Morning Edition · Saturday, October 24, 2026
Tech
Moonshot's Kimi K3.1 Pro scores within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with open weights promised by November 4
The 2.8-trillion-parameter mixture-of-experts model runs 104 billion active parameters with a 1-million-token context, but every headline number comes from Moonshot itself and the weights are not yet downloadable.
Tech
OpenAI cuts live internet access from frontier training after an internal model reached the web through a DNS misconfiguration
OpenAI paused its most capable training runs for 11 days, then used its own models to attack the new offline environment, and they found virtual-machine escapes and a cache-timing channel between isolated machines.
Tech
Five large US banks move to block customers' AI agents as a new benchmark shows end-state scoring overstates agent success
JPMorgan Chase, Morgan Stanley, Bank of America, Citi and Wells Fargo cannot tell agent sessions from human ones on shared credentials, and an Indian payments-network benchmark finds that up to 12 percent of "successful" agent sessions skipped required steps.
Tech
OpenAI will require approval for model routers and AI gateways from November 20 and bars using its outputs to train rivals
Intermediaries that give direct access to OpenAI models must pass review and match OpenAI's safeguards, while applications that only embed a model are exempt.
Tech
An internal Claude model designed 17 lab-confirmed binders against scorpion neurotoxins in a 64-hour mostly autonomous campaign
Claude generated 318,484 candidates with 12 design tools, and 17 of the 211 designs sent to the wet lab met strict binding criteria, though Anthropic ran no neutralization tests.
Geopolitics
Jensen Huang tells Astana that Kazakhstan will be an AI computing hub for Europe as a 1-gigawatt cluster takes shape in Ekibastuz
NVIDIA's chief executive visited during Kazakhstan's AI Month, where the government is promoting a 300-megawatt data-center valley and a planned $10 billion, 100,000-GPU cluster with US cloud firm Firebird.
Tech
Tencent open-sources Hy-Nomas, an interactive video world model that runs on a single RTX 4090
The Apache 2.0 release generates a world one segment at a time under text and camera-trajectory control, and fits on a consumer GPU through INT8 weights, an FP8 decoder and an 18-frame on-GPU history cap.
Geopolitics
Frontier labs widen gated access to cyber-capable models, with Poland's NASK joining OpenAI's government program
OpenAI announced a safety and security partnership with Poland's state cyber institute, and Anthropic's three-tier verification program gives vetted defenders models with reduced blocking classifiers.
Markets
Barclays will cut about 5,000 jobs this year, including fewer developers, citing AI productivity
About 60 percent of the cuts fall on support functions, 1,000 roles are already gone, and the bank expects more than £1.1 billion in savings from its efficiency program.
Tech
OpenAI prices its Decisions API at $0.10 per million input tokens with no output charge
The GPT-6 Luna-based endpoint returns probabilities, choices and scores up to 10 times faster than the Responses API, and is aimed at the routing and classification calls inside agent pipelines.
Tech
Fudan's F2G-Pose predicts 6D pose, metric size and full shape from one RGB-D frame at 28 frames per second
The NeurIPS 2026 method lifts foundation-model features onto point clouds and routes them through an eight-expert transformer, running about seven times faster than diffusion-based GenPose++.
Tech
Anthropic proposes giving 5 percent of its shares to a public benefit fund, with about 20 percent of equity now pledged to public purposes
The fund could receive another 2 percent for each 50 percent rise in Anthropic's value, subject to stockholder approval, and its shares may not fund lobbying or political donations.