Google released Gemini 4 Argon on 30 September, calling it its most powerful model yet. The frontier model is going only to a select group of cyber partners through its Fairwind Program, according to TechCrunch.
BMW opened US orders on 30 September for its first electric 3 Series, the 2027 i3 50 xDrive at $61,500, which undercuts the petrol M350 xDrive by $4,400 before a $1,350 destination charge.
OpenAI said on 1 October that it shut down a campaign to extract the hidden reasoning behind its models, and that the same technique kept working for weeks on Microsoft Azure. The same day brought three new open or open-weight model releases and a fresh accusation that frontier safeguards do not travel with model weights.
Nvidia launched its Open Agent Safety Platform on 28 September, an open software stack plus BlueField-4 DPU hardware that can quarantine misbehaving AI agents in milliseconds, according to Tom's Hardware. The same week, Spanish researchers completed 10,000 charge-discharge cycles on an iron-lead flow battery, and a Texas A&M trial found creatine helped middle-aged adults gain muscle without exercise.
On 1 October, OpenAI's Dots personal agents, restricted to ChatGPT Pro subscribers, collided with Meta's free Muse while Nvidia pushed an open agent safety platform and Glow Security revealed more than 13,000 internal screenshots leaked by coding agents at 343 organizations.
OpenAI said on 1 October that it disrupted a coordinated campaign to extract reasoning from its models, attributing a core cluster of the activity to individuals associated with China's Moonshot AI, the developer of Kimi. The company said the activity surged to 16,000 requests from more than 4,000 users over two days in late July.
OpenAI has parted ways with three researchers on its safety team for allegedly sharing confidential company information with a third party AI safety organization, The Wall Street Journal reported on Thursday 1 October.
A 145-million-parameter model pretrained by volunteers on donated hardware went live on 30 September, the same week Google restricted its most advanced model to vetted cyber partners and Amazon open-sourced a 2B decision model.
OpenAI says it shut down an adversarial distillation campaign that pulled hidden reasoning out of its models, but independent researchers found the same method still worked on Microsoft Azure as late as 13 September, according to a study update published on 1 October.
OpenAI has parted ways with three members of its safety team for allegedly sharing confidential company information with a third-party AI safety organization, the company confirmed on Thursday.
Open-weight releases kept arriving through Tuesday and into Wednesday, from a 145M-parameter model trained by volunteers to a 9B classifier built in a day, while the FTC opened its first enforcement action into rogue AI agents.
Modal and Carnegie Mellon's Full Stack Data Lab said on 30 September that their Quail engine processed over a billion tokens per minute on a single H100 GPU, more than 10x their vLLM baseline, and priced that work at under 6 cents per billion tokens on Modal.
OpenAI told reporters on Tuesday it will not go public until it can "make confident safety decisions", while researchers gathering in New York on 30 September argued that no company should be the last word on whether a model is safe.
OpenClaw opened the code for an enterprise control plane for persistent AI agents on 30 September, the same day two separate open source projects published deterministic guardrails and an agent identity protocol. The three launches point at the same gap: agents are easy to run, hard to audit.
Nvidia published Kumo Tabular, an open foundation model for tabular prediction, in a Hugging Face post dated 29 September, while Fermion Research released the 164 MB speech model Phonon-2 a day later. Both arrive as the FTC opens an industry-wide investigation into AI labs and OpenAI holds back its own IPO.
Modal and Carnegie Mellon's Full Stack Data Lab published Quail on 30 September. The inference engine plans SQL queries and schedules LLM execution together, and on a single H100 it clears one billion tokens per minute, more than 10x a vLLM baseline.
Two open source projects published on 30 September claim the same prize from different directions. Quail reports over a billion tokens processed per minute on a single H100 by planning SQL queries against the inference engine. Magnitude says it runs open models up to 2x faster than llama.cpp by compiling kernels on the user's own hardware.
Fermion Research released Phonon-2 on 30 September, an open speech recognition model that fits in a 164 MB download and averages 5.21% word error across seven English sets on the Open ASR Leaderboard. Within hours, Nvidia's Kumo Tabular and a volunteer-trained 145M-parameter language model landed alongside it.