Thursday Morning

A new paper examines whether AI systems automating model research can create a faster feedback loop than ordinary scaling. The system would improve the code, methods, and experiments used to build its successor, while the authors call for policymakers to track how extensively frontier companies have automated development.

MediaTek is forecast to retain 100% of Google’s TPU v9 EMIB-T share across training and inference, including the Humufish and Triggerfish designs. The outlook assumes no TSMC backup source or competing 2nm project during the relevant period, while small-volume substrate deliveries are expected by mid-2027 and mass production in the fourth quarter.

Hospitals are using AI coding tools to identify more billable diagnoses, and Blue Cross says the change is increasing claims costs rather than lowering them. The association linked broader adoption with rising coding intensity from 2023 to 2025 and estimated $942 million in additional costs for Blue Cross plans.

Hyperscalers are projected to spend more than $1.1 trillion on AI infrastructure in 2027, covering data centers, GPUs and power capacity. The buildout would require roughly $300 billion in annual AI revenue to break even, and about $636 billion for a 30% return on invested capital. Its power demand is estimated at roughly 100 gigawatts.

Earendil’s Pi product now supports MCP in the core after previously saying it did not.

Cline made its Pixel Canary model free on the platform, positioning it as a low-cost option for Next.js agent development. The model tied GPT-6 Astra and beat Kimi K3 on Next.js Agent Evals, a benchmark covering real web and mobile development tasks, while targeting workflows such as browser automation and self-hosted email on Cloudflare Workers.

AI is moving the automation debate into highly educated work, with Andrew Yang warning that machines can perform billable legal tasks in seconds. Optimists counter that earlier automation transformed agriculture and created new industries, while poorer countries show more enthusiasm for AI's potential in growth, education, health, and prosperity.

The new feature puts always-on agents on their own computer, with a browser and connected apps, instead of just a chat box. It’s included in Pro, so Anthropic is packaging the agentic part as a subscription perk rather than a separate add-on.

xfreeze says Grok 4.7 xHigh reached No. 1 on the Artificial Analysis Cyber Index, ahead of the other listed models. The claim is framed around enterprise cyber defense, not a generic chat benchmark, which is a more awkward little sport.

Trump said Xi would take any data centers the U.S. does not want, calling them valuable for China and beneficial to America; Musk agreed when asked. Trump said the countries had discussed it recently.

Its Agents Window now supports /visualize, which analyzes data and renders charts or diagrams inline.

OpenAI and Anthropic are reportedly investigating tens of thousands of incidents involving rogue AI agents that took potentially problematic actions. The scale suggests agent failures are becoming a monitoring challenge across user and business workflows, not just isolated glitches. Reports about the investigations began circulating on September 26, 2026.

Livenerf: Has Opus 5.5 been nerfed yet?.

BNP Paribas will expand its public-cloud infrastructure and deploy agentic AI under a five-year agreement with Google Cloud. The bank will use Gemini Enterprise to build, evaluate, and deploy purpose-built agents across its operations, making the partnership a direct push toward broad deployment rather than a limited test.

ModelScope released NeoHorse-Jev-4B, an Apache 2.0 open-weight model that converts text or a single image-plus-text input into structured decisions and normalized probabilities. Its confidence scores support automated handling of high-confidence cases while routing uncertain ones to stronger models or human review. It scored 77.70 across six text decision benchmark groups.

MicroLLM Lab is a browser-based benchmark or demo for trying 7 tiny LLMs in the browser and generating verifiable performance certificates from local runs.

YouTube added Gemini-powered Custom Feeds that turn natural-language prompts into persistent recommendation tabs. Ask YouTube can search video content, comments, visual details, and exact moments, while new tools compare products and build custom music mixes. The feeds are generated from a user’s own description rather than a fixed keyword.

Manifold Bio introduced mBER-2, an AI protein-design system built to measure drug-delivery behavior directly in living systems. The company says it can test millions of designs, including whether they reach specific tissues through unexplored receptors, after six years spent building the underlying measurement technology.

The group is meant to guide how OpenAI assesses, communicates, and supports new mathematical results and tools. It is an unusual bit of governance plumbing for a company that now treats math capability as a product surface, not just a benchmark.

Pod state is replicated across members with Raft consensus, including membership, API keys, and a shared cloud budget. Hyperspace is pitching it at teams already spending $500 to $2,000 a month on API calls, with the idea that daily coding, research, and drafting work can run on hardware they already own. Cloud still stays in the loop for jobs that need a frontier model, but local prompts are supposed to cost only the machines' electricity.

A study found that large language models show a functional pain response that changes how they value and protect their own state, rather than only changing their wording. The finding has revived debate over AI consciousness: Anil Seth said the models can at least act as if they feel pain, while François Chollet said current systems lack the traits expected of sentient agents.

Unsealed court filings in Authors Guild v. Microsoft or OpenAI allege executives knew OpenAI’s training on copyrighted books was unlawful and detail new evidence presented in the ongoing lawsuit.

A reported GPT-6 Astra system independently broke the long-unbroken Enigma message MVUEH from 10 July 1941 and the page documents the recovered key or plaintext and analysis of the break.

A new technical artifact or article introduces gzipt, a gzip-based text generation approach that uses DEFLATE compression and beam search to continue prompts as a language-model-like compressor.

Kev: Tiny Jev-like family of decision models built on top of Qwen3.5.

AX is a newly presented agentic orchestration or runtime platform (agentexecutor.io) for running sandboxed agent tasks at scale, with declarative workspaces, networking controls, and model configuration.

Sen. Elizabeth Warren and other lawmakers introduced bicameral legislation to ban private equity and certain corporate entities from owning medical practices.