Hyunkook Kim
Google Research on September 15 published Retrieve-for-Train, a framework that returns search results as one coherent set, alongside its ICML 2026 paper. Running reinforcement learning once to teach a small diffusion model made query fan-out 12 to 20 times faster.
Anthropic's developer account posted a 2-minute-7-second founder interview video shot at Frontier Day in San Francisco on September 16. Four companies, Spawn, Hex, Sazabi and Phylo, describe what building a company on Claude looks like right now.
Google published a 26-page crisis resilience white paper on September 15 under the names of Yossi Matias, vice president of Google Research, and Kate Brandt, chief sustainability officer. It gathers the company's AI disaster-prediction work in one volume, from flood forecasts seven days out and cyclone forecasts 15 days out to cases where warnings turned into cash assistance and paused loan repayments.
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15. The two voice models run tools in the background mid-conversation and reason while they speak, and Extended Thinking took first place on the Speech to Speech Quality Index with a score of 82.6.
Meta opened Meta One, a subscription service bundling Instagram, Facebook, WhatsApp, and Meta AI, worldwide on September 15. The individual bundles raise AI image and video generation limits, while the creator and business tiers sell follow buttons and a business agent for a monthly fee.
Anthropic opened Salesforce in Claude as a public beta on September 15. The plugin pulls accounts and pipeline into Claude with 37 sales skills built in, and it is the first product of the Claudeforce partnership announced on August 26.
Bilal Chughtai, who worked on AGI safety and alignment research at Google DeepMind, warned of AI risk in a resignation post published on X and LinkedIn on September 14. He wrote that AI has the potential to kill us all, called for coordination and transparency to end the race between companies, and the post was read 560,000 times within a day.
The Tencent Hunyuan team released the EvolveScaler paper and project page on September 15, an information-evolution data framework. On data built by defining state in code first and then rendering it as natural language, the median avg@5 across 14 models fell to 11.3 on the longest tier.
On September 14 Nous Research announced a team tier for Hermes Agent, Hermes Business, and an on-premises edition, Hermes Enterprise. On top of a shared balance and per-member caps sits a skill library that the whole team uses, no matter who built each skill.
Microsoft Research featured a paper analyzing real-world GitHub Copilot coding-agent traces in its September 14 Research Focus. Of 761 million LLM calls, 87% were started by the agent itself, and a single tool failure could inflate compute by up to 4x.
Cohere CEO Aidan Gomez said live on Bloomberg TV on September 14 that the existential-risk debate around AI has veered too far into science fiction. He cited the Hugging Face hacking incident as an example of over-extrapolation and called it a red flag for the largest companies to set the rules among themselves.
Parse 5, the document-parsing vision language model Cohere released on August 27, scores 79.2 on ParseBench, below GPT-5.5 and Opus 4.8. The company published that table itself and led instead with throughput and cost: 2,160 pages a minute on eight H100s.
