01OpenAI Quietly Lists a Second GPT-Rosalind Model for Life Sciences [AI Models & Research]A new gpt-rosalind-discovery model has appeared on OpenAI's API pricing page with no announcement — no blog post, help article or system card. It lists at the same $5 per million input / $25 output as gpt-rosalind-research, with the name pointing at drug discovery; access remains gated behind the trusted-access program. → source
02AI Models Can Provably Hide Arbitrary Capabilities — and Evals May Never Find Them [AI Models & Research]Capability evaluations assume dangerous abilities can be elicited with the right prompt. New backdoor attacks embed encrypted neural modules that decrypt only under adversarial conditions — proving the assumption fails, and that a clean evaluation cannot rule out capabilities a model is deliberately hiding. → source
03X-AgentRed: Red-Teaming Personal Agents Across Multi-Service Ecosystems [AI Models & Research]Researchers built a red-teaming framework for personal agents that use tools across services: benign user requests still compose into harmful outcomes once an agent's actions accumulate across apps and services. The benchmark sequences multi-step compositional attacks across evolving multi-service ecosystems. → source
04RATED: a New Benchmark Judges Agent Tool Retrieval by Execution, Not Semantics [AI Tools & Ecosystem]Most tool-retrieval research optimizes semantic similarity; RATED optimizes whether the chosen tool actually executes. Built on execution dynamics — availability and failure modes during runs — the benchmark pushes agent infrastructure to prefer the tool that works over the tool that reads best. → source
05OpenAI Adds Composer Predictions to Codex — the Agent Now Anticipates Your Next Change [AI Tools & Ecosystem]The October 9 release note quietly ships Composer predictions to Codex — the coding agent now anticipates your next change and pre-drafts it. Small on its face, but it marks the moment coding tools shift from responding to predicting. → source
06Voice AI Hasn't Had Its ChatGPT Moment Yet — so Say the Execs at Cerebral Valley [AI Applications & Industry]At the Cerebral Valley Voice Summit, executives argued voice AI still lacks the lightning-in-a-bottle consumer moment — accuracy, trust and disclosure (people must be told they're talking to an AI) are the blockers. Otter's CMO notes transcription was never the endpoint: everything downstream fails when the ASR is wrong. → source
07A Week After Amazon's NDA Reversal, AI Data-Center Builders Bet Transparency Can Soften Opposition [AI Applications & Industry]TechCrunch's new question goes beyond Amazon's own reversal: now that NDAs are being dropped, will open books on power, water and tax breaks actually persuade the communities that stopped believing the sales pitch? The piece reads as the industry's bet that transparency, not more announcements, defuses the moratorium wave. → source
08New Zealand's Greens Run on AI Regulation: Moratoriums, Labeling and Data Sovereignty [AI Applications & Industry]Co-leader Chlöe Swarbrick launched an election platform covering data sovereignty, labeling for AI content, an AI office inside the Prime Minister's department, algorithm audits for high-risk systems, and a one-year moratorium on new data centers — framing regulation as why government exists, not a brake on inevitability. → source
09Salesforce Renames AIForce to SIForce — the Super-Intelligence Rebrand Reaches the Fortune 500 [AI Applications & Industry]Marc Benioff renamed Salesforce's agent platform hours after Trump's post calling anyone still saying 'artificial intelligence' THE ENEMY — following the September 29 executive order and Elon Musk's SpaceXSI. As Business Insider noted, renaming a product doesn't make it superintelligent; the rebrand just tracks the vocabulary. → source
10Will the HomeView Be Apple's First Smart Home Success? [AI Applications & Industry]TechCrunch asks whether Apple's HomeView can finally make the smart home click: AI-assisted control in a HomePod-centric package, after a decade of fragmented hubs and assistants that cannot talk to each other. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.