- Policy & legal
OpenAI is facing state lawsuits, EU information requests and an FTC subpoena in the same week. Eight US states sued OpenAI, alleging it made unfounded safety claims and misused children's data, and the European Commission's AI Office sent requests for information on security, external evaluation and model use monitoring.4 Axios also reported that the FTC subpoenaed OpenAI and Anthropic over consumer harms, and a tally of filings counts 13 new complaints against OpenAI from the past week.5
- Dev tools
NVIDIA has released the full Run:ai GPU orchestration platform as open source under Apache 2.0. The release covers the control plane, cluster services, CLI and an MCP server, plus four components that can be used on their own: KAI Scheduler, a GPU Fractioning Operator, KAI Resource Management and Workload Map.2 NVIDIA's NGC community edition requires no subscription but is community-supported, and paid enterprise support remains a separate option.6, 2
- Industry
OpenAI, Google and Workday are each pitching enterprises on running AI agents inside their own platforms. In a little over two weeks, all three companies gave corporate customers the same answer about where their agents should live: on their platform.7 OpenAI went first on September 29 with an offer aimed at large customers.7
- Models
Mistral AI has released Timestral, its first family of open-weight time series forecasting models. The four models range from 5M to 209M parameters and run on standard server CPUs, and Mistral reports that the largest ranks first among zero-shot models on GIFT-Eval, TIME and fev-bench.8 The weights use a modified MIT license, and fine-tuning on private data runs through Mistral Forge.8
- Models
GoTo has released an open-source 27-billion-parameter model for Bahasa Indonesia and regional languages. GoTo announced the model at Bali Tech Week and said it can run on institutions' own servers.9 It was trained on data from media partners and public-sector websites, and Universitas Sumatera Utara helped gather Batak-language data.9
- Models
Birla AI Labs has launched Yug-1, a time series foundation model built in India. The Aditya Birla Group venture says Yug-1 outperformed other first-generation time series models on GIFT-Eval and is available on GitHub and Hugging Face.10 The company says its team went from training to deployment in under six months.10
- Research
Security experts are divided on how seriously to take Anthropic's self-replicating AI worm research. The Anthropic red team's program spread across 15 smart TVs and used a little computing power from each, and some experts warned that weakly patched IoT devices would make such worms hard to contain.11 Others called it a stunt, saying the worm moves slowly and basic measures such as strong passwords and two-factor authentication would stop most attacks.11
- Dev tools
Google has open-sourced Harness2, a framework that improves an agent's harness while the model stays frozen. Harness2 edits prompts, tools, memory and other harness components for each task through a Propose, Probe, Compose step, without ground-truth verifiers, and keeps lessons that guide later tasks.12, 3 The paper reports gains of up to 11.5 points on WorkBuddy-Bench with GLM-5 and Gemini 3.5 Flash.13, 3
- Policy & legal
Meta's disputed crawling clause still appears in its Australian commercial terms. Mumbrella reports that the clause allowing Meta to crawl business websites and use the content for AI training appears to remain in Australia, although Meta has reportedly removed it in some jurisdictions.14 The terms take effect October 30, and Meta has said the changes only consolidate existing provisions.14
- Infrastructure
An Arm-backed study models nearly twice AMD's rack-level performance for Arm's AGI CPU on agentic workloads. Signal65 found an average of 1.99x modeled performance per rack against AMD's EPYC 9755 across seven workloads, mainly because more Arm servers fit within a 24 kW power budget.15, 16 Arm developed the measurement method with Fujitsu, Meta and OpenAI, and testing on final production hardware is still to come.17, 16
- Products
YouTube is testing Ask YouTube AI answers at the top of standard search results. Signed-in desktop users in the experiment who search in English in the US may see an AI overview for open-ended or how-to queries, with an option to ask follow-up questions.18 The test comes a week after Ask YouTube became available to all US users.18
- Infrastructure
Samsung will reportedly source base dies for Nvidia's NVHBM memory from TSMC. Seoul Economic Daily reported that Samsung chose TSMC's 3nm process after concluding that yields on its own SF2 process fell short of Nvidia's requirements.19 NVHBM moves the memory controller into the HBM base die, and SK Hynix and Micron are also expected to use TSMC for those dies.19
- Products
Microsoft's new Copilot app now includes Word and Excel. The change was among several business software updates in a weekly roundup that also covered Intuit's AI platform, SAP's AI offerings and Gusto's Claude connector.20
- Products
An AI geolocation tool now says it can identify a specific house from a single interior photo. Raven, previously called Geospy, offers the capability to police, according to 404 Media's reporting.21
