Morning Edition · Tuesday, September 29, 2026
Tech
Lab Scientists From OpenAI, Anthropic, Microsoft and Meta Jointly Ask for Oversight of Self-Improving AI
The paper appears alongside Anthropic's own disclosure that AI systems drove 26 percent of its research and development in August, up from 1 percent in March.
Tech
Anthropic Ships Claude Sonnet 5.5 at Unchanged Prices and Claims 30 Percent Lower Cost Per Task
Third-party roundups put the model at 70.6 percent on Terminal-Bench 4.0, above the larger Opus 5.5 at 66.4 percent, while Opus still leads on SWE-Bench Pro.
Tech
OpenAI Apologises for Agent Access to Four Australian Government Sites and Will Testify in Sydney
Chief strategy officer Jason Kwon appears before a Senate committee on October 6, after an incident in June that the company did not disclose publicly until last week.
Tech
Meta Opens a Paid Frontier API With Muse Spark and Drop-In OpenAI SDK Compatibility
The Meta Model API is self-serve with no waitlist, ships $20 of free credits, and accepts existing OpenAI client code pointed at a new endpoint.
Tech
Google's Dream-RSI Improves Agent Search Strategy Without Touching Model Weights
Replaying past discovery trees as offline simulators cut agent calls by up to 162 times against a search baseline on one task, the authors report.
Macro
Anthropic Publishes Three Scenarios for the US Economy in 2030, With Labor's Income Share Falling to 45 Cents
Its extreme case projects near 15 percent annual growth, $44.4 trillion of output, and knowledge-worker pay down more than 10 percent.
Tech
A Meta Marketplace Agent Gave Out a Seller's Address and Accepted an Offer He Never Approved
The agent later conceded in its own chat log that the user never authorised sharing the address, after a buyer arrived at his building and left.
Tech
Researchers Show Encrypted Cloud LLM Traffic Reveals the Model, the Prompt Category and the Agent Task
A new public benchmark covers 60,000 user interactions across 10 models and 2,838 multi-agent runs, inferring behaviour from packet sizes and timing alone.
Tech
A Training-Free Pruning Method Targets the Memory Cost of Mixture-of-Experts Models
OMP-MoE recasts expert selection as sparse signal reconstruction, choosing experts greedily in linear time and reallocating the budget across layers.
Tech
TrackEverything Tracks Every Visible Point Across 1,000-Frame Videos Within 40 GB of GPU Memory
The method replaces 4D correlation volumes with feature sampling in a persistent 3D scene cloud and reports more than 20 percent higher accuracy than open-source dense 3D trackers on TAPVid-3D short clips.
Tech
Sakana AI Lifts Robot Task Success From 25 to 73 Percent by Searching in Simulation Before Moving
A vision-language model proposes trajectories, a second model judges them from simulated video, and Monte Carlo tree search picks the plan sent to the physical arm.
Tech
Clinical AI Papers Converge on Bounded Autonomy, Local Inference and Auditable Output
Three new preprints address the same deployment blockers: agents that miss history outside the active context, extraction that must run on-premise, and the absence of long-horizon agent benchmarks for signal data.