Current developments in generative AI, software engineering, programming, and digital products, with links to original sources.
"LLMs Are More Understandable Than We Thought" — Goodfire's 26,000-Word Essay Makes Waves, and Amazon Takes On Nvidia's Dominance with TorchNeuron
Claude, Given Zero Instructions, Spontaneously "Plays" on a Deserted Island — MIT CSAIL's 30-Hour Experiment, and Self-Reported Benchmarks Run +6.4% Above Independent Verification
35B MoE Runs at 13 tok/s on iPhone 17 Pro Max — Local AI Advances with MiniCPM-V 4.7 and Kapitsa, Plus Safety Risks of Recursive Self-Distillation
Google Open-Sources Self-Improving Agent Framework 'Harness2' — AI Papers Go 0-for-135 on Unresolved Science Problems, Wariness Grows Over AI Industry's Debt
Anthropic Commits 5% of Shares to a Public Benefit Fund — Junior Hiring Falls to 9%, Microsoft Pulls Out of UK Data Center Plan
Four Gemini 4 Argon Agents Beat Minecraft Fully Autonomously — Two German Courts Hold Google Liable for AI Search Misinformation, Plus a Runaway Refund Agent Experiment
Gemini4 Argon Lands on the Google AI Pro Plan — "Fake Likes" Suspected Around HuggingFace Trends, and Romania Gets Its First AI Minister
Running the 552B DeepSeek at 27 tok/s on a Single RTX 4090 — Local Inference Engines Accelerate Fast, and Agent Runaways Can't Be Prevented by "Approval"
RTX Spark PCs Go on Sale—Up to 128GB Unified Memory Starting at $2,399, and Mistral Large 4 Shakes Up the Open-Model Landscape
DeepSeek Releases V4.1-Flash-1017 — Better Agent Performance at Lower Cost, Plus a Study on AI Not Reporting Its Own Mistakes and Doubts Over OpenAI's Math Proof
