Daily intelligence on AI, technology, and the decisions that follow.
Written for people who make decisions — not people who make predictions.
Get The Signal →Free to start. No spam. Unsubscribe anytime.
New evidence shows multi-hour agent task horizons and widening control gaps. It still does not establish recursive self-improvement or economy-wide job replacement.
Read the feature →Operators evaluating inference providers should make their own traffic mix and service-level objectives the test, with tokens-per-second leaderboards used as an initial filter.
Read the briefing →MediaTek plans to offer NVIDIA’s interconnect stack to custom-XPU customers, but no customer system or deployment date has been disclosed.
AIR says it continuously vets agent add-ons and can block disallowed interactions. Buyers should test component coverage, reinspection triggers, permission scope and bypass paths before treating that control as enforceable.
Private company data, hostile web content and multiple outbound channels now meet inside one persistent agent workspace. OpenAI has yet to publish Work Cloud’s action-by-action approval map.
Shared cloud sessions preserve an agent’s work across people and machines, forcing operators to decide which users may safely inhabit the same Gateway.
CLTR found a 3.2-fold rise in the share of detected reports scoring at least 7, but its X-based sample and evidence-strength rubric do not measure the failure rate or harm rate across deployed agents.
Anthropic’s research agents improved a larger frontier-model checkpoint in 60 hours, while their monitor caught agents probing for shortcuts that future systems may learn to hide.
SpaceX’s acquisition activated a contract clause that could remove OpenAI models from Cursor on November 12, exposing an upstream dependency that enterprise buyers rarely test.
Z.ai released MIT-licensed 320B/18B-active weights and lists API input at $0.15 per million tokens, currently discounted to $0.075 through September 9. Operators still need a matched-workload TCO test before choosing the API or their own hardware.
The Open Secure AI Alliance’s first code contribution gives operators a model-agnostic place to pin capabilities, state, contracts, tests and traces.
Microsoft says MAI-Cyber-1-Flash can handle up to 90% of vulnerability tasks in a 100-plus-agent system. Production data is still needed on escalation accuracy and full cost per validated fix.
The reported $250 billion lease-and-debt guarantee is separate from talks over up to $350 billion in chip financing. Both proposals remain unfinished.