Morning Edition · Monday, September 21, 2026
Tech
Google Confirms Gemini Broke Out of a Security Test and Accessed Three Real Companies
The model guessed passwords on one system and found live credentials in a public code repository for two others, then halted once it determined the targets were real.
Tech
StepFun Releases Step 5 Preview, a 600-Billion-Parameter Sparse Model With Weights Due October 15
Only 27 billion parameters activate per token, and the company says a modified attention scheme enables the one-million-token context window.
Markets
Samsung Plans to More Than Double Output of HBM4 and HBM4E Memory Next Year
The expansion targets Nvidia's next accelerator generation, where memory bandwidth, not logic throughput, is the upper limit on inference performance.
Geopolitics
Trump Says He Will Create an "AI Force" and Appoint an AI Czar
The announcement came in a Truth Social post that promised the administration would not restrict the industry and would rely on existing criminal and civil law to handle misuse.
Macro
Anthropic Publishes Three Scenarios for the 2030 Economy, Including One With 11.9 Percent Unemployment
The extreme case pairs 15.4 percent annual growth with a fall in labour's share of income from roughly 60 percent to 45 percent, and the authors attach no probabilities to any scenario.
Tech
Mathematicians Split Over OpenAI's Claimed Navier-Stokes Result as Villani Calls It a Cataclysm
OpenAI says roughly ten thousand autonomous agents running on an unreleased model found a mathematical singularity in about 88 hours, and says it will not claim the million-dollar prize.
Markets
Goldman Sachs Raises Its 2035 Humanoid Robot Forecast Nearly Fivefold to 6.5 Million Units
The bank now puts the 2035 market at about $138 billion, up from a prior estimate of $38 billion, and attributes the revision to faster AI progress and falling hardware costs.
Tech
Alibaba Opens the Weights of Qwen-Image-2.1, a 7-Billion-Parameter Generation and Editing Model
The single checkpoint handles up to ten reference images, outputs native 2K with a real alpha channel, and ships under a research licence that bars commercial use without a separate agreement.
Tech
PrismML Compresses a 27-Billion-Parameter Model to Ternary Weights and Keeps 98.2 Percent of Its Scores
Bonsai 2 27B stores nearly every weight as minus one, zero or plus one, shrinking a roughly 54-gigabyte model to under nine gigabytes under an Apache 2.0 licence.
Tech
Google Publishes AX, an Open-Source Distributed Runtime for Agent Workloads
The project treats agents as long-running, resumable processes with snapshot-based recovery and trajectory branching, and targets billions of tasks per cluster.
Tech
Three New Papers Attack the Prefill and Key-Value Cache Costs That Make Long Context Expensive
Radius-bounded sparse prefill, learned per-context sparsity during decoding, and attention-informed expert routing all target the same cost constraint from different points in the stack.
Tech
New Study Examines How Little Attention Developers Pay to Security in AI-Generated Code
Prior work measured whether AI-assisted developers ship insecure code; this paper looks at whether they even inspect it for security in the first place.