The Pulse — March 26, 2026

This is a concrete example of “post-training NAS + distillation + RL” being used to reshape a big base model into a cheaper/faster production model while keeping quality—especially targeting KV-cache + long-context bottlenecks.

KV-cache bandwidth/memory is increasingly the real limiter for long-context + reasoning workloads. TurboQuant frames a path to aggressive KV compression without finetuning, with reported large speedups on H100.

This is signal that “computer-use” (agents operating real UIs) is now a core frontier, not a demo feature. They’re hiring for perception/interaction talent and explicitly measuring progress against OSWorld.

Benchmarks are shifting from “can you answer?” to “can you learn efficiently over time with sparse feedback?” That’s closer to real agent work: exploration, memory, adaptation.

This is part of the emerging “behavioral API contract” layer: explicit chain-of-command, hard rules vs defaults, and public legibility for steering + governance.

As agents do computer-use + browse + ingest docs, they inevitably pull untrusted content into context. This piece argues “guardrails” become security theatre unless you restructure the system (permissions, human approvals, or code generation).

AI buildouts are hitting physical-world constraints (power, permitting, moratoriums). That upstream constraint can reshuffle winners (on-device, quantization, efficiency).

Creators are showcasing UI-acting agents; upstream, Anthropic is explicitly optimizing OSWorld and warning about prompt injection.

“Hacking” here often means bypassing product constraints. This is exactly where the upstream safety model (prompt injection, permissions, external side effects) becomes relevant.

There’s a continuing open-source scramble around “cowork/agent shells” that sit between a model and real-world tools. Worth watching for the interfaces that win (tool protocol, logging, permissions, audit).