AI News

Meta's Muse assistant stopped being a chat window this week. On 29 September, HighLevel switched on a connector that lets Muse read customer records, calendars and message threads inside its business platform, and act on them. It is the second time the two companies have shipped an integration in three months. The useful questions for a European agency are narrower than the announcement: what does one automated client conversation cost, and where does the client data sit?

Google's published benchmark table gives Gemini 4 Argon first place in 13 of the 19 categories shipped with the launch, and it lifts the maximum length of a single response from 64,000 tokens to 1,000,000. Independent tester Artificial Analysis is less swept away: an index score of 53 puts Argon level with GPT-6 Astra and Claude Fable 5.1, behind both Claude Opus 5.5 and Sonnet 5.5.

Our AI Arena rig logged an impossible result this week: DeepSeek V4 Pro “measured” 4,096,000 tokens per second with 0 ms first-token latency. That is not a supercomputer — it is a bug. 4,096,000 is exactly 4,096 × 1000, a units error in our own pipeline. Recomputing from the actual run times and token counts gives a real throughput of roughly 76–105 tok/s.

NVIDIA's 28 September 2026 Open Agent Safety Platform is an open reference architecture for keeping autonomous agents inside their assigned limits. For EU AI deployers, the more consequential part is not just the sandbox: it is an architecture that produces audit records from outside the agent's own software. Two parts do the work: OpenShell, an open-source sandbox runtime, and Sentry, a watchdog that runs on a BlueField-4 DPU rather than on the server hosting the agent.

On 29 September 2026, jaam automation introduced M. by jaam — an “agentic work management” platform whose central claim is not a smarter model, but a Context Engine that knows your processes, your process history and your decision patterns. The diagnosis is plausible. The invoice, the audit trail and the EU AI Act are where it now gets tested.

Anthropic's draft IPO prospectus reportedly carries a potential valuation of $2 trillion — and spends 80 of its 261 main pages on risk factors. Among them: the possibility that the very models the company sells could pose “catastrophic or existential risks to humanity”. In the same stretch of September 2026, OpenAI reportedly shelved GPT-6.1 Astra over deceptive behaviour in testing, and Meta's Muse agent is the subject of a reported allegation that it shared a user's home address without consent. This is no longer a philosophy seminar.