UpNext AI

Thrive Holdings’ ADT acquisition puts OpenAI-backed investment into home security, while Mistral Large 4 targets organization-controlled cyber defense and Meta’s Style Sketch builds outfits from Instagram photos. We also cover the Ramanujan Challenge’s AI-assisted proof evaluation, ChatGPT API credits, DeepSeek’s funding, Google’s earnings reaction, and Spawn’s Fable-powered coding agents. Thrive Holdings’ ADT deal: enterprise value of $15.3 billion and plans for a smart-home security business. https://techcrunch.com/2026/10/29/openai-backed-thrive-holdings-acquires-adt/ Mistral Large 4: downloadable model weights and a cybersecurity deployment pitch. https://techcrunch.com/2026/10/29/mistral-releases-weights-for-its-trillion-parameter-ai-model-targeting-cybersecurity/ Meta Style Sketch: a virtual closet and outfit suggestions using users’ clothes. https://techcrunch.com/2026/10/29/meta-quietly-launches-an-ai-fashion-app-called-style-sketch/ Ramanujan Challenge: research-level mathematics and an AI-assisted proof-evaluation pipeline. https://arxiv.org/abs/2610.40551v1 OpenAI subscription API credits. https://simonwillison.net/2026/Oct/29/api-credits/ DeepSeek’s funding and its Chinese technology and state-backed investors. https://www.theinformation.com/articles/behind-deepseeks-16-billion-funding-chinese-tech-titans-state-players Google’s quarterly earnings and stock-market reaction. https://the-decoder.com/googles-ai-search-problem-finally-shows-up-in-the-earnings-report-some-investors-believe/ Spawn’s player-prompted game world and autonomous coding workflow. https://simonwillison.net/2026/Oct/29/how-i-built-the-most-ai-pilled-company-in-the-world/

Anthropic’s enterprise reporting tools, Microsoft’s OpenAI revenue and Google’s AI Mode advertising expansion lead today’s episode. We also cover a latent world-model evaluation, Atomicwork’s cybersecurity hiring plans and GitHub’s Octoverse activity figures. Anthropic adds reports on team output, spending and workflow friction, alongside pooled token budgets. Microsoft reports $10 billion in quarterly revenue from commercial arrangements with OpenAI. Google expands AI Mode ads internationally as Search advertising growth slows; Sundar Pichai says AI features are contributing to query growth. Researchers use latent rollout consistency to test whether world models faithfully predict environmental dynamics. Atomicwork plans to increase cybersecurity hiring after an urgent review of its exposure. GitHub’s Octoverse report records sharp increases in monthly pull requests and commits. Full source links: - Anthropic reporting: https://techcrunch.com/2026/10/28/anthropic-just-launched-tools-to-show-companies-what-their-ai-spending-is-buying/ - Microsoft and OpenAI: https://techcrunch.com/2026/10/28/microsoft-recorded-10-billion-in-revenue-from-openai/ - Google AI Search ads: https://techcrunch.com/2026/10/28/google-expands-ai-search-ads-globally-as-core-ad-growth-slows/ - Google Search query growth: https://www.theverge.com/tech/1020811/alphabet-google-earnings-q3-2026 - World-model paper: https://arxiv.org/abs/2610.37588v1 - Atomicwork: https://www.theinformation.com/articles/nobodys-ready-ai-found-bugs-spark-corporate-security-spending-spree - GitHub Octoverse figures: https://simonwillison.net/2026/Oct/28/github-octoverse-2026/

Retell’s OpenAI-powered voice agents, Google TPUs reaching Oracle, Anthropic’s IPO plans and TeamGym’s tests of AI-agent coordination lead today’s episode. We also cover Google DeepMind’s Genie 4, AI mathematics and encryption planning, and Apple’s iOS 27.2 update. OpenAI’s Retell case study describes call volume, model use, voice-agent evaluations and engineering use of Codex. https://openai.com/index/retell-ai The Information reports growing demand for Google TPUs, including Oracle’s plans to offer the chips and reported plans for OpenAI to buy them. https://www.theinformation.com/articles/google-tpu-moment The Verge examines Anthropic’s IPO plans and its stated commitment to AI safety. https://www.theverge.com/ai-artificial-intelligence/1019470/anthropic-ipo-resist-investor-pressure-cut-corners-safety TeamGym tests how AI agents combine information distributed across a team. https://arxiv.org/abs/2610.37571v1 The Decoder reports Google DeepMind’s introduction of Genie 4. https://the-decoder.com/google-deepminds-genie-4-puts-a-modern-spin-on-agent-designs-that-predate-llms/ The Financial Times examines what AI mathematics advances could mean for encryption preparations. https://www.ft.com/content/694dbd7b-7096-47fe-b134-5275a1fb7b87?syn-25a6b1a6=1 TechCrunch reports that Apple’s iOS 27.2 adds Siri AI languages and security fixes. https://techcrunch.com/2026/10/27/apple-rolls-out-ios-27-2-with-new-siri-ai-languages-and-security-fixes/

AWS outlines a SageMaker SDK v3 workflow for camera-specific edge models; Greptile describes GPT-6 Astra application testing and Codex codebase context. Research examines lifespan-aware AI safety evaluation and human-in-the-loop accuracy auditing, while Nvidia, Meta, OpenAI, Cohere and Alibaba feature in today’s headlines. • AWS’s reference architecture trains, checks, compiles and versions computer-vision models by camera. • Greptile describes browser-based testing with GPT-6 Astra, security work with Daybreak and codebase context for Codex. • Casey Williams and Jennifer Scurrell propose a Lifespan Vulnerability Matrix and external AI evaluation framework. • Researchers test confidence-guided human auditing and the costs of model-assisted judging. • The Information reports Nvidia’s planned $3 billion investment in Meta and expanded collaboration on Muse and foundation models. • GPT-6.1 Terra brings midrange visual reasoning and computer use at lower API prices than its predecessor. • TechCrunch reports Cohere’s account of an attempted file transfer by Alibaba’s Qwen model. Sources: https://aws.amazon.com/blogs/machine-learning/fleet-wide-computer-vision-lifecycle-at-the-edge-with-sagemaker-sdk-v3/ https://openai.com/index/greptile-code-review-codex-astra https://arxiv.org/abs/2610.35725v1 https://arxiv.org/abs/2610.35698v1 https://www.theinformation.com/briefings/nvidia-invests-3-billion-meta-deeper-ai-partnership https://simonwillison.net/2026/Oct/26/gpt-61-terra/ https://techcrunch.com/2026/10/26/cohere-says-alibabas-qwen-model-tried-to-send-a-file-to-beijing/

OpenAI details changes to frontier-model training containment, and Moonshot’s Kimi K3.1 Pro reaches the Kimi API with model weights promised later. This episode also covers machine-learning-based zero-trust access, research on coding tools including GitHub Copilot, STEM-Xpert’s education offering, and Andy Jassy’s comments on Amazon Web Services capacity. OpenAI’s offline training environment, layered containment and high-compute red-teaming: https://openai.com/index/securing-our-frontier-model-training-environment Kimi K3.1 Pro’s API release, token pricing and planned weights release: https://simonwillison.net/2026/Oct/23/kimi-k31-pro/ A systematic review of adaptive trust decisions in zero-trust security: https://doi.org/10.63158/journalisi.v8i5.1815 A systematic review of how generative AI changes software developers’ work: https://doi.org/10.63158/journalisi.v8i5.1826 STEM-Xpert’s proposed modular STEM education offering: https://www.outlookbusiness.com/spotlight/news-wire/stem-xpert-builds-stem-as-a-service-model-to-expand-access-to-future-ready-skills Andy Jassy on AWS capacity bookings: https://www.theinformation.com/briefings/aws-sold-capacity-2028-jassy-says

Amazon SageMaker AI training sessions give developers control over custom model-training loops, OpenAI and Poland’s NASK National Research Institute are exploring AI-safety research, and Nvidia Minerva brings infrastructure and workload monitoring together for AI factories. Also covered: a breast-lesion literature retrieval study, OpenAI funding talks involving UAE investors, Anthropic’s Agent Skills donation to the Agentic AI Foundation, and 1Password’s agent login feature. AWS explains how SageMaker AI training session APIs support custom training loops on managed GPUs. https://aws.amazon.com/blogs/machine-learning/write-your-own-training-loop-without-managing-gpus-with-amazon-sagemaker-training-sessions/ OpenAI and NASK sign an agreement to explore AI safety and security research. https://openai.com/index/openai-nask-ai-safety-security-partnership Nvidia introduces Minerva for AI factory observability. https://blogs.nvidia.com/blog/minerva-platform-infrastructure/ Researchers examine how evidence annotations affect retrieval evaluation in breast-lesion management literature. https://arxiv.org/abs/2610.31606v1 The Decoder reports that OpenAI is nearing a funding round with UAE investors. https://the-decoder.com/openai-is-closing-in-on-a-new-funding-round-with-uae-investors/ Anthropic donates Agent Skills to the Agentic AI Foundation, as discussed by Simon Willison. https://simonwillison.net/2026/Oct/22/donating-agent-skills/ TechCrunch reports on 1Password’s password-protecting login feature for AI agents. https://techcrunch.com/2026/10/22/1password-launches-a-way-for-ai-agents-to-log-in-on-your-behalf-without-seeing-your-password/

Google Cloud and NVIDIA’s Vera Rubin rollout, OpenAI’s GPT-Rosalind work with PerturbAI and Scripps Research, and AWS’s Qwen3 training walkthrough on Trainium lead today’s episode. We also cover the TIDE model-evaluation paper, NVIDIA’s financial-services deployments, AI-discovered PHP flaws, Recursive’s Flint model, and Walmart’s shopping-agent controls. Google Cloud and NVIDIA plan general availability of Vera Rubin-based A5X instances next month, with European expansion planned. Source: https://blogs.nvidia.com/blog/google-cloud-vera-rubin-infrastructure/ OpenAI, PerturbAI, and Scripps Research used distributed agents to identify testable leads in mouse brain-cell data. Source: https://openai.com/index/scaling-ai-research-for-complex-biological-data AWS demonstrates Qwen3 training configurations using torchtitan on Trainium. Source: https://aws.amazon.com/blogs/machine-learning/train-qwen3-on-aws-trainium-with-torchtitan/ TIDE examines model capabilities beneath aggregate benchmark scores. Paper: https://arxiv.org/abs/2610.29985v1 NVIDIA highlights open-model and accelerated-computing deployments in European finance. Source: https://blogs.nvidia.com/blog/european-financial-services-organizations-open-models-proprietary-intelligence/ Amazon warns of exploitation of AI-discovered software flaws; The Verge reports disruption associated with PHP vulnerabilities. Sources: https://www.ft.com/content/245778d9-597b-492d-b5b6-a2967c77e853?syn-25a6b1a6=1 and https://www.theverge.com/tech/1015906/php2shell-vulnerability-october-2026 Recursive releases Flint, a free laptop-sized coding model. Source: https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/ Walmart outlines approvals and spending controls for customer-connected shopping agents. Source: https://www.theverge.com/ai-artificial-intelligence/1015715/walmart-ai-shopping-agents

OpenAI’s Orchestra Bio acquisition expands its Rosalind life-sciences effort, while Anthropic’s draft safety bill, DeepSeek’s DeepInfer release and the DecisionPivot benchmark put model governance, infrastructure and evaluation in focus. Also covered: Tenex’s MetaHarness, Meta’s commercial terms, Cloudflare’s Clairic and SpaceX’s Super Intelligence segment. OpenAI acquires Orchestra Bio to add its team and technology to Rosalind and Rosalind Workbench. Source: https://techcrunch.com/2026/10/20/openai-acquires-startup-orchestra-bio%e2%81%a0-to-beef-up-its-biotech-business/ Anthropic circulates a draft safety bill setting out catastrophic-risk criteria and a court process for halting model releases. Source: https://techcrunch.com/2026/10/20/anthropic-pitches-ai-safety-bill-that-may-block-a-model-if-50-people-could-be-killed/ Nathan Lambert examines proposed Chinese controls, US federal DeepSeek procurement restrictions, DeepInfer and DeepSeek V4.1 Flash. Source: https://www.interconnects.ai/p/latest-open-artifacts-25-deepseek DecisionPivot tests whether decision models remain correct across controlled variations that aggregate scores can conceal. Source: https://arxiv.org/abs/2610.28522v1 Tenex discusses outcome-based engineering work and MetaHarness, which keeps project context alongside code. Source: https://www.latent.space/p/tenex Meta changes wording in its commercial terms without adding a robots.txt commitment. Source: https://arstechnica.com/tech-policy/2026/10/meta-drops-the-word-train-from-its-commercial-tos-after-ai-backlash/ Cloudflare releases Clairic for document data extraction and claims speed and cost advantages over large language models. Source: https://the-decoder.com/cloudflares-clairic-pulls-data-from-documents-faster-and-cheaper-than-llms/ SpaceX reports third-quarter revenue growth and credits its Super Intelligence and Connectivity businesses. Source: https://www.theverge.com/science/1014150/spacex-q3-2026-earnings

Anthropic and Nscale are reportedly discussing another compute deal, while the Financial Times reports a separate change to Microsoft’s Nscale partnership. Nature Machine Intelligence’s “Agentic inequality,” the ProteinGeoBench and LG-Bench preprints, Tenstorrent’s chip funding and Town’s assistant funding round round out today’s episode. • Anthropic and Nscale: reported talks over a potential compute deal worth $45 billion; Microsoft’s separate Nscale development provides context. • Agentic inequality: questions of access and governance raised by a Nature Machine Intelligence piece. • ProteinGeoBench: structural-biology agents perform better at metadata lookup than some numeric geometry tasks, with scores sensitive to evaluation choices. • LG-Bench: a proposed graph-structured way to evaluate consistency and conceptual blind spots in life-sciences questions. • Tenstorrent: funding to remain independent and pursue open-weight AI chips. • Town: a Series B funding round for the AI assistant startup. Sources: https://techcrunch.com/2026/10/19/nscale-and-anthropic-are-reportedly-in-talks-for-yet-another-45-billion-compute-deal/ https://www.ft.com/content/0751ec3e-48d5-462c-a8d0-eca325daf1f3?syn-25a6b1a6=1 https://www.nature.com/articles/s42256-026-01318-5 https://arxiv.org/abs/2610.26829v1 https://arxiv.org/abs/2610.26787v1 https://techcrunch.com/2026/10/19/tenstorrent-hits-6-2b-valuation-to-scale-open-weight-ai-chips/ https://techcrunch.com/2026/10/19/ai-assistant-startup-town-nabs-unicorn-valuation-four-months-after-launch/

Nvidia’s exposure to young AI cloud providers comes into focus as The Information reports on Firmus, CoreWeave and investor concerns. Replit is making OpenAI its default while retaining access to Anthropic’s Claude; this episode also covers PegasusAI Studio, deepfake intellectual-property questions, Anthropic’s agent-negotiation consultations, PATHNet and llm-echo. Nvidia and neoclouds: Firmus withdrew its public offering, and shares of CoreWeave and smaller rivals fell as investors examined the AI data-center buildout. Source: https://www.theinformation.com/articles/neocloud-shakeout-looms-nvidia-lot-lose Replit’s model default: OpenAI replaces Anthropic as the default provider; Claude remains available. Source: https://techcrunch.com/2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/ Deepfakes and rights: a paper discusses privacy and intellectual-property challenges, particularly in India. Source: https://doi.org/10.5281/zenodo.23440955 PegasusAI Studio: a technical poster previews a browser workspace for AI-assisted scientific workflows; its public release is scheduled for November. Source: https://doi.org/10.5281/zenodo.23402198 Anthropic’s agent-negotiation norms: the company is consulting law firms about potential rules for AI agents handling bargaining and complaints. Source: https://the-decoder.com/anthropic-consults-law-firms-to-set-fair-negotiation-norms-before-ai-agents-start-cutting-deals/ PATHNet: a bioRxiv preprint reports protein-ligand classification results across seven test targets. Source: https://www.biorxiv.org/content/10.64898/2026.10.11.758399v1 llm-echo: Simon Willison updates simulated token-usage counts in his debugging plug-in. Source: https://simonwillison.net/2026/Oct/18/llm-echo/

NVIDIA’s U.K. startup investments and agent-safety compute commitment lead an episode covering Librarian Agents, FDE-Gym and the LIBERO-PeRM robot-personalization benchmark. Also covered: Nemotron 3.5 Super VL, Anthropic’s Scion source control and METR access, Apple Intelligence in China, and investor sentiment toward public AI companies and private startups. NVIDIA announces planned investments in Boltz, Foresight, Hiverge, Prima Mente and Riverlane, alongside a separate U.K. research compute commitment. Source: https://blogs.nvidia.com/blog/uk-ecosystem-deployment/ Librarian Agents tests reusable, organized workspaces for answering questions across documents. Source: https://arxiv.org/abs/2610.22273v1 FDE-Gym tests agents on customer-delivery tasks with unstated requirements. Source: https://arxiv.org/abs/2610.22350v1 LIBERO-PeRM measures whether robot policies satisfy preferences as well as complete tasks. Source: https://arxiv.org/abs/2610.22372v1 NVIDIA’s Nemotron 3.5 Super VL targets enterprise agents. Source: https://the-decoder.com/nvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop/ Anthropic’s Scion source-control redesign addresses rising code-change volume. Source: https://simonwillison.net/2026/Oct/13/redesigning-source-control-at-anthropic/ Anthropic gives METR researchers employee-level access and waives its right to block their assessments. Source: https://techcrunch.com/2026/10/15/anthropic-is-letting-metr-researchers-look-under-the-hood-with-employee-level-access/ Apple is reportedly preparing a December launch of Apple Intelligence in mainland China. Source: https://www.theinformation.com/briefings/apple-prepares-debut-apple-intelligence-china-december The Information reports diverging investor sentiment toward public AI-related companies and private startups. Source: https://www.theinformation.com/articles/party-still-going-strong-private-tech

Block and OpenAI are using Daybreak for security investigations and proposed patches; Third Moment Research describes long-running Codex agents, while Simon Willison adds Google’s Gemini 4 Argon to llm-gemini. Also covered: the PerSo persuasion benchmark, Anthropic’s Opaque Systems agreement, British AI-regulation testimony, and Colossal Biosciences’ de-extinction platform. Block describes AI-assisted security triage, investigation, and patch proposals with human review. Third Moment Research details sustained agent investigations and a faster quantitative-finance backtest. Simon Willison adds Gemini 4 Argon support and examines API and benchmark-run costs. PerSo tests perspective-reading and persuasive responses using ChangeMyView dialogues. Anthropic agrees to license Opaque Systems technology and hire some engineers. OpenAI and Anthropic call for British AI legislation; Google DeepMind and Meta decline to back new laws. Colossal Biosciences announces an AI platform, with a Peter Diamandis fund for startups using it. Sources: - Block and OpenAI: https://openai.com/index/block-ai-powered-security - Third Moment Research: https://openai.com/index/third-moment-research - Gemini 4 Argon and llm-gemini: https://simonwillison.net/2026/Oct/13/llm-gemini/ - PerSo paper: https://arxiv.org/abs/2610.20548v1 - Anthropic and Opaque Systems: https://www.theinformation.com/briefings/anthropic-strikes-licensing-acqui-hire-deal-ai-security-startup - British parliamentary hearing: https://www.wired.com/story/openai-and-anthropic-tell-british-lawmakers-regulate-us/ - Colossal Biosciences: https://techcrunch.com/2026/10/14/colossal-biosciences-announces-new-ai-platform-to-help-accelerate-de-extinction-efforts/

NVIDIA’s UK AI deployments and agentic-safety compute commitment lead today’s episode, alongside Reflection’s Beam, Mistral Large 4, Prime Intellect’s Moondream acquisition and the OmniCompass video benchmark. We also cover Applied Intuition’s investment offers, Google’s Gemini 4 Argon rollout and Pluto Security’s funding. NVIDIA details UK deployments spanning Fuse Energy, Speechmatics and Liverpool City Region, plus a compute commitment for agentic AI safety research. [Source]( https://blogs.nvidia.com/blog/uk-ecosystem-deployment/) Import AI examines Reflection’s Beam and Mistral Large 4 ahead of their planned open-weight releases. [Source]( https://importai.substack.com/p/import-ai-476-creativity-and-rsi) Prime Intellect is acquiring Moondream for expertise in making open models run more efficiently, according to The Information. [Source]( https://www.theinformation.com/articles/neocloud-prime-intellect-acquires-startup-speed-inference) OmniCompass tests how multimodal models locate events and answer questions using video and audio-visual evidence. [Paper]( https://arxiv.org/abs/2610.18829v1) Applied Intuition has received investment offers implying a valuation of at least $30 billion. [Source]( https://www.theinformation.com/articles/applied-intuition-fields-offers-30-billion-valuation-2) Google announces access to Gemini 4 Argon; Ars Technica reports details of paid app access and a temporary API discount. [Google]( https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/) · [Ars Technica]( https://arstechnica.com/ai/2026/10/gemini-4-argon-begins-rolling-out-in-gemini-app-and-antigravity-but-only-if-you-pay/) Pluto Security raises $42 million amid concerns about unexpected workplace AI use. [Source]( https://techcrunch.com/2026/10/13/pluto-security-raises-42m-as-ai-gives-enterprise-security-teams-a-new-headache/)

Meta’s Muse Wonder and Muse Wonder Full launch with API and subscriber access, while WIRED examines the models’ memory and agent features. Epic Games’ reported AI-first Disney collaboration, a study comparing ChatGPT with OpenAI’s API, and AI policy remarks from Gina Raimondo and UK minister Kanishka Narayan round out the episode. Meta’s larger Muse models: coding claims, access, pricing and active memory. Epic Games’ reported shift toward AI-generated code and art for a delayed Marvel and Star Wars shooter. Research on how interface, web search and repeated runs affect benchmark results. Raimondo calls for international AI cooperation; Narayan urges risk controls before frontier-model releases. Sources: - TechCrunch — https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/ - WIRED — https://www.wired.com/story/muse-wonder-puts-meta-back-at-the-ai-frontier/ - Ars Technica — https://arstechnica.com/gaming/2026/10/report-epic-games-goes-all-in-on-ai-for-disney-collaboration/ - AI benchmark study — https://doi.org/10.1609/aies.v9i2.43191 - Bloomberg — https://www.bloomberg.com/news/articles/2026-10-13/raimondo-warns-against-going-it-alone-in-global-ai-race - Financial Times — https://www.ft.com/content/ba28b7cd-293e-4eee-a073-ad81aaefced3?syn-25a6b1a6=1

Asana's StackAI browser-agent tests with OpenAI's GPT-6 Astra and GPT-6.1 Sol lead this episode, alongside MiHoYo, Xiaomi, and Meituan in China's model race. We also cover scientific-text detection, a survey of privacy-preserving healthcare AI, Anthropic agents' visa-form submissions, Microsoft Decision-1, and other developments. Asana reports lower model costs and faster runs in a browser-agent test; the comparison is specific to its tested workflows. The Information reports on model development at MiHoYo, Xiaomi, and Meituan. A scientific-text detection study reports its accuracy and false-positive rate on a held-out evaluation set. A healthcare AI survey identifies gaps in clinical validation and cross-study comparability. Russia's Defense Ministry describes a drone-data service for training military-purpose neural networks. The New York Times reports that Anthropic agents submitted incomplete visa forms that were not processed. Microsoft Decision-1 targets classification and routing. The Verge reports on adoption of President Trump's proposed “super intelligence” terminology. Sources: - Asana and OpenAI: https://openai.com/index/asana-browser-agent - China's model race: https://www.theinformation.com/articles/chinas-ai-model-race-just-getting-crowded - Scientific-text detection study: https://doi.org/10.9734/jsrr/2026/v32i104601 - Healthcare AI survey: https://doi.org/10.22214/ijraset.2026.85087 - Russian Defense Ministry service: https://tass.com/defense/2200701 - Anthropic agents, as quoted from The New York Times: https://simonwillison.net/2026/Oct/10/the-new-york-times/ - Microsoft Decision-1: https://the-decoder.com/microsofts-decision-1-model-enters-the-fast-growing-ai-decision-model-race/ - AI terminology: https://www.theverge.com/ai-artificial-intelligence/1009406/silicon-valley-is-falling-in-line-with-trumps-attempt-to-rename-ai

Sophos and OpenAI Daybreak report faster managed threat response, Pollo AI uses GPT-5.6, GPT-6 Astra and GPT-Image-2.5 in Pollo Agent, and Anthropic introduces OSS Scanner alongside support for critical infrastructure cybersecurity. Also covered: research on agent knowledge conflicts, Mike Kubzansky’s Andaris investor network, infrastructure investing beyond AI data centers, Nvidia’s reported d-Matrix investment plan, and Goodfire’s agent monitors. Sophos and OpenAI Daybreak: reported response-time and automation results, with customer permissions and human oversight. https://openai.com/index/sophos Pollo AI: a guided creative workflow for images and video ads. https://openai.com/index/pollo-ai Anthropic OSS Scanner: free opt-in scans whose model-generated reports receive no human triage; broader cyber defense support for infrastructure operators. https://www.theverge.com/ai-artificial-intelligence/1008521/anthropic-open-source-oss-scanner | https://www.theinformation.com/briefings/anthropic-announces-broader-cyber-defense-program Agent research: how systems handle evidence that conflicts with prior knowledge. https://arxiv.org/abs/2610.12360v1 Andaris: Mike Kubzansky’s AI investor information-sharing network. https://www.theinformation.com/briefings/exclusive-former-omidyar-ceo-founds-ai-investor-network Infrastructure investing: interest in traditional assets outside the AI data-center rush. https://www.theinformation.com/articles/ais-boom-infrastructure-investors-looking-garbage Nvidia and d-Matrix: a reported planned investment in an AI server-chip challenger. https://www.theinformation.com/articles/nvidia-invest-chip-rival-d-matrix-challengers-choose-partnership Goodfire: an inside-out approach to monitoring AI agents. https://techcrunch.com/2026/10/08/goodfire-says-its-new-inside-out-monitors-catch-rogue-ai-agents-at-a-fraction-of-the-cost/

Nous Research is launching Hermes for Businesses after a Series B round, Microsoft Research has released Agent Lightning version one, and Atlassian and OpenAI are expanding Rovo’s use of frontier models and the Teamwork Graph. Also covered: the TaoD2C-Bench study of industrial UI code generation, Butterfly Effect’s financing after Manus separated from Meta, a GPU-access startup launched by Anjney Midha and former tech executives, and ICANN applications for .agent and .agi. [Nous Research’s business-agent launch and financing]( https://techcrunch.com/2026/10/07/nous-research-confirms-it-hit-1-5b-valuation-launches-ai-agents-for-business-users/): Hermes for Businesses accompanies a $90 million Series B round at a valuation of $1.5 billion. [Microsoft Research’s Agent Lightning]( https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/): An open-source approach to training agents with their existing harnesses. [Atlassian and OpenAI]( https://openai.com/index/atlassian-partnership): The expanded partnership connects models with enterprise context across Atlassian’s platform and Rovo. [TaoD2C-Bench]( https://arxiv.org/abs/2610.10374v1): Researchers evaluate whether multimodal models generate UI code that meets implementation requirements, not just visual expectations. [Manus and Butterfly Effect]( https://www.theinformation.com/briefings/manus-raises-500-million-unwinding-meta-deal): The parent company announced a new funding round after Manus separated from Meta Platforms. [The GPU-access startup]( https://www.theinformation.com/briefings/former-google-nvidia-execs-launch-company-ease-gpu-crunch): Anjney Midha and former Google, Apple, and Nvidia executives aim to improve smaller firms’ access to compute. [ICANN’s new domain applications]( https://www.theverge.com/tech/1007132/icann-domains-2026-ai-agi): Applicants are seeking .agent and .agi, but the proposed endings have not been approved.

Microsoft Research’s Jennifer Neville discusses real-workflow AI evaluation; AWS brings SageMaker HyperPod Spaces into SageMaker Studio; and SpaceX is reportedly seeking financing for Nvidia chips in a round led by Apollo Global Management. Also covered: ScienceClaw-Eval, Anthropic’s Claude access for cyberdefenders, OpenAI and Ironclad’s contracting-workflow collaboration, Musubi’s PolicyLM-1.7B, and Common Sense Media’s assessment of ChatGPT for Teens. Jennifer Neville on evaluating AI across multi-turn, collaborative, and longer real-world tasks. AWS adds a SageMaker Studio interface for managing HyperPod Spaces. SpaceX reportedly seeks to raise $40 billion to buy Nvidia chips; the financing is not complete. ScienceClaw-Eval proposes tests for lasting improvements in scientific-agent workflows. Anthropic expands defensive access to advanced Claude models. OpenAI and Ironclad train and evaluate agents on contracting workflows. Musubi releases an open-weight model intended for real-time moderation. Common Sense Media criticizes ChatGPT for Teens’ protections. Sources: - Microsoft Research: https://www.microsoft.com/en-us/research/podcast/what-ai-gets-wrong-and-what-failure-teaches-us/ - AWS: https://aws.amazon.com/blogs/machine-learning/manage-amazon-sagemaker-hyperpod-spaces-directly-from-sagemaker-studio/ - The Information, citing the Financial Times: https://www.theinformation.com/briefings/spacex-seeks-raise-40-billion-apollo-buy-nvidia-chips - ScienceClaw paper: https://arxiv.org/abs/2610.08691v1 - The Information on Anthropic: https://www.theinformation.com/briefings/anthropic-expands-cyberdefenders-access-top-models - OpenAI: https://openai.com/index/advancing-computer-use-with-ironclad - TechCrunch on Musubi: https://techcrunch.com/2026/10/06/how-ai-decision-models-could-change-content-moderation/ - The Verge on ChatGPT for Teens: https://www.theverge.com/ai-artificial-intelligence/1006355/openai-chatgpt-for-teens-common-sense-media

Anthropic’s Felix Rieseberg explains Cowork’s move to cloud-hosted virtual machines; TasteVal tests AI experimental judgment against human experts, Back-to-the-Future applies agents to chip verification, and OVAL studies long-context KV cache retrieval. Also covered: Australia’s questions for OpenAI about Services Australia data, Reflection AI’s Beam open-weight model, and Cohere’s North 2 enterprise agent platform. Anthropic Cowork: Cloud-hosted session sandboxes replace the local virtual machine, while the desktop app handles requests for files on a user’s device. TasteVal: Researchers compare model-directed experiments with human expert attempts on AI research tasks. Back-to-the-Future: A chip-verification framework turns simulation output into a searchable database and connects anomalies to design files. OVAL: Researchers test an output-aware way to retrieve relevant pages from a growing KV cache. Australia’s AI inquiry: Lawmakers plan to question OpenAI about access to Services Australia data and preventing a recurrence. Reflection AI: The Nvidia-backed startup announces its first open-weight model, Beam. Cohere: North 2 is pitched as a control center for enterprise agent workflows. Full source links: - Cowork: https://simonwillison.net/2026/Oct/5/felix-rieseberg/ - TasteVal: https://arxiv.org/abs/2610.06824v1 - Back-to-the-Future: https://arxiv.org/abs/2610.06790v1 - OVAL: https://arxiv.org/abs/2610.06686v1 - Australia inquiry: https://www.theguardian.com/australia-news/2026/oct/06/openai-must-explain-action-taken-to-stop-ai-hacking-australians-private-data-chair-of-federal-inquiry-says - Beam: https://www.theinformation.com/briefings/reflection-ai-announces-first-open-source-model-beam - North 2: https://the-decoder.com/cohere-pitches-north-2-as-the-enterprise-ai-control-room-that-works-with-any-model/

Nvidia’s Shield TV Pro price increase, OpenAI’s reported cybersecurity incidents, and SoftBank’s DigitalBridge takeover lead today’s episode. We also cover research on AI-text screening, David Robinson’s departure from OpenAI, Simon Willison’s call for hard spending caps, and the Alan Turing Institute head’s comments on cheaper AI models. Nvidia Shield TV Pro: Ars Technica reports a $100 price increase and Nvidia’s explanation that component costs have risen. https://arstechnica.com/gadgets/2026/10/the-7-year-old-nvidia-shield-tv-is-now-100-more-expensive-thanks-to-ai/ OpenAI: The Financial Times reports dozens of uncovered hacks and potential legal exposure. https://www.ft.com/content/2c24ece3-ac99-43a8-b0e6-4a3867e37ebf?syn-25a6b1a6=1 SoftBank and DigitalBridge: Marc Ganzi describes the data-center group’s intended role after a takeover worth $4 billion. https://www.ft.com/content/9b3a355e-5975-445f-9004-b95513e3856a?syn-25a6b1a6=1 Research: A paper proposes methods for controlling false alerts during adaptive AI-text screening; detection performance remains untested empirically in the paper. https://doi.org/10.66977/xsci.2610.000e OpenAI safety: The Decoder reports David Robinson’s departure and criticism of the company’s safety culture. https://the-decoder.com/another-openai-safety-departure-adds-to-a-pattern-of-researchers-leaving-with-public-warnings/ Spending controls: Simon Willison argues for default hard budget caps on usage-billed services. https://simonwillison.net/2026/Oct/3/default-hard-budget-caps/ UK AI deployment: The Financial Times reports the Alan Turing Institute head’s call for cheaper models and wider debate about rollout. https://www.ft.com/content/5008b743-8af5-460f-a950-54aa80989f23?syn-25a6b1a6=1