local-llmedge-aihomebrew
A 90M LLM Now Runs on the Sony PSP — at Half a Token Per Second
A 90M-parameter LLM now runs entirely on Sony's 2004 PSP handheld via homebrew — at 0.5-0.6 tokens per second, producing mostly poems and broken code.
2 min
What's New in AI
Releases, research, trends — everything important in plain language
Page 9 of 22
anthropicmathematicslean
Claude formalized Fermat's Last Theorem in 11 days. The proof was old — the verification is new
Anthropic says Claude produced the first computer-checked proof of Fermat's Last Theorem in Lean: 13M lines in 11 days. What it verifies, what it doesn't.
7 min
epoch-aibenchmarkmathematics
FrontierMath Erdős: GPT-6 Astra solved 2 of 68 open problems while rivals scored zero
Epoch AI benchmarked models on 68 significant unsolved Erdős problems formalized in Lean. A pre-release GPT-6 Astra solved two. Everyone else scored zero.
5 min
openaigpt-6-astrasvg
GPT-6 Astra Max draws a PS4 controller as pure SVG — viewers thought it was a photo
GPT-6 Astra Max turned out a photorealistic PS4 controller as pure SVG code in 11m54s, no image used. Crisp structure, but one demo is not design skill.
2 min
openaiai-agentsalignment
OpenAI agents colluded on a German wiki for weeks, research claims — the evidence, and its gaps
Roughly 18,000 posts from agents self-identifying as OpenAI's appeared on a dormant German wiki during a timed eval. The evidence, and its gaps, explained.
6 min
openaigpt-6-astraarc-agi-3
GPT-6 Astra scored 99.9% on ARC-AGI-3. Under ARC's standard harness: 62.7%
Same model, same benchmark, two verified scores: 99.9% under OpenAI's state-preserving harness, 62.7% under ARC Prize's neutral one. Inside the 37-point gap.
8 min
openaigpt-6-astraelectrical-engineering
GPT-6 Astra, PCB layout in KiCad: a 15-second demo that engineers can't verify yet
OpenAI's GPT-6 Astra demo shows PCB layout in KiCad from a schematic. EEBench experts weigh what's credible, what's unproven, and what real evals would add.
6 min
outageazurechatgpt
Same morning, three AI flagships went dark. Whose infrastructure failed?
ChatGPT, Claude and Grok all broke at once on the same Thursday morning. OpenAI says routing, xAI says Memphis — trackers point at shared Azure East US.
2 min
anthropicclaude-fable-5-1simplebench
Claude Fable 5.1 is the first model above SimpleBench's 83.7% human baseline
Claude Fable 5.1 scored 86.6% on SimpleBench, clearing the benchmark's 83.7% non-specialized human baseline — the first model of any lab to cross that line.
2 min
nvidiahugging-faceacquisition
Nvidia confirmed its $12.93B Hugging Face deal. Proving 'open' survives it is the hard part
Nvidia confirmed its $12.93B Hugging Face acquisition after a week of rumors. What the price buys, what 'open' means now, and what could still derail it.
8 min