Stay ahead on AI.
All in ONE place.

OpenAI has launched the Grid Protectors Program, a national initiative to help its future AI compute campuses reduce electricity demand when the US grid is under strain. Working with utilities, regulators, and policymakers, OpenAI will design facilities to cut power use during peak periods. A pilot agreement with Georgia Power at the planned Effingham County campus includes up to 1,000 megawatts of flexible demand response, enough to power roughly 700,000 households, with Georgia Power projecting approximately $950 million in annual customer savings beginning in 2029.

Israeli startup 42AI has exited stealth with $62.5 million in seed funding to build a behavioral world model that predicts human decisions. The transformer-based system converts observations of user behavior into mathematical trait representations, then calculates decision probabilities under various conditions. Currently deployed at 10 large financial brokerages, it helps manage liquidity, margin requirements, and hedging exposure. The company claims 98% accuracy in predicting trading flows and an F1 score of 0.8–0.85, and says one brokerage generated $30 million in additional value through dynamic hedging.

Elon Musk confirmed that SpaceX cut Grok Bot usage limits after costs became unsustainably high relative to revenue. Over Halloween weekend, some users found their weekly limits maxed out despite not using the agentic AI overnight, with usage jumping from as low as 50% to 100%. Musk acknowledged the issue in an X reply but has not addressed it on his main feed. SpaceX's Super Intelligence division turned a $431 million profit, though $1.2 billion in one-time costs from its Cursor acquisition reduced gains.

Hume AI has launched its Audio Decision API, an audio-native model that answers custom questions about speech and sound in real time. Unlike transcript-based systems, it processes raw audio directly, returning structured probability scores for yes/no, multiple choice, scale, or span questions — no retraining required. Use cases include detecting caller frustration, identifying AI-generated voices, and searching 40,000 recordings in minutes. The model achieved 70.7% agreement with human panel judgments, outperforming existing methods on accuracy, speed, and cost.

Starting in December, OpenAI will merge usage limits for ChatGPT and Codex into a shared token allowance across Plus, Pro, and Business plans. This change comes as OpenAI integrates Codex's agentic capabilities into the main ChatGPT experience. Usage is measured in tokens, with more advanced models and complex tasks consuming more of the allowance. Instant chat will not draw from the shared pool. Users can monitor combined usage in Settings → Usage, and may wait for resets, upgrade, or buy credits if limits are reached.

xAI has added invisible watermarks to text outputs from supported Grok models to comply with the EU AI Act. Starting November 7, 2026, the feature is enabled by default across all API endpoints for models including grok-4.7, grok-4.6, and grok-4.5. The watermark works by subtly shaping token selection during generation, adding no visible characters. Developers can check watermark status via the x-watermark response header. Team admins can opt out per model in the xAI Console, with all changes logged in the audit log.

OpenAI has published court briefs from dozens of supporters arguing that AI training on copyrighted works constitutes fair use in its ongoing lawsuit with The New York Times. Filers include the U.S. government, the American Library Association, the American Council of the Blind, legal scholars, and independent creators. Supporters argue AI training is transformative, causes no cognizable market harm, and that restricting it would hinder medical research, accessibility tools, and scientific progress while threatening American innovation and economic competitiveness.

Google's Project Suncatcher satellite reached a major milestone this week, generating the first-ever AI response from a data center class TPU in orbit. Launched October 1, 2026, the prototype satellite received an uplinked prompt during a ground-station pass near Svalbard, Norway, and ran Google's Gemma 4 open model to produce a haiku about space. Next steps include thermal stress and radiation testing of the TPUs, with twin interconnected satellites planned for 2027.

Google has announced it will voluntarily adopt the IETF AIPREF standard for AI training controls, joining Bing, OpenAI, and Anthropic. The new mechanism lets website owners set AI training preferences across multiple providers in a single robots.txt step using a shared policy token with user-agent __usage_ai-training, rather than listing each provider individually. Google plans to implement support in early 2027. Existing controls like Google-Extended will continue functioning alongside the new signal.