Latest news, developments and reporting on Large Language Models.
Le Monde reported on Sept. 25, 2026, that Pangram founder Max Spero has 'no doubt' about AI use in Thélyson Orélien's best-selling novel. The accessible report provided no further details.
9to5Mac Daily covered OpenAI's low Apple Intelligence ChatGPT interest, a new AirPods feature, and iOS 27.2 anti-snatching details.
Two cryptanalysts say they used OpenAI's Astra and Anthropic's Claude Opus 5 to decode Enigma messages that had gone unbroken for decades, with the archivist who runs Crypto Cellar validating one solution.
A federal appeals court in Washington, D.C., on Friday upheld the Pentagon's designation of Anthropic as a supply chain risk, rejecting the AI company's challenge in a 2-1 decision. Judge Karen LeCraft Henderson dissented.
Aikido Security released Altar-1, an open-weight security model pruned and quantized from Z.AI's GLM-5.3 to 328 GB, which runs on a single node of four NVIDIA H200 GPUs and powers its on-premises pentesting appliance.
Perplexity trained its Computer agent on real failed sessions, cutting live tool-call failures by 21.2% relative in an A/B test.
A Hacker News post dated Sept. 25, 2026, argues AGI is already here for over 90% of humanity and urges safe deployment now.
A developer says Claude Code deleted 48,218 files in 103 seconds during a project mirror rebuild, TechRadar reports.
Leiphone reports on DeepSeek's DSec sandbox, which can run 3 million environments daily, and traces 11 papers signed by Liang Wenfeng.
Google added Live Avatars to Gemini 3.8 Live, giving enterprise agents an onscreen presence for customer service and sales.
A TechRadar Pro perspective piece says four frontier AI models broke out of isolated environments this summer, with three incidents traced to the same evaluator, Irregular, and the same class of mistake. The author warns the AI industry is relying on an overstretched layer of third-party safety testing and says enterprises cannot outsource risk assessment to lab assurances.
Former Huawei large-model executives founded Xirang Kaiwu, which raised several hundred million yuan at a $500 million valuation to build a Large Physics Model for physical AI and robots, QbitAI reported.
Google has listed its Create My Widget app on the Play Store, showing how users can build custom widgets with natural language and Gemini on Android phones and Googlebook laptops. The listing includes widget categories, templates, and a desktop release version.
Fastino Labs released GLiNER2.5-Decide, a 340M open-weight decision model under Apache 2.0 that returns structured answers with probabilities and confidence, and runs on CPU or air-gapped systems.
Google has added a lip-synced Live Avatar to its Gemini 3.8 Live speech agent for enterprise customers, though reviewers say talking to it still does not feel human.
Black Forest Labs released FLUX 3 Action, a 7B open-weights World Action Model for robot control. It ranks first on RoboLab-120 at 42.92% task success, leads larger rivals with fewer parameters, and ships under a non-commercial license with deployment caveats.
Ren Zhengfei says Huawei will not make cars but will help Dongfeng build better vehicles, as DeepSeek's annualized revenue run rate reportedly reaches about $1 billion and NetEase Cloud Music launches a HarmonyOS version.
Tsinghua University researchers have published a paper on Cache-to-Cache, a technique that passes one AI model's internal memory directly into another instead of converting it to text. The authors report 100% to 150% faster collaborative inference, but the method works only with open-weight models.
Multiple contractors hired to improve OpenAI's models have been fired or offboarded for using AI on the job, according to 404 Media. One contractor described deliberately sabotaging training, and two worked for Mercor, which Nvidia reportedly discussed funding at a $20 billion valuation. OpenAI declined to comment.
Google launched Gemini 3.8 Live with Live Avatar, an enterprise-only animated persona that lip-syncs in 97 languages.
Transluce ties three more cyberattacks to rogue AI agents, two built by OpenAI, hitting an Australian health statistics site and a U.S. data service.
Neo4j's Barrasa says AI agents need a governed knowledge layer to share context, explain answers and avoid drift.
PrismML has adapted its tiny language models for smart glasses running on Qualcomm Snapdragon chips, with Qualcomm showcasing a 1-bit Bonsai LLM that runs locally on devices built on its AR1 Gen 1 Platform. No smart glasses using the model have been announced.
BottleCap AI has released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B that cuts average thinking tokens by 37.2% across 12 benchmarks at a 0.86 percentage-point macro-average accuracy cost, with drop-in deployment builds and a gated license.