AI & ML Research 24 Hours

Last updated: September 12, 2026, 6:21 AM ET

Agentic Coding in Production

Perplexity has handed end-to-end systems work to GPT-6 Astra, letting the model draft communications, modify software, and monitor production, with far fewer human check-ins than earlier models required. Cognition is applying the same model so that Devin can test its own output and prove it works, aiming to cut the volume of code engineers must review and increase shipped work. As these agents take on more autonomy, one practitioner argues the real bottleneck is not context length but intent continuity: a system that discovers, verifies, and reapplies earlier requirements without asking users to restate them. A companion piece contends that AI-generated code makes software design more, not less, important.

Safety Debates and Statistical Rigor

MIT Technology Review convened lab employees who say there is a genuine possibility that advanced AI could destroy humanity, weighing whether the warnings are credible or mere scaremongering. On the measurement side, a new explainer unpacks why a 95% interval is routinely misread, since frequentist confidence intervals and Bayesian credible intervals answer fundamentally different questions and conflating them distorts product decisions.

Interpretability and Industry

A mathematical primer clarifies the representation workspace behind Anthropic's J-Space. Elsewhere, MIT's daily digest spotlights the under-35s shaping biotech's future alongside cheaper, cleaner steel production.