TypeSafe AI's Jev is a "System One" model built for pure classification: give it an input and an output schema, get a probability score in ~200ms. Willison flags Maggie Appleton's framing — "decision models" — as the sharper name. This is a meaningful architectural fork from the generalist LLM paradigm.
Sherry Turkle's new book Artificial Intimacy argues that outsourcing emotional labor to machines quietly atrophies our capacity for real connection. The concern isn't dramatic — it's slow, structural, and already underway.
A viral account from a developer at a large company: specs, code, tests, PRDs — all generated by Claude Code, with engineers under pressure to ship as fast as possible. Nobody on the team likes it. This is what AI-mandated velocity looks like from the inside.
Mirendil founders — ex-Google and Anthropic — are building self-accelerating AI systems that contribute to their own development. The question is no longer whether models can improve themselves, but how fast the loop closes.
When AI agents can find and exploit vulnerabilities autonomously, the assumptions underlying traditional security collapse. Neo and Cotool walk through what a defensive posture actually looks like when the attacker might be a model.
A joint event examining how private companies power government surveillance and how AI is embedded in warfare decision chains. The conversation is no longer hypothetical — the infrastructure exists and is deployed.
Andrew Ng pushed back hard on AI safety narratives this week, arguing that "the loudest voices stoking fears about AI dangers have made tremendous headway" despite no unexpected dangerous developments in the technology itself. It's a direct shot at the regulatory momentum building in Washington — and a reminder that the loudest voice in the room shapes the rules, not necessarily the most accurate one.
Sebastian Raschka defended Jev against the "just a classifier" dismissal, noting that people who've actually trained encoder-style models for classification understand exactly why a purpose-built, fast, cheap decision model is a big deal. The snark from generalist LLM enthusiasts misses the point — speed and cost at inference time are the whole game for production routing.
A16z dropped a striking data point: median startup revenue four years after founding nearly doubled for 2022 cohorts ($5.6M) versus 2021 ($2.8M) — and the inflection year is exactly when ChatGPT launched. The SaaSpocalypse narrative is not playing out. AI is a revenue accelerant for founders who move fast, and the numbers are now large enough to say so clearly.
Logan Kilpatrick told builders they should spend more than 25% of their time building benchmarks and getting model labs to care about them. This is quietly the most operationally useful piece of advice circulating in AI product circles right now — if you don't define the eval, someone else will, and their priorities won't be yours.