The Brief
New briefs daily around 7 AM Eastern
Sandbox Windows agents with OS-enforced file and network policies
Microsoft made Execution Containers generally available on Windows. Admins and developers declare which files and network destinations an agent may use, and the operating system enforces that policy at runtime.
- 01Read
Nadella says to contain frontier models like insider risks
Microsoft CEO Satya Nadella argued that controls over what a model can access and do must sit outside the model and its harness, with tamper-proof logs, independent audit, model diversity and an emergency stop.
- 02Read
Sakana's MASS trains a 27B model on its own best agent teams
Sakana AI and UC Berkeley researchers had one 27B open-weights model run as a team of subagents, pick its best team runs by self-judgment and train on them. After two cycles, score per output token rose 1.2-1.6x.
- 03Read
ElevenLabs ships Architect alpha to diagnose and fix voice agents
ElevenLabs released ElevenAgents Architect in alpha, a built-in assistant that analyzes transcripts, failing tests and Spotlight findings, proposes configuration changes, writes simulations and saves versioned drafts for approval.
- 04Watch
Akamai demo: speculative decoding cut vLLM speed from 58 to 16 tok/s
In an AI Engineer workshop, Akamai's Du'an Lightfoot and Khaja Omer showed speculative decoding dropping a vLLM-served model from about 58 to 16 tokens per second, and used it to argue for measuring every serving change.
- 05Read
Karpathy says the gap in perceived LLM capability is widening
Andrej Karpathy described a funnel of LLM experience, from billions with little contact to a few thousand with internal frontier access, and tied the widening perception gap to problem ambition, jagged capability and tiered access.
A dated brief from the vibe-coding frontier. Today’s Intel.