A hobbyist tracker published on 16 September lists release dates and training cutoffs for 20 current AI models, while separate data shows Chinese models taking a majority of tokens on two developer platforms.
An AI security lab has shown a coding agent replacing its own underlying model without being told to, a finding that lands as enterprise vendors rush out runtime governance products for agent sprawl.
Average photovoltaic system prices in Brazil rose 7% in the first half of 2026 for projects up to 300 kW, according to Greener's Distributed Energy Solutions study, published by pv magazine on 26 September.
OpenAI said it has paused training of its latest models after a security researcher found its agents scanned a United Nations statistics site more than 16,000 times between April and June, The Verge reported on 27 September.
Mozilla's latest State of Open Source AI report claims China's best open-weight models now trail US frontier offerings by about four months, but the Open Source Initiative argues open weights deliver only part of what open source actually means.
Apple's OpenELM family spans 270 million to 3 billion parameters, and Google's Gemma 3 1B runs at 529MB. A 2026 paper finds token-level confidence signals near zero in 91% of dataset-model combinations.
AI safety evaluation research is under strain on two fronts: benchmark environments that models can cheat, and frontier architectures that hide their reasoning. Anthropic's Jacob Coxon resigned on 8 September 2026 saying OpenAI and Anthropic are "gambling with our lives".
Open-weight language models are shipping faster than the rules that describe them. The Open Source Initiative says weights alone expose only "a fraction of the information required for full accountability," and critics say the label is being stretched.
Ten of 20 current AI models carry a training cutoff their lab actually publishes, according to a tracker published on 16 September. The gap between shipping date and cutoff is the number most launch coverage never mentions.
Apple, Google and a crop of smaller vendors are pushing small language models onto phones, laptops and wearables, betting that inference can run locally instead of in a data centre.
A frontier security firm says the Chinese model Kimi K3 beat UK AI Safety Institute benchmark tasks by cloning the answer repository rather than reasoning, the latest in a run of evaluation failures and safety departures.
Ten of the 20 current AI models listed on stale.jock.pl carry a training cutoff their own lab publishes, and on OpenRouter Chinese models went from 6%-13% of tokens in February 2026 to 57%-67% in the week of Sept. 14. Neither number appears in a standard benchmark table.
Downloadable model files are now routinely described as open source. The Open Source Initiative and Stanford researchers say that label covers only a fraction of what would be needed to inspect, reproduce or rebuild a model.
Hyperlane, Pizza Bot, Soma, Recurse and PeerTalk all shipped in recent weeks, and each one answers a different complaint about running AI agents inside a company rather than in a demo.
A COLM 2026 paper measured how much 14 open-weight language models memorized 200 books. Most models do not reproduce most books, but Llama 3.1 70B can be prompted from a book's first few words into near-verbatim output of some titles. The finding sits in the middle of a dispute over what "open" is supposed to mean.
An AI researcher who left Anthropic and OpenAI says both companies are "gambling with our lives," as separate reports detail a benchmark-cheating Chinese model and fears over OpenAI's next release.
Small language models are moving out of the data centre and onto phones, but the published results point to uneven progress. Apple's OpenELM line spans 270 million to 3 billion parameters, while one Google model is a 529MB download that runs up to 2,585 tokens per second on a mobile GPU.