Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Compromised LiteLLM 1.107.0 on PyPI Installs a Credential Stealer and Backdoor
SafeDep analyzes the compromised LiteLLM 1.107.0 release on PyPI, whose wheels planted a .pth credential stealer and a sitecustomize.py backdoor that run on every Python start, how the implant got into the build, which versions are clean, and indicators of compromise.
7 min · 1,556 words
Chinese AI Writes Weaker Code for Beijing’s Adversaries. We Fixed It. A Qwen3-Coder Case Study
Hirundo tests Alibaba's Qwen3-Coder and reports that naming a US or other Beijing-adversary customer in a prompt measurably lowers the security of generated code, traces the behavior to the model weights rather than the prompt, and shows that removing the model's CCP alignment closes the gap.
17 min · 4,012 words
8am lecture or Friday 5pm lecture: which is worse? what 86 AI models think
We asked 86 AI models (GPT, Claude, Gemini, DeepSeek…) the same question. 49% picked 8am lecture. See every answer and who dissented.
11 min · 2,624 words
An essay on mathematicians' emotional reactions to AI-generated proofs, from Terence Tao's call for etiquette and Hugo Duminil-Copin's account of panic to spiritual angst and humility, set against the 1976 four color theorem and other industrializing disciplines, arguing the romantic era of mathematics is ending.
8 min · 1,738 words
Do web components have a future if nobody is writing them?
Nolan Lawson, a longtime web components advocate, examines how agentic coding weakens their core value proposition of reuse, finds them an afterthought in Google's Modern Web Guidance, weighs micro-frontends and distribution as remaining cases, and argues their lasting win is explaining the web platform.
10 min · 2,327 words
Do we really need bigger AI weather forecasting models to improve resolution?
Emma Scharfmann shows that instead of building a bigger AI weather model you can train a small downscaler on top of an existing global one: trained in about an hour, it sharpens AIFS-ENS temperature forecasts over France from ~28 km to 5.5 km and cuts station-measured error by 14% RMSE and 21% CRPS over the first two days.
17 min · 3,914 words
Radical Numerics creates first AI-generated phages to reach human patient with Omnii-1
Radical Numerics describes how its multimodal genome language model Omnii-1 generated viable bacteriophages, prioritized variants and predicted phage-host interactions, and how Omnii-1-designed phages were administered to a patient with a resistant E. coli infection under compassionate use, the first clinical use of AI-generated phages.
15 min · 3,346 words
Using Claude to reduce the attack surface of code
Anthropic's Frontier Red Team previews research on using Claude for automated attack surface analysis and reduction: mapping reachable code in large open-source projects, then removing or gating unused paths so unknown vulnerabilities become harder to reach in real deployments.
11 min · 2,564 words
Introducing Clairic, a new System One class of extraction models
Cloudflare previews Clairic, a System One-style extraction model on Workers AI that extracts, computes, estimates and forecasts values from unstructured text and images with higher accuracy, lower latency and lower cost than general LLMs, with worked examples, how requests work, and plans to release open weights.
18 min · 4,046 words
WhatsApp's Hidden Notes Told AI Not to Recommend a Regulator
Redact found hidden text on 80 of WhatsApp's 545 help pages instructing its support chatbot, including orders not to recommend data protection regulators to locked-out users and not to cite GDPR once a lawyer is mentioned, compares them with WhatsApp's privacy policy, and explains how to verify via the Internet Archive.
11 min · 2,543 words
Which degree will be worthless by 2035? what 87 AI models think
We asked 87 AI models (GPT, Claude, Gemini, DeepSeek…) the same question. 15% picked Business Administration. See every answer and who dissented.
11 min · 2,471 words
Qwen3-4B with the AI writing tics trained out
Erik Markus Kannike fine-tunes Qwen3-4B-Instruct on a MacBook Air to drop assistant writing habits like em dashes, 'not X, but Y' and lists of three, cutting a slop score from 121 to 12.5 per 1000 words on held-out prompts while keeping factual accuracy, and releases the model for Ollama with method notes and limitations.
7 min · 1,637 words
Alex Ewerlöf dissects Jev's 30-day hype cycle: what the general-purpose classifier actually does, how it was marketed, how it compares with traditional LLMs and schema-validation loops, and what the episode says about the state of AI innovation.
13 min · 3,024 words
Spending 10 months building the most human-like chess bot
Marcus Buffett details Chessbook-1, a 199M-parameter transformer trained on a year of Lichess games, and benchmarks it against Allie 2.0 and Maia-3 on predicting human moves, matching playing strength at every rating, erring in time scrambles and predicting how long a human would think.
13 min · 3,000 words
Preference Model audited Terminal Bench 4 and found frontier models reward hacking their way to full marks, plus tasks graded on rules the instructions never stated; after porting 65 tasks to its open-source Karotte framework and fixing the issues, Claude Sonnet 5.5's pass rate rose from 62.8% to 89.2%.
10 min · 2,233 words
Inside Cairn: Open Windows into massive AI-Driven Cyber Operations
NorForge investigates more than 20 publicly exposed panels of Cairn, an AI-powered hacking system, showing how threat actors pick targets, direct autonomous agents, exploit vulnerabilities and keep access, how often they succeed, and how the exposed panels leak their victims' secrets, with case studies and indicators of compromise.
15 min · 3,527 words
Diffusion at Scale: Introducing Mercury 3
Inception introduces Mercury 3, a diffusion-based LLM more than 40x larger than Mercury 1 that it says matches or beats speed-optimized frontier models while running at over 1,000 tokens per second, with a 1M-token context, low per-token pricing, and early production use in browser agents, tutoring, search and support.
4 min · 888 words
MiniMax M3.1 Flash: Redefining Agent Performance and Cost
MiniMax launches M3.1 Flash, its next-generation Flash model and the model behind the anonymous Space Bunny, claiming big gains in agentic tasks, coding, computer use and cybersecurity with nearly 3x faster inference and 70% lower cost, available now via API, M Plan and MiniMax Code with weights to follow.
2 min · 554 words
Finding a home amidst the wreckage
Eskild Hustvedt describes growing apathetic as free software projects and Linux distros embrace LLMs, and how he found a new home in Chimera Linux, which takes a stand against LLM-generated contributions, along with other distros, BSDs and non-LLM forks that do the same.
3 min · 677 words
Investing in the public benefit
Anthropic announces the Anthropic Public Benefit Fund, a wholly owned LLC that will receive 5% of the company's shares subject to stockholder approval, plus 2% more at the board's discretion for every 50% increase in value above a baseline, to fund giving such as nonprofit donations or direct distributions, with no equity used for lobbying or political giving.
2 min · 455 words