live
- your daily tech & design digest, one email a day -
Microsoft ships Decision-1, a model built for fast decisions [Xtra!]A purpose-built 'decision' model signals a shift from general chat models to narrow, latency-optimized agents.
A deep technical tour of speculative decoding [Xtra!]Speculative decoding is one of the main levers for cutting LLM inference latency and cost - this is the long-form explainer worth bookmarking.
The case that self-improving software is inevitable [Xtra!]A sober argument for why code that modifies and optimizes itself is coming regardless of how uncomfortable that makes engineers.
Chainalysis used AI to trace the $387M Bitget hack to North Korea [Xtra!]A concrete case of AI-assisted blockchain forensics attributing a massive hack in near real time.
Coinbase built and post-trained an LLM fraud agent [Xtra!]Real production numbers (30% fraud reduction) on layering an LLM agent onto existing ML/rules systems rather than replacing them.
Microsoft will let Copilot act on local files on Windows PCs [Xtra!]Giving an AI agent write access to local files on hundreds of millions of PCs raises the stakes on sandboxing and rollback.