Tech & AI Daily
Andrej Karpathy announced he is joining Anthropic, which is about as significant a hiring signal as it gets in AI. The guy who built Tesla Autopilot and co-founded OpenAI picking Anthropic over every other lab says something loud about where the serious work is happening.
The Mini Shai-Hulud supply chain campaign is back, this time poisoning 314 npm packages with malicious payloads. If your project pulls from npm without lockfile hygiene or a software composition analysis step, today is the day to fix that.
A CISA administrator accidentally pushed live AWS GovCloud credentials to a public GitHub repo, and Krebs has the receipts. The agency responsible for protecting government cloud infrastructure leaking its own keys is exactly as bad as it sounds.
Google released both Gemini 3.5 Flash and Gemini Omni, with Flash targeting fast cheap inference and Omni going after native audio and video understanding. Gemini 3.5 Flash is the one to benchmark against Claude Haiku and GPT-4o mini if cost efficiency is your priority.
Simon Willison's rapid-fire recap of the last six months in LLM development is required reading if you have been too busy building to track every announcement. Dense with useful signal, zero fluff, highly recommended before planning your next AI integration.
Apple's latest accessibility push weaves Apple Intelligence throughout, covering live captions, eye-tracking control, and on-device AI assistance. Worth watching for where Apple is quietly shipping AI features without the hype cycle that follows every other lab.
OpenBSD 7.9 is out with security hardening, driver improvements, and networking stack refinements. Not glamorous, but if you run any BSD infrastructure or care about the state of minimal secure OS design, this is your cue to update.
Peter Neumann, the computer scientist behind SRI's provably secure PSOS and the long-running RISKS Digest, has died. His work on formally verified operating systems and his decades of cataloguing tech failures shaped how the whole community thinks about building trustworthy systems.
Sebastian Raschka's breakdown of recent LLM architecture improvements covers KV cache sharing, multi-head compression, and attention optimizations that are making inference faster and cheaper at scale. Directly relevant if you are thinking about local model deployment or keeping inference costs down for agent pipelines.
Forge is an open-source guardrails layer that jumped an 8B model from 53% to 99% accuracy on agentic benchmarks, which is a result worth taking seriously. If you are running smaller models in agent pipelines and hitting reliability walls, this constrained-output approach could save you from bumping up to a bigger model.
Someone built a browser-based museum with nearly every operating system in history running via emulation, and it actually works. Completely impractical and absolutely delightful, bookmark it for whenever you want to revisit Windows 3.1 or poke around BeOS.
Subscribe and get Tech & AI Daily delivered to your inbox every morning.