Social Media Briefing — February 12, 2026

Karpathy writes a detailed post about using DeepWiki MCP + GitHub CLI to extract specific functionality from codebases. He used an agent to 'rip out' torchao's fp8 training into 150 lines of clean self-contained code that runs 3% faster. Argues software should become more modular 'bacterial code' and that 'libraries are over, LLMs are the new compiler.'

Karpathy releases a new project: training and running inference on GPT in 243 lines of pure, dependency-free Python, calling it the full algorithmic content of what's needed with everything else being for efficiency.

Following yesterday's Research coverage of the Opus 4.6 system card, Anthropic announces they're delivering on their commitment to write sabotage risk reports for frontier models, starting with Claude Opus 4.6. They noted when releasing Opus 4.5 that future models would be close to ASL-4 threshold for autonomous AI R&D.

Anthropic announces commitment to cover 100% of electricity price increases from their data centers, pay grid upgrade costs, bring new power online, and invest in systems to reduce grid strain.

Following yesterday's Research coverage of the Opus 4.6 system card, Anthropic shares link to sabotage risk report for Claude Opus 4.6, explaining they decided to preemptively meet the higher ASL-4 safety bar rather than debating blurry thresholds.

Sam Altman announces an update to GPT-5.2 (the instant model) in ChatGPT, describing it as not a huge change but hopefully a little better.

Jason Warner (Poolside CEO) writes an extensive essay arguing that intelligence is the new critical infrastructure, comparing AI model providers to cloud hyperscalers. Key arguments: (1) Google is the only full-stack AI player, (2) large companies MUST train their own models or die, (3) frontier model training costs ~$500M not billions, (4) first-party agents will always beat third-party ones, (5) app companies are just arbitraging current model capability gaps. Promotes Poolside's 'Model Factory' approach.