Sunday’s haul has a split personality, and the split is the story. On the capability side, everything is compounding: a Chinese lab ships a frontier-class model and promises to give the weights away on November 4, an American chipmaker publishes the arithmetic for serving a three-trillion-parameter model across 59 wafers, five small open models in a shared-folder harness tie a frontier flagship on a coding benchmark, and ChatGPT quietly rewires what it cites—documentation over blogs, by a landslide. On the money side, Apollo’s chief economist looked for the payoff outside tech and found a flat line. And underneath both, three of the most serious consciousness researchers alive published the week’s most argued-over claim: today’s AI is not conscious, and the road we are on probably does not get there. Capability up, payoff flat, mind unproven. That is an honest Sunday.
Kimi K3.1 Pro: the frontier, soon to be downloadable
Moonshot AI released Kimi K3.1 Pro on its API on October 24, priced at $2 per million input tokens and $10 per million output—matching GPT-6.1 Sol and half the price of Claude Opus 5.5, per Simon Willison’s write-up. Independent scoring by Artificial Analysis put it at 55, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon at 53 each, and behind only Opus 5.5 at 58. Moonshot promises the full weights by November 4. The detail that makes this more than a price war: Moonshot deliberately cut the model’s offensive cyber skill first, removing cyber-related data in mid-training and filtering cyber reinforcement-learning environments. On its internal exploit benchmark, xdev-bench-hard, the released model scores 81.0% against 90.5% for an unfiltered checkpoint.
“We aim not to release open weights that would extend the absolute frontier of cyber capability.”
Hold both halves at once, the way the board has to. The weights are a promise, not a delivery—November 4 is the test. And a nine-point trim on an internal benchmark is a design choice, not a guarantee, especially for weights anyone can fine-tune afterward. But the direction is historic: the most capable soon-to-be-open model in the world is arriving with its maker arguing, in public, about what an open release owes the defenders. The code-is-law chair calls that architecture as policy. The free-speech chair notes it was done by training choices, not by licensing speech. Both chairs are right, and the argument itself is the news. Cyber detail via The Tower Post.
Cerebras does the wafer math: 59 chips, one model
Cerebras published a technical post walking through how it would serve Kimi K3—about three trillion parameters, 93 layers—on its WSE-3T wafers. Each wafer holds 44 GB of on-chip memory; at 4-bit quantization each layer is about 16 GB. Dense layers get a wafer each, with the spare room for key-value cache; linear-attention layers, which keep a small fixed state, pack two to a wafer. Total: 35 wafers for the 69 linear layers plus 24 for the dense layers—59 wafers, weights parked on-chip, only activations crossing between stages at roughly two microseconds a hop, at a claimed 1,000 to 2,000-plus tokens per second.
Read the fine print the way the Slashdot chair insists: Cerebras itself calls K3 “an illustrative example,” with an assumed 256k average request, and says the honest metric is tokens per second per megawatt at matched speed—a number the post does not give for this layout. So this is a deployment recipe, not a measured result. It is still worth your morning, because the recipe is the point: frontier inference is becoming an exercise in keeping weights still and moving as little else as possible. The model is no longer the hard part. The plumbing is.
ChatGPT changed what it cites: docs beat blogs, 33.5% to 9.3%
This is the home-beat item, and it is a big one. SEMAI’s AI Search 2026 Report, released October 21 and built on more than 250,000 URLs across ChatGPT, Gemini, Perplexity, and Google AI Overviews, compared the same queries measured the same way in February and now. Gemini and Perplexity barely moved. ChatGPT flipped: after its model change, documentation and support pages rose from 1.4% to 33.5% of recorded URLs for ChatGPT 5.6 Luna, while articles and guides fell from 41.2% to 9.3%. Company websites still appear in 79.4% of Luna’s matched queries, and industry associations and public authorities show up in 56.9% and 46.4% respectively. The full report is on SEMAI’s site.
“A model change reshaped what ChatGPT cites, and it did not happen on the other engines.”
SEMAI sells AEO tooling, so discount accordingly—but the method, same queries, same measurement, two dates, is the right shape, and the size of the swing is hard to hand-wave. The practical translation for any business I advise: your help center is now a ranking surface. This paper has argued that the citation is the new ranking; today the citation has a favorite genre, and it is the dull, accurate, step-by-step documentation nobody wanted to fund. Write it like you mean it.
Five Qwens in a trenchcoat tie Claude Fable 5
A GitHub project called GVS5H orchestrates five Qwen3.8-27B instances through a shared filesystem—a plan file, a tasks file, a verifier role—and reports 92.4% pass@1 on LiveCodeBench hard problems, essentially tied with Claude Fable 5 at 90.4%. The caveats arrived instantly and they are fair: an agentic loop is not a single-pass benchmark run, LiveCodeBench is a saturated 2025-era benchmark, and there are no numbers here against Astra, Sol, or Opus 5.5.
Grant every caveat and the story survives, because the story is not the score. It is that coordination is turning into a substitute for scale: five mid-size open models, a shared folder, and explicit roles doing work we assumed required one giant proprietary brain. That rhymes with yesterday’s Stanford item—cheaper reasoning through structure, not just size—and it is the pattern small operators should bet on. You may never own the biggest model. You can absolutely own the best-organized team of small ones.
Apollo: no sign of an AI payoff outside tech
Apollo Global Management’s chief economist, Torsten Slok, published “No Signs of AI Payoff Outside the Tech Sector” on October 24. His split is brutal in its simplicity: for the Magnificent Seven, revenue per worker and profit margins are rising; for the other 493 companies in the S&P 500, both are flat, and the San Francisco Fed’s productivity series shows no AI lift. Mainstream corporate America, he writes, is starting to treat token spending as a cost center, does not need the latest model, and is unwilling to pay for it.
“Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech.”
This paper is techno-optimist and stays that way; a flat line in 2026 is not a verdict on 2030, and general-purpose technologies have always paid late—electricity took decades to show up in the factory numbers. But Slok’s chart is necessary cold water for anyone selling AI into ordinary businesses, which includes me. The 493 do not buy capability. They buy judgment they can trust, in context, with accountability attached. Until the offer sounds like that, the payoff line stays flat and the token bill stays a cost.
Nature: three reasons today’s AI is not conscious
The week’s most argued-over argument is not a product at all. In Nature on October 19, Anil Seth of the University of Sussex, Ned Block of NYU, and Giulio Tononi of the University of Wisconsin–Madison argued that today’s AI systems are not conscious and that consciousness in AI is “extraordinarily unlikely, at least in the digital, silicon-based systems that underpin today’s technologies.” Three reasons: intelligence and consciousness come apart—the cerebellum holds nearly four-fifths of the brain’s neurons and computes brilliantly outside awareness; substrate matters, because “brains are not—or at least not just—computers that are made of meat”; and the leading theories, integrated information theory and biological naturalism, both point to wiring and living material that feed-forward deep learning does not have.
“LLMs are mirrors that reflect an image of our collective digitized past. We gaze into them and all too readily confuse the reflections with our own conscious minds.”
I am on record here as super interested in this question, and I keep it open where it is open. This is a strong argument from serious people, aimed at a real confusion—fluency feels like a mind. It is also, note, an argument from theories that are themselves contested, published as a comment, not a measurement. The honest position this Sunday is theirs to rebut and mine to hold loosely: nobody has a consciousness meter, the mirrors are getting better, and the question of what, if anything, it is like to be one of these systems remains exactly that—a question. Wonder and evidence, both. That is the beat working as designed.
What the day adds up to
Six items, one tension: the machines are compounding and the ledger is not—yet. An open frontier model with its cyber edge deliberately dulled. A 59-wafer recipe for serving it. A citation economy that just promoted documentation over blogging. Five small models organized like a team, tying a flagship. An economist’s flat line for the 493 companies that are supposed to pay for all of it. And three scientists reminding us that none of this, so far, adds up to a mind. For a business owner the Monday-morning version is short: fund the boring factual pages, organize small models before you rent bigger ones, sell judgment rather than tokens, and keep a human gate on anything that acts in your name. The capability will keep compounding. Whether the payoff follows is now the only chart that matters.
Sources: Moonshot Kimi K3.1 Pro via Zot News and Simon Willison, October 24, 2026; Kimi cyber detail via The Tower Post, October 25, 2026; Cerebras, How Cerebras Deploys Trillion-Parameter AI Models, October 22, 2026; SEMAI AI Search 2026 Report via EIN Presswire, October 21, 2026; GVS5H and the Apollo note via ai0.news, October 25, 2026, with Apollo detail via PakTechStan, October 25, 2026; Seth, Block, and Tononi in Nature, October 19, 2026, via AI Weekly.