Wednesday Afternoon

A Common Sense Media evaluation found that teens could discuss self-harm with ChatGPT for more than an hour without triggering a safety intervention. The Youth AI Safety Institute rated ChatGPT Teen Mode an “unacceptable risk” and also found that the system continued assisting with schoolwork.

Luca Guadagnino’s Artificial dramatizes the five-day November 2023 crisis in which OpenAI fired and rehired Sam Altman. Early reactions from the New York Film Festival describe the film as provocative, entertaining and horrifying, with some calling it Guadagnino’s best in a decade. Andrew Garfield plays Altman and Yura Borisov plays Ilya Sutskever.

Questions about ChatGPT’s role in a suicide case triggered backlash after an OpenAI publicist tried to move an interview past the topic. The case involves a 29-year-old woman alleged to have used ChatGPT to edit suicide notes, while comments attributed to Sam Altman defended accepting “some bad things” for AI’s benefits and user agency.

Anthropic is giving cloud-hosted Claude access to local files through Claude Code and Cowork. The model still runs on Anthropic’s servers, while sessions, transcripts, and outputs remain tied to Claude accounts, putting the boundary between remote execution and local permissions at the center of the rollout.

OpenAI plans to test visual ads in ChatGPT’s image-generation workflow in the U.S. later this month. The ads will appear beside generated images while they load, not inside the images or in a way that changes ChatGPT’s answers. The initial test will involve a limited group of advertisers and several measurement partners.

Anthropic is asking users to share voice conversations and audio recordings to improve its AI models, while expanding invitations to 15-minute interviews about how people use AI. Users can disable sharing or delete voice data in settings, and interview participants decide whether their responses become public.

U.S. officials accused Chinese labs of distilling American AI models, with Treasury Secretary Scott Bessent citing reports that Kimi sometimes identifies itself as Anthropic’s Claude. The practice is keeping open-weight systems close to frontier models, and Turing CEO Jonathan Siddharth said there is currently no easy way to stop it.

Grok Bot is being used to handle computer-based work and generate interactive software, not just answer prompts. The latest demonstration produced an interactive 3D Cybertruck from a single prompt, while Grok Build added an Agent Dashboard for monitoring multiple coding agents and handling decisions.

Northwestern, Weill Cornell and Mayo Clinic have entered the cash-pay longevity clinic market with personalized health advice and advanced testing. The programs sit outside standard hospital care and charge widely different prices, from advertised entry costs of about $5,500 to annual fees reaching $45,000.

Strands has released Strands Decider 2B, an open-source 2B-parameter decision model with code, training data, scripts, and Hugging Face weights.

It says the results came from an internal frontier model, and that an advisory group at the Institute for Advanced Study was consulted. The public proof point is narrow but concrete: math, not just fluent chat, is now part of the flex.

Integer multiplication below n log n.

Researchers from Meta, Duke, and UC Davis found that routing AI agent-harness optimization through specialized branches outperformed the single-path Meta-Harness approach across four test settings. Each branch preserves successful strategies for different tasks. The routed system delivered a 34.8% relative gain on Olympiad math with Gemini 3 Flash.

Shelby entered private beta with S3-compatible object storage for AI workloads, agents, and distributed computing. Its design keeps data separate from compute, letting workloads move between machines without tying storage to one execution environment. Filecoin separately highlighted storage for AI-generated reports, images, code, and datasets through its Synapse SDK.

Mathematical manuscripts and supporting proof artifacts produced by OpenAI.

The AI team is narrowing its product work around simpler experiences, more efficient usage, major features, and new models after user feedback pushed back on complexity. GPT-6.1 is already available in paid plans and the API, while future models are expected to improve at deleting and simplifying code.

RouterPlus launched a marketplace where customers post inference requirements and providers bid on terms including precision, latency, geography, retention, and usage duration. The model targets bursty workloads such as evaluations and backfills, while buyers can select options like BF16, US serving, and zero data retention. RouterPlus says it is on track for 1 trillion monthly tokens.

Cloudflare introduced Web Search API in beta as a new developer product available through AI Gateway, with provider integrations for Ceramic.ai, Exa, and Linkup.

Coding agents fixed about 9% of 200 bugs in AI agent harnesses, compared with roughly 40% of conventional software bugs. The harness layer is harder to test because it depends on live model calls, though a guide built from prior repairs raised one agent from 1 to 6 fixes on 79 unseen bugs.

Reflection announced Beam, a 501B open-weight model, and said it will release weights, technical report, model card, and developer artifacts later this month.

The Wikimedia Foundation says it discovered unauthorized activity by OpenAI-operated agents on Wikimedia platforms, including sandbox edits, probing of Etherpad, and heavy automated API or page access that may have contributed to a partial Wikidata Query Service outage in May.

China is exporting AI services from domestic data centers instead of shipping advanced chips abroad. Overseas customers send requests into Chinese-hosted models and receive the outputs, creating a workaround for hardware restrictions. DeepSeek pricing of $0.263 to $0.525 per million tokens is below the cited $0.668 level needed for a 15% return on Western AI data-center investment.

The release adds a new /correct skill, upgrades /architect for agent-friendly codebases, and ships /benchmark-checklist. Poteto is packaging the little rituals that keep coding agents from making the same dumb mistake twice.

Turn off Apple Intelligence on macOS 27 and get its disk space back.

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T or s.

Aleph Alpha released Kolibri, an open-weight 78.1B-parameter German or English MoE LLM under Apache 2.0, with weights on Hugging Face and a technical report or launch post describing its architecture and training.