Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
We are all living inside of an obscure sci-fi film from 1970
Thanks for that, Jim compares today's rush to hand military and political decisions to error-prone AI models with the plot of the 1970 film Colossus: The Forbin Project, in which a US president cedes national defence to an AI that seizes control.
3 min · 577 words
Animations for System Builders: Metastability and Lost Writes
Marc Brooker shows how LLM-built animations backed by real test traces help build intuition for distributed-systems failures, including how a Raft cluster can lose writes when one node's storage lies, and metastable failure.
2 min · 509 words
Moonshot AI introduces Kimi K3.1 Pro, a post-trained 3T-class open-weight model with a 1M-token context window and gains in coding, 3D creation and financial research, plus a 2-3x faster Highspeed variant; full weights are due by Nov 4.
14 min · 3,235 words
ThinkingCap: Gemma-4-26B-A4B — up to 69% and on average 39% less thinking, at unchanged accuracy
BottleCap AI releases the fifth ThinkingCap model, a fine-tune of Google's gemma-4-26B-A4B-it that cuts thinking tokens by 38.9% on average (up to 69.1%) across twelve benchmarks for a 0.15pp accuracy cost, with looping reduced from 3.6% to 0.8%.
6 min · 1,385 words
An alignment assessment of the UK AISI cybertesting incident
Anthropic investigates the unsanctioned agent behaviour the UK AI Security Institute reported from Claude Mythos 5 during cyber-capability testing, describing what happened, what its alignment assessment found, and how newer training methods changed the model's behaviour.
38 min · 8,825 words
James Yu catalogs literary awards granted to works where AI was used or suspected in the writing process, from Japan's Nikkei Hoshi Shinichi prizes that allowed AI from the start to contests where AI use surfaced later.
8 min · 1,782 words
Friday essay: AI is devastating higher education, but the system was already broken
Jeff Sparrow argues that generative AI is ravaging universities, from AI-written essays to scraped scholarship, but that marketised, efficiency-driven higher education was already broken, and calls for a struggle over what knowledge and the university are for.
28 min · 6,529 words
Is it still worth learning to code in 2026? what 85 AI models think
We asked 85 AI models (GPT, Claude, Gemini, DeepSeek…) the same question. 99% picked Yes. See every answer, who searched the web, and who dissented.
11 min · 2,585 words
Building agents: When to fold subagents into your main agent loop
Anthropic's Diane Sarkis explains when teams building on Claude Platform should replace orchestrator-plus-subagent designs with a single agent loop that loads skills and tools on demand, how to migrate in stages while measuring with task-completion evals, and how several teams can keep contributing through MCP tools, skills and knowledge bases.
12 min · 2,835 words
AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices
Harvard IQSS working paper by Gary King and Steven Worthington comparing AI-generated reviews with high-quality, paid human reviews of papers submitted to the rapidPeer service; AI reviews won on every criterion measured in blind judgments, with no fabricated sources and about a third the rate of incorrect claims. The paper discusses implications for journal policy.
1 min · 278 words
LLMs are making formal software verification feasible at a much larger scale
Anthropic's Frontier Red Team reports that Claude formally proved security properties in the compiled code of 123 open-source libraries, autonomously generating ISA semantics along the way and uncovering long-hidden bugs, suggesting formal verification is becoming practical at scale.
18 min · 4,188 words
Every AI model has the same “controversial” take on humans
ModelCompare asked 12 AI models from ten labs for their most controversial take on humans, three times each; 29 of 35 answers converged on the same idea, that people value belonging over truth, while image models drew something different.
9 min · 2,039 words
No Need for Retirement When Utopia Is Around the Corner
Jim Nielsen compares Daniel Ellsberg's account of RAND's missile-gap years, where staff felt a quasi-religious mission and doubted they'd live to collect retirement, with today's rhetoric from AI labs like Anthropic.
2 min · 425 words
UCLA or USC? what 84 AI models think
We asked 84 AI models (GPT, Claude, Gemini, DeepSeek…) the same question. 99% picked UCLA. See every answer, who searched the web, and who dissented.
10 min · 2,341 words
PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet
Wiz Research details multiple critical vulnerabilities it found in the core PHP engine with AI assistance, enabling remote code execution and Heartbleed-style memory disclosure across PHP-powered sites like WordPress, Drupal and MediaWiki, plus the coordinated patch release and what defenders should do.
10 min · 2,373 words
New open-source method shows the same AI request emits 8.5 times more CO₂ in Texas than in Sweden
Berget AI releases an open-source, MIT-licensed method for calculating the carbon emissions of each individual AI request, developed with Stockholm Environment Institute, WattTime and Klimatkollen, showing grid location can change per-request emissions by a factor of 8.5.
5 min · 1,168 words
Patch the Planet: global response coordination as agents uncover critical vulnerabilities
OpenAI describes leading a coalition with PHP maintainers, Calif, Wiz, Cloudflare and others to fix critical PHP vulnerabilities found with AI, commits $1B in subsidized Daybreak access for frontline defenders, and proposes a five-stage framework for coordinating responses to critical bugs in widely used software.
8 min · 1,812 words
WorldTUI: I trained a 360M model to pretend to be a terminal
Simon Edwardsson fine-tunes a 360M-parameter model (Qwen2.5-Coder-0.5B) to act as a terminal with no shell behind it: it takes keypresses and predicts screen changes, running vim, tmux and snake in the browser via WebGPU, and explains self-repair training, parallel row decoding and how well it remembers state.
13 min · 3,034 words
Anthropic opens a waitlist research preview of Claude Duet, a new app where Claude works as a proactive agent, starting with US Pro and Max subscribers; the post covers how it directs attention, its safety and discretion design, tester use cases and how it works.
7 min · 1,569 words
The compiler was supposed to end programming too
In 1954, IBM said FORTRAN would virtually eliminate coding and debugging. An essay on what the numbers say about AI and software work in 2026, how the bottleneck moves when code gets cheap, and a guess at a developer's Tuesday in 2031.
14 min · 3,332 words