AI digest: Infrastructure gets serious
Memory compression breaks new ground, agents get proper auth protocols, and the industry splits on whether current AI is actually intelligent.
This week’s standout theme is infrastructure maturing fast. We’re seeing real solutions to the practical problems holding AI back.
Together AI cracks the memory problem with 2-bit quantisation
Together AI’s OSCAR uses attention-aware compression to squeeze KV caches down to 2.28 bits per element whilst keeping accuracy decent. This matters because memory bandwidth is the real bottleneck for long-context models, not compute. Smart move making it open source.
WorkOS tackles the agent authentication mess
The new auth.md protocol lets AI agents register with web apps properly instead of the current bodge of humans filling forms. It’s just a Markdown file that apps publish to tell agents which OAuth flows they support. Simple but essential as agents move into real workflows.
Microsoft’s web agent ditches clicking for code
Webwright generates reusable Playwright scripts instead of clicking around like a human. Hit 60.1% on the Odysseys benchmark, up from 33.5% with basic GPT-5.4. The terminal-native approach feels right, even if 1,000 lines of code sounds optimistic for real websites.
The intelligence argument heats up
Hassabis reckons we’re “in the foothills of the singularity” whilst LeCun says current AI isn’t actually intelligent. Meanwhile, George Hotz warns coding agents will be “one of the most costly mistakes” because they create subtle bugs. The industry’s still working out what these systems can actually do reliably.