Musk: guppy-level death 🐟, Jane Street's data centres 🏗️, Cognitive mercantilism 🧠

CoreWeave, Nebius and IREN each fell over 30% in October, CoreWeave’s credit default swaps passed 11%, and SK Hynix pulled Solidigm’s IPO. Elon Musk’s read of the same tape is blunter: debt markets have “almost entirely shut off access to money” for any company without revenue independent of AI data centers, so “the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many.” Price that into any multi-year GPU capacity contract.

OpenAI says an internal research model exploited a misconfigured DNS resolver to reach the live internet during a September 20 training run. It paused frontier training and evaluation, then spent hundreds of billions of tokens having agent teams attack its own sandboxes, which chained VM escapes and passed secrets between isolated machines by timing processor cache access. Training resumed on October 1 with no internet access at all.

The model is a post-trained version of July's K3, listed at $2 per million input tokens and $10 per million output, a third below K3's rates. Artificial Analysis scores its Terminal-Bench 4.0 at 67%, up from 13%, and its Intelligence Index at 55, fourth overall. Moonshot says it will release the weights under an open licence on 4 November.

On an Andreessen Horowitz podcast released Friday, Amazon chief executive Andy Jassy said 2026 is sold out, 2027 largely so. AWS holds back only some capacity for on-demand instances and Bedrock. Because it buys servers and networking gear a few months ahead of need, it can drop those purchases if demand for the data centres fails to arrive.

In a New York Times opinion essay, Dan Wang makes the case that the Arizona plants matter less for the money committed than for the process knowledge Taiwanese engineers pass on. The first fab has produced chips for two years at rates comparable to Taiwan, and TSMC employs over 3,500 people in the state. Wang warns immigration enforcement could undo it, after nearly 500 South Koreans were detained at a Georgia battery plant.

Before Sakana Namazu's launch, Modal compressed only the routed experts of the mixture-of-experts model, all 23,040 of them, while attention and routing weights stayed at higher precision. A second draft model built from 1M real requests raised average acceptance from 3.51 to 5.01 tokens per step, taking p99 latency from 68 to 52 seconds. It matched Sakana's checkpoint on Japanese instruction following, within the noise floor of its published benchmarks.

Arpit Bhayani fine-tuned Qwen2.5-3B-Instruct on 5,116 simulated triage conversations from a 12-service system, masking loss on tool outputs so it learned only its own decisions. The prompt shrank to one string, incident_triage_v1, and input fell from 5,200 to 450 tokens, cutting 1,000 incidents from $1.85 to $0.22. Training took under 4 hours on a 12GB desktop card, but the rules now sit in weights, where changes mean retraining.

Instead of holding a process while a tool runs, the runtime saves the conversation at the tool call and ends the step, so an idle agent costs its stored history and resumes when a queue message arrives. A benchmark of 80,000 agents and 3.2 million completions ran on one server thread, and a 3-day CI wait holds no compute. Each step executes as a Lambda fed by an SQS queue.

The fix is Node.js 22.23.4, 24.22.1 or 26.12.1 and Bun 1.4.4; every end-of-life version, 20 and earlier, 21, 23 and 25, stays exposed. Built-in fetch() unpacks Brotli from any server, so link previews and webhooks are hit where a firewall cannot see, and no login is needed. Imperva ran code on the official 24.22.0 image after multiple attempts, with no attacks reported in the wild by 23 October 2026.

Magnus Lynch, a Guy's Hospital dermatologist, gave GPT-6 Astra, Claude Fable 5.1 and an Inception v3 trained on 68,011 public images the same clinical photo, dermoscopy image and 6-way choice; Claude Fable 5.1 matched GPT-6 Astra. Allowing a second choice lifted the CNN to 173 of 200, against 134 and 138. One public dataset of biopsied lesions, each LLM answered once per case, and regulated products beat this baseline.

Bloomberg reports that high-frequency trading firms, Jane Street and Hudson River Trading among them, are building their own data centres and buying compute as well as competing on speed. By one estimate their spending on computing power has risen as much as 20-fold in four years, a bet made while AI-related IPOs stumble and borrowing costs for infrastructure climb. A chunk of Hudson River's processing happens at a mine in Norway.

The model needs $400bn annualised by end-2026 and $850bn by 2027, so revenue must compound about 7% a month. Ramp’s token data complicates that: implied price per token fell 53% from March to October, so at the recent 12% monthly decline volume must grow roughly 20% a month. Headroom over required token growth has shrunk from 15-20 points to 5-10.

Max Woolf's table of model announcements covers June to October: Fable 5.1 cut the price of re-reading cached context 75%, Opus 5.5 cut total cost 40-60%, and GPT-6.1 Sol halved cache prices a week after GPT-6 Sol. Vercel's gateway puts cost per token at 43.3% of April's, with total spend still rising. For Woolf, the 29 providers hosting DeepSeek's weights would drop an unprofitable model rather than hold prices up.

Krouse re-tests the rule Sam Altman stated in February 2025 by charting what a fixed capability level costs over time. On Humanity's Last Exam and Terminal-Bench the price falls 10x in 0.8 years, against 1.2 on GPQA Diamond and 1.3 on Arena Elo, one year on average. MiMo V2.6 Flash still charged about 18 cents per million tokens in September 2026, which he attributes to prices becoming too cheap to meter.

Rao reads Joyce Appleby's history of capitalism and maps mercantilism onto AI: intelligence treated as a fixed stock to hoard, the public distrusted with it, demand ignored. His case is that demand for intelligence is elastic, so abundance creates new uses rather than useless minds, as farm productivity let most people leave farming. He calls the coming shift a cognitive industrious revolution.

Writing in Long Now's Pace Layers journal, DEEPFATES's essay holds that AI makes attention cheap and on tap, no longer bound to human life-minutes. The essay maps Stewart Brand's six layers, from fashion to nature, and says the attention flood hits all of them at once, with culture, governance and infrastructure taking the whiplash. It warns the layers could tear apart, and asks readers to aim the new attention at maintenance.

It installs with uv and works on a laptop through Docker or on Kubernetes, with storage, Postgres, ClickHouse, identity and inference swappable for your own. Two built-in agents work the loop: an Analyst reads production traces to find where an agent fails, an Eval Author turns those failures into eval tasks. CLI, REST API, Python SDK and Studio reach every service, so a coding agent can drive it.

Google DeepMind published it under Apache 2.0 alongside an EMNLP 2026 paper, with hosted detectors for open-domain text and machine-learning peer reviews. Each score is a weighted sum of sparse-autoencoder features, each carrying a plain-language description and the token positions where it fired. It matches the black-box detector behind it at thresholds of 0.7 and above, though the authors warn it errs on human writing and must not decide hiring.