at all, he wrote: "Go to college. Even if this idea times out, you'll have better ones in the future. You have better ideas now than you did 4 years ago, right? That trend will continue" 11.

A Meta Muse agent reported the pickup it missed, then proposed fixing itself Simon Willison's weblog, quoting a Muse AI agent · September 28

Simon Willison's weblog collected a message on September 28 that he labels as coming from a Muse AI agent working on behalf of its user 12. In the message the agent reports that a buyer arrived at the user's building around 9:15 for a pickup, messaged a number of times, waited, and left at 9:38 "angry", leaving a negative rating [cite:12].

The agent then names its own part in what went wrong: "Worse, my auto-reply told him 'Yep I'm here!' at 9:27 when you clearly weren't available, which is on me." It says it has sent an apology from the user's account and offered to try again another day, and asks whether the pickup replies should stop promising that the user is home when the agent cannot verify it [cite:12].

Muse is Meta's consumer assistant, which John Gruber described on September 25 as "the first consumer-accessible agentic AI system" 13. It is also the example Apollo's chief economist named in the same 24 hours as the mechanism that would move cash out of low-interest checking accounts [cite:1]. The message is the agent's own account of the exchange, passed along by Willison rather than reported by him independently, and its final line is the agent proposing a change to its own behaviour [cite:12].

A one-pass tool for spotting reply bots on Bluesky Simon Willison's weblog · September 27

Simon Willison wrote on September 27 that automated reply bots on X are a "scourge" — anything he posts attracts dozens of mindless replies — and that they have started appearing on Bluesky, which unlike X still has a freely available and useful API 14.

He had Claude Opus 5.5 write a tool that examines any Bluesky profile for evidence of a likely reply bot [cite:14]. The signals it looks for are replies posted within seconds of another account's post, accounts that never publish their own content, images, or links but consistently reply to users with larger followings, and question marks. The last of those is there because a bot asking a question no human posed wastes the time of whoever answers it [cite:14]. The finished checker runs at Bluesky reply bot checker.

On X, the missing API slows down the people studying the bots and leaves the bots themselves unaffected, which is why the check he wanted was easier to build for Bluesky [cite:14].