Morning Edition · Friday, October 2, 2026
Tech
OpenAI Fires Three Safety Researchers Over Alleged Leak to an Outside Safety Group
The dismissals land three days after OpenAI's president signed a White House accord promising internal verification teams and external audits of frontier models, with no enforcement attached.
Tech
Google's Gemini 4 Argon Raises Output to One Million Tokens and Leads 12 of 18 Disclosed Benchmarks
Argon scores 77.9% on the DeepSWE v1.1 long-horizon coding suite but trails GPT-6 Astra by 10.5 points on the harder FrontierSWE v2, and general access is gated behind a vetted-defender program.
Tech
OpenAI Ships an Ultrafast Serving Mode for GPT-6 Astra, Claiming Up to 8x Faster Token Generation
The speedup comes from inference software OpenAI optimized with its own internal models on Nvidia Blackwell hardware, and it is live in the API at low rate limits.
Markets
Trump Rules Out Nationalizing AI Labs and Floats Intel-Style Equity Stakes Instead
The president pointed to the government's roughly 10% position in Intel as the template for OpenAI, Anthropic and xAI, while Anthropic is reported to be absent from the talks.
Tech
The Agent Harness Is Becoming Infrastructure, With DeepSeek and Earendil Both Shipping Open Runtimes
DeepSeek's MIT-licensed harness treats models, tools and sandboxes as swappable plugins, while Earendil's Pi reached a stable 1.0 with native Model Context Protocol support and a durable runtime for long-lived agents.
Markets
Nvidia Puts a Price on AI Factory Capacity: Roughly $60 Million per Megawatt
The company's return-on-investment argument rests on keeping expensive capacity busy with mixed workloads, a case it is making as data-center capex faces harder questions.
Tech
Bilibili Open-Sources Index-Translate, a Qwen3.5-Based Translation Family Covering 150 Languages
The Apache-2.0 release spans 2B, 9B and a preview 35B mixture-of-experts model, plus variants for voice-preserving dubbing, syllable-matched subtitles and document-length consistency.
Tech
Reading an LLM Judge's First Token Inflates Position Bias by 42 Points, Paper Finds
Three Qwen3 judges failed to lead with a verdict token on 12% to 49% of pairs, and the cheap logit readout (scored from the model's raw output probabilities before any text is generated) distorts bias measurements while leaving accuracy nearly unchanged.
Tech
Two Papers Try to Make Agent Authority Auditable Before Self-Evolving Agents Reach Production
One proposes a harness that separates proposing an action from proving its external effect, the other maps how safety properties break when agents keep rewriting their own memories, tools and skills.
Tech
Berkeley Lab Cuts Beamline Image Analysis From a Month to About 15 Minutes With Meta's Open Vision Models
Fine-tuned SAM 3 and DINOv3 run across 300 A100 GPUs inside a Department of Energy project, enabling 3D reconstruction while experiments are still running.
Tech
An MIT Machine-Learning Loop Found an mRNA Formulation That Survives a Year at Room Temperature
The algorithm screened about 50 approved excipients and proposed ratios the team tested two at a time, producing solid-state lipid nanoparticles that kept full bioactivity after more than two months at 37 degrees Celsius.
Tech
A Deployability Study Puts Nine On-Device Entity Recognizers Against Latency and Output Validity
The comparison spans spaCy, GLiNER encoders and locally run Qwen3 and DeepSeek-R1 models from 13 million to 8 billion parameters, scored partly against labels built by a cross-family judge panel.