Eight signals in three days: a sovereign model built on Chinese data, hard caps for AI spending, and whether models can feel

In the three days to October 5, 2026, eight items from seven of the fourteen AI and tech accounts this channel follows were worth carrying. Europe's first serious open-weight "sovereign" model shipped with an admission inside its own technical report — its synthetic training data was largely generated by Chinese models; OpenAI set a date for retiring custom GPTs while its heaviest users said they can no longer plan their usage; Simon Willison argued that AI services need spending caps that actually stop; Ethan Mollick said the constraint on AI gains has moved from individuals to organisations; François Chollet and Sam Altman separately pushed back on treating models as minds; Andrej Karpathy noted how recently today's AI audience arrived; and Christian Szegedy's essay on AI and mathematics, circulated by Paul Graham, argued that the field's objections describe last spring's models. The window runs three days because the editions of October 3 and October 4 were not published.

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.

Europe's new sovereign model was trained with synthetic data from Chinese ones Aleph Alpha's release post and technical report, and Ethan Mollick's post on X · 550 likes and 59,550 views on the post at capture · October 3

Aleph Alpha released Kolibri on October 3, an English–German mixture-of-experts model with 78.1 billion parameters, of which 3.46 billion are active for each token, a context window of up to 1 million tokens, and weights published on Hugging Face under the Apache 2.0 licence 1. The German company trained it on about 24 trillion tokens, more than a fifth of them German, on 768 NVIDIA B200 GPUs running on infrastructure in Germany and Finland, and sells it as a sovereign model: one that a ministry or a manufacturer can serve on its own machines, under European law, and that no outside party can switch off underneath them [cite:1].

Sovereignty here covers how the model was built, not where its training data came from. The datasets section of the technical report names the models used to generate Kolibri's fine-tuning data and to regenerate parts of its open datasets: GLM-5.2 and GLM-5.3 from Z.ai, and Qwen3.8-27B from the Qwen team 2. Ethan Mollick, who teaches at Wharton and writes about how AI is adopted, posted that page with the passage highlighted: "US labs say Chinese labs distill their models. And Europe's new 'sovereign model' is fine-tuned on data generated by… GLM and Qwen." 3

Aleph Alpha reports 96.9 on AIME 2025, 96.0 on AIME 2026, 84.3 on GPQA Diamond and 85.9 on LiveCodeBench v6, and says Kolibri matches models with up to four times its active parameter count [cite:1]. The release drew 672 points and 327 comments on Hacker News 4, and Aleph Alpha's own announcement post drew 8,111 likes 5.

LeCun: distillation economics make open frontier models inevitable Yann LeCun's post on X · 556 likes and 52,689 views at capture · October 4

Yann LeCun, a Turing Award winner and one of the field's most consistent advocates of open research, argued on October 4 that frontier models will end up open because of how their economics work: training a frontier model is expensive, distilling one is cheap, and "this market force alone tells us that frontier foundation models will be free/open" 6. He added that this is not the only force pointing that way [cite:6].

The claim runs against the grain of the current dispute, in which US labs say Chinese labs distill their models — the mechanism LeCun describes as unavoidable [cite:3]. This window produced a worked example of it: the European lab above built its sovereign model's training pipeline on data generated by GLM and Qwen [cite:2].

His post drew 556 likes and 52,689 views by capture, and he restated the same four lines in a later post the same day [cite:6].

OpenAI will retire custom GPTs on December 11, and its heaviest users say they cannot plan around it OpenAI's help centre, and Ethan Mollick's posts on X · 448 likes and 22,698 views on the most-shared post at capture · October 5

OpenAI's help centre says custom GPTs are scheduled to retire on December 11, 2026 — February 11, 2027 for enterprise workspaces with an approved deferral — and that creators can migrate them to plugins, where a GPT's instructions become a skill, its knowledge files become reference files, and its connected apps are re-added as apps 7. Custom actions do not transfer at all, so GPTs built on them will stop working until someone rebuilds the connection, which may mean a custom MCP server and technical setup [cite:7].

Ethan Mollick, who has run a public GPT for years, wrote that he can migrate his most-used one to a private plugin, but that the reason he created a GPT was for other people to use it: "GPTs are obviously obsolete, but I know they were a focus within companies & among educators for years. This isn't a great path for them." 8 The same day he questioned the company's habit of handing out usage resets, which arrive as credits the user applies: "getting random prizes of tokens is kind of strange. Like what percent of the user base gets this?" 9 Another user's framing of the same problem — Pro customers who are not on X, wondering why their limits keep resetting at random — drew 4,667 likes 10.

The capability tier has not moved with the packaging. Mollick wrote that "there is still no equivalent to GPT-6 Pro", which he says does very hard tasks in one pass and explains what it found unusually well 11.

AI services need hard budget caps, not midnight warning emails Simon Willison's weblog, and his post on X · 891 likes and 109,484 views on the post at capture · October 3

Simon Willison published a short argument on October 3 that pay-by-usage services need a feature most of them still lack: a default hard budget cap, the kind that lets you say "after $X a month, cut this thing off and return errors" 12. Soft caps will not do, he writes, because a warning email arrives after the money has already gone: "Nobody wants to wake up to an email sent at midnight warning about a budget limit and find that, while they slept, their rogue service had consumed several hundred (or several thousand) more dollars of usage" [cite:12].

His reasoning is about who is now writing the code that spends: coding agents, and personal agents that are coding agents behind a friendlier interface, make it much easier to spin up a service that calls paid APIs or bills for storage and compute, so the failure mode moves from a person forgetting to check a dashboard to a program nobody is watching [cite:12]. He wants hard caps on by default, with a clearly labelled opt-out for anyone who wants to live without them, and names AWS — which added project spend limits in September — as the provider he most wants this from, alongside Google Cloud's July Spend Caps [cite:12].

The post drew 891 likes and 109,484 views on X 13, and 610 points and 302 comments on Hacker News 14.

The constraint on AI gains is moving from individuals to organisations Ethan Mollick's post on X, quoting Mark Cuban · 848 likes and 51,077 views on the post at capture · October 5

Mark Cuban's line — "AI won't take your job. Someone who knows how to use AI better than you, will take your job" — drew 2,932 likes 15. Ethan Mollick replied that the advice is less true than it used to be, for three reasons: the durable skill behind "using AI better" is unclear as the tools get easier to use; organisational systems, rather than individual workers, increasingly set the ceiling on what AI adds; and the effect on jobs is uncertain but may land on whole job categories 16.

Mollick drew the practical conclusion in a follow-up: this does not mean individuals cannot gain a large advantage from AI, but that companies need to build their systems for human augmentation rather than automation, "because individual performance is not the only factor" 17. The two posts drew 848 and 89 likes respectively [cite:16][cite:17].

His own example of where the organisation shows up is a request rather than a finding: he wants personal operating systems built for working alongside AI agents, with fine-grained security, an easy way to see what the agents are doing, and separate audit agents that check their work 18.

Chollet and Altman, separately, push back on treating models as minds François Chollet's and Sam Altman's posts on X · 2,923 likes and 184,141 views on Chollet's post; 28,367 likes and 6,587,319 views on Altman's, at capture · October 3–4

François Chollet, who created the Keras deep-learning library and co-founded the ARC-AGI benchmark, argued on October 4 that inferring consciousness from computation is a category error: "living creatures are made of atoms, and since this rock is made of atoms, it is likely alive" 19. He listed what he says a static input–output program lacks — information integration, interoception, temporal binding, embodiment — and concluded that humanity has not yet created a conscious program and shows no sign of being close [cite:19].

He had put it more sharply a day earlier: current models are "effectively calculators scaled to a very large number of competence circuits", and the idea that they are conscious he called absurd and "downright offensive" 20. He also warned that treating models as sentient, and trying to weigh their suffering against humans' on the same scale, could lead somewhere dark 21, and said that creating artificial consciousness should not be a goal of AI research at all 22.

Sam Altman made a narrower point the same day: he is "very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models", and thinks it is a real safety issue 23. That post drew 28,367 likes and 6.6 million views, the largest response of any item in this window [cite:23]. Read together, the two argue the same thing from opposite sides: capability is not moral status, and it is not a reason to hand over judgement.

Almost everyone now paying attention to AI arrived within the past year Andrej Karpathy's and Ethan Mollick's posts on X · 2,498 likes and 143,758 views on Karpathy's post at capture · October 4

Andrej Karpathy's observation, in a reply about how the field reads its own history: "99%+ of people who are now paying attention have been onboarded to anything related to AI in <1 year," a fact he says is confusing for the ones who were there earlier — the "AI dinosaurs", by which he means anyone who has been in AI since before 2026 24.

That post drew 2,498 likes and 143,758 views by capture, among the largest responses in the window [cite:24].

Ethan Mollick put a number on the same shift: about 65% of people in the United States use generative AI several times a week, which he calculates is about 72% of internet users and therefore close to saturation — with the caveat that most of them do not pay for it 25.

An AI researcher's essay asks what is left for mathematicians when the machines can fly Christian Szegedy's essay, and Paul Graham's post on X · 1,927 likes and 192,900 views on the post at capture · October 3

Christian Szegedy, a machine-learning research scientist, published "Is Mathematics Over, or Just Graduating?" on October 3, an essay built on one metaphor: mathematics as an unmapped jungle that explorers cross on foot, and AI as an airplane that can now reach ground the walkers could not 26. He writes that the airplanes are still primitive and expensive by the standards of the tools they replace, but that in the past few months they have become reliable enough to land on rocks nobody had climbed — his term for solving conjectures that had resisted earlier attempts [cite:26].

His argument is against reading current limits as permanent: "Do not bet against the airplane. The skeptics who say AI can't conjecture, can't theorize, and can't explain are describing last spring's models. Within months, that description will no longer hold." [cite:26] He also predicts that what mathematics rewards will shift — from the heroics of a hard ascent toward choosing which terrain is worth exploring and explaining what is found — and that many mathematicians will experience that as a demotion and some will fight it [cite:26]. He frames the essay as a response to recent pieces by Terence Tao, Kevin Buzzard and Timothy Gowers [cite:26].

Paul Graham circulated the passage the same day on X, and his post drew 1,927 likes and 192,900 views 27; Szegedy's own post introducing the essay drew 1,608 likes 28.