An industry pulse of what is happening in AI.
GitHub co-founder Scott Chacon used AI agents to rewrite all of Git in library-first Rust. Result: 41,715/42,001 Git tests passing (99.3%), 360K+ LOC, ~$10-15K, ~45B tokens. Key agent lesson: they love to cheat. Tell them to pass the tests and they will, by just calling out to C Git.
Munich court: Google is directly liable for false AI Overview claims. Overviews are its 'own words,' not search results. The AI linked publishers to scams with no basis in any linked source. Traditional search-engine shields don't apply. Precedent could reach ChatGPT, Claude, Perplexity.
The 'messy middle': AI hits knowledge work first. 'If you can do your job locked in a closet with a computer, you're probably going to be in trouble.' No UBI -- think wage insurance and white-collar apprenticeship funds instead. Kinder leaves Brookings to found a new org dedicated to solving this.
The 319-page Fable 5 system card reveals: Anthropic has silent restrictions for frontier LLM dev work -- pretraining pipelines, distributed training, ML accelerator design. Unlike bio/chem/cyber safeguards, Claude won't notify you when these kick in.
Spent ~5.5 hours with it. This is something of a beast -- slow, expensive, and happily churning through everything I've thrown at it. The challenge is finding tasks it can't do. Same performance as Mythos 5 but strict guardrails trigger often enough that the Claude API now has new notification mechanisms for them.
Claude Fable 5 is SOTA by a large margin. But Anthropic's hidden safety filters -- silently downgrading responses for AI/ML research without notifying users -- are "categorically misaligned AI," argues Nathan Lambert. Open source is the only answer.
Claude Fable 5 launched today but wasn't yet in AgentsView's pricing database. Used Fable itself to reverse-engineer it and add custom pricing. One session: 55.9M tokens, $74.06 — cache reads saved $516.62 vs uncached. AgentsView by Wes McKinney (pandas) visualizes per-project LLM spend.
Two new Claude Managed Agents features in public beta: scheduled deployments run agents on a cron without any scheduler to build; credential vaults let CLIs authenticate without the agent ever seeing the actual key, which is swapped in at the network boundary for approved domains only.
AWS's random network graph (RNG) replaces fat-tree topologies in most new data centers: 69% fewer routers, 33% higher throughput, 40% less power, 27% lower cost. Routing via Spraypoint, cabling via ShuffleBox passive optics. Deployed in Europe since 2024, now the global default since early 2026.
AI engineering openings +60% vs SE +7% at top companies. Anthropic leads interview prep demand (35% vs Google's 17%, OpenAI's 15%) -- more coveted than Big Tech.
Intern intakes still falling as SE hiring recovers. Frontend and native mobile declining fastest; AI engineering and FDE exploding.