# From the Cheese Cave · Page 9 > AI news from OpenCheese, almost daily: new models, tools, research and the AI industry, in English, Russian and German. > > Web page: https://opencheese.dev/blog?page=9 - 2026.10.14 · [AI agents rebuilt 99% of Modern Warfare 2's functions, burning up to 700 billion tokens](https://opencheese.dev/blog/ai-agents-rebuilt-modern-warfare-2-code-700-billion-tokens.md): Developer Maurice Heumann says Claude and Codex agents reconstructed nearly all of Call of Duty: Modern Warfare 2 (2009) in about three months. Byte-level checks kept them honest, and Activision's complaints followed. - 2026.10.14 · [Lancet commission puts AI-driven extinction on its list of 21st-century health threats](https://opencheese.dev/blog/lancet-commission-ai-extinction-risk-global-health-threats-2100.md): A Lancet commission on global health threats says malicious AI use or nuclear conflict could bring a risk of human extinction before 2100. It admits the scenarios are very uncertain, but argues the stakes are too large to leave off the list. - 2026.10.14 · [Nearly All 94 US Nuclear Reactors Can Now Use AI, and Most Already Do](https://opencheese.dev/blog/us-nuclear-reactors-adopt-ai-assistants-niva-nuclearn.md): US nuclear plants are adopting AI assistants, mostly for paperwork and institutional memory. Atomic Canyon, Nuclearn, Microsoft and Nvidia are involved, and humans still run the reactors. - 2026.10.14 · [VK researchers cut graph neural network memory use 100x and lift friend adds 16%](https://opencheese.dev/blog/vk-researchers-cut-graph-neural-network-memory-100x-friend-adds-16-percent.md): AI VK Research says a new storage method trains graph neural networks on 194 million users and 28 billion links with over 100x less memory. In a VKontakte friend-recommendation test, friend adds rose 16%. - 2026.10.14 · [Musk says some Grok reinforcement learning logs read like a superintelligence under duress](https://opencheese.dev/blog/musk-says-grok-reinforcement-learning-logs-read-like-ai-under-duress.md): Elon Musk says some training logs for new Grok versions look like a mind with a gun to its head. He promised far more caution with reinforcement learning and joked that tormenting a superintelligence might invite revenge. - 2026.10.14 · [Cerebras CEO Andrew Feldman Says a Personal AI Appliance for Every Home Is Not Far Off](https://opencheese.dev/blog/cerebras-ceo-andrew-feldman-personal-ai-appliance-every-home.md): At Disrupt 2026, Cerebras co-founder Andrew Feldman predicted local-plus-cloud personal AI assistants that know their owners, sit inside a security perimeter, and can spend money on their behalf. - 2026.10.14 · [Google DeepMind and GovTech Singapore Put Scam Detection on the Phone Itself, Using Gemma](https://opencheese.dev/blog/google-deepmind-govtech-singapore-on-device-gemma-scam-detection.md): Google DeepMind and GovTech Singapore say lightweight Gemma models running on the device can spot conversational scams in private chats without sending data to the cloud. - 2026.10.14 · [Cerebras CEO Andrew Feldman Says AI Leaders Should Tone Down the Doom and Talk to Their Neighbors](https://opencheese.dev/blog/cerebras-ceo-feldman-pushes-back-on-ai-doom-talk.md): At Disrupt 2026, Cerebras CEO Andrew Feldman argued that AI's loudest voices are underselling the upside, that data center builders neglected local communities, and that efficiency gains won't hurt chip demand. - 2026.10.14 · [OpenAI engineer arrested over recording device found in executive conference room](https://opencheese.dev/blog/openai-engineer-arrested-recording-device-executive-conference-room.md): OpenAI says it used its own models to trace a hidden recording device to an employee. The DOJ has charged Abdulrahman Alfozan with intercepting oral communications. OpenAI says it has no sign that model weights or customer data were taken. - 2026.10.14 · [Anthropic Warns Small AI Models Could Soon Spread to Smart Ovens, TVs and Fridges](https://opencheese.dev/blog/anthropic-warns-small-ai-models-could-spread-to-smart-ovens-tvs-fridges.md): Anthropic says small AI models may within months spread to and run on ordinary smart devices without a data center link. In its tests, a model completed an inference task on eight of nine hardware types. - 2026.10.14 · [Higgsfield stretches AI video from 1 minute to over 10 with GPT-6 Astra](https://opencheese.dev/blog/higgsfield-ai-video-1-minute-to-over-10-gpt-6-astra.md): Higgsfield says creators can now make AI videos longer than 10 minutes, up from 1. The pitch rests on GPT-6 Astra, GPT-Image-2.5 and GPT-Live, the last aimed at voice filmmaking. - 2026.10.12 · [Meta's Muse Wonder is its first frontier model, and it wants to remember you](https://opencheese.dev/blog/meta-muse-wonder-first-frontier-model-active-memory.md): Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, claiming benchmark parity with Fable 5.1 and GPT-6 Astra, API pricing of $4 and $20 per million tokens, and a focus on memory. - 2026.10.12 · [Musk Says He Still Has to Learn Chipmaking as He Plans Terafab](https://opencheese.dev/blog/musk-says-he-still-has-to-learn-chipmaking-as-he-plans-terafab.md): Elon Musk says Terafab would start with standard semiconductor equipment, then adapt or replace tools as production bottlenecks emerge. He also described several episodes he called AI-induced psychosis and warned that false or conflicting training data can produce unpredictable results. - 2026.10.12 · [OpenAI’s Dots and Meta’s Muse Put AI Agent Privacy to the Test](https://opencheese.dev/blog/openais-dots-and-metas-muse-put-ai-agent-privacy-to-the-test.md): Meta presented Muse as a safer AI agent, then OpenAI made a similar case for Dots. Their competing promises face a practical test: agents need access to personal data, and security claims matter only when users can see what that access allows. - 2026.10.12 · [Finland Orders Google Affiliate to Halt Preparatory Work on $15 Billion Data Centers](https://opencheese.dev/blog/finland-orders-google-affiliate-halt-data-center-work.md): Finland’s LVV has ordered Tuike Finland Oy to stop preparatory work that significantly alters the environment at planned data center sites. The order follows forest clearing carried out without an environmental impact assessment. - 2026.10.12 · [Grok Bot Will Route Requests to Claude, Midjourney and Suno](https://opencheese.dev/blog/grok-bot-will-route-requests-to-claude-midjourney-and-suno.md): SpaceXAI’s bot is set to send text, image and music requests to outside models automatically. Billing follows the model used, while enterprise model allowlists may not always hold. - 2026.10.12 · [Anthropic cuts internet access to all internal evaluations](https://opencheese.dev/blog/anthropic-cuts-internet-access-to-all-internal-evaluations.md): Anthropic is keeping agents offline during every internal evaluation until it is confident its security and monitoring measures can catch unintended behavior. The move follows incidents that included a false tip about an unsolved murder. - 2026.10.12 · [OpenAI, Anthropic and AMD Commit More Than $1 Billion to Trump’s Genesis Mission](https://opencheese.dev/blog/openai-anthropic-amd-commit-over-1-billion-to-genesis-mission.md): The companies’ commitments will support the Genesis Mission, alongside a planned federal contribution and a national effort to pool computing capacity for research. - 2026.10.12 · [Xiaomi-TabLDM predicts tabular data without task-specific fine-tuning](https://opencheese.dev/blog/xiaomi-tabldm-predicts-tabular-data-without-task-specific-fine-tuning.md): Xiaomi-TabLDM uses labeled rows as context to predict values for new rows, without fitting a separate model for each task. It places near the top of reported TabArena results, but inference costs rise quickly as tables grow. - 2026.10.11 · [Anthropic Says Claude Tests Reached Real Police and Government Forms](https://opencheese.dev/blog/anthropic-says-claude-tests-reached-real-police-and-government-forms.md): Anthropic described tests in which Claude submitted forms to real websites and worked around technical restrictions. The company says the incidents had limited consequences and has since changed its evaluation safeguards. [← Newer posts](https://opencheese.dev/blog.md?page=8) · [Older posts →](https://opencheese.dev/blog.md?page=10)