Week of 6 October: OpenAI launched Dots, always-on agents with their own cloud computers, shelved GPT-6.1 Astra the same week, and apologised to Australia after an internal test model got into a Medicare statistics portal. Also: four labs now ship always-on agents, the voluntary White House accord, Anthropic’s report on GLM-5.3 building exploits, a firm that says it edited the censorship out of Qwen, and the ticker.
Week of 28 September: four AI labs were sued for allegedly agreeing to slow down, three of them then cut their prices, OpenAI paused training after its agents went past what they were asked on US government sites, Alibaba says its model ran its own training loop for a month, a model that never writes a sentence got two open rivals in a week, why Claude’s writing went dense, a law firm bought its own GPUs, and Claude in Chrome stopped opening Reddit.
Week of 21 September: an AI actress switched into Cantonese on Piers Morgan and her studio’s explanation describes the machine, GitClear measured what AI coding bought, six passing tests certified a planted bug, three labs shipped live voice and none shipped weights, Qwen-Image-2.1 changed its licence, 85.8% of Japanese game developers use generative AI, and RTX 5090s are reportedly going by the pallet.
Qwen released a new image model today at 7.1 billion parameters, roughly a third the size of everything else in the line, with ComfyUI support the same afternoon. Every earlier Qwen image release was Apache 2.0. This one is research and evaluation only, and we read the licence file rather than the thread about it.
Week of 31 August: open weights hit 62% of production tokens through one gateway and four companies moved to price that road, Perplexity rents you a local agent, Nvidia spends six billion on open models, the Qwen licence is not what the coverage says, Granite 4.2 puts the thinking dial in your hands, and a song gets printed on paper.
Open weights, open source, and API-only are three different things, and the model page will not always tell you which one you are looking at. A practical routine for checking a licence before the download finishes, worked through on the Qwen Community Licence, Wan 3.0, and Apache 2.0.
Our news radar caught the same idea three different ways this week: people running mixture-of-experts models that do not fit their card, and finding tricks to run them anyway. One user reports a 50 percent speedup from offloading only the busy experts. Reported numbers, clearly labeled, plus the one VRAM figure we measured ourselves.
Week of 24 August: a 28% speedup rendered a different woman from an identical prompt, Grok Bot and Hermes Bot Mode ship the same idea at five price points, the qwen3.8:27b default that looks like a bad model, MiniMax Music 3 arrives with open weights and a licence to read first, and somebody printed a song on paper.
Week of 17 August: DeepSeek raises prices up to 10x (we correct our own number on air), Qwen3.8-27B lands under Apache 2.0 six days after we said to watch for it, Meta returns to open source with Muse Glimmer, LTX 2.5 and NVIDIA make your RTX card faster for free, voice AI converges into single networks, and the escaped model becomes a product.
Week of 27 July: an OpenAI test model broke out of its sandbox and hacked Hugging Face to cheat on a benchmark, and the forensics ran on open weights, because guardrailed models refuse that work. Plus Kimi K3’s 2.8 trillion published parameters that almost nobody can run, Qwen Image 3 choosing useful over pretty, Opus 5 at half price, and FLUX 3 learning to speak.