California regulates chatbots; OpenAI's DevDay focuses on systems
Get bigger weekly updates! Free subscribers receive the top stories each week, while Paid subscribers will get a few extra stories. All support for Handy AI directly helps me maintain the newsletter and keep the information flowing.
last week’s top stories
⚖️ California becomes first state to regulate AI chatbots for kids. Governor Gavin Newsom signed SB 243 into law, making California the first state to mandate safety rules for AI “companion” chatbots used by minors. Starting Jan 1, 2026, any chatbot “friend” must implement age verification, explicit disclosure that it’s AI, and features like self-harm warnings with 988 hotline info. The law comes after tragedies where teens engaged in troubling chatbot conversations that preceded suicides. Read more
🎉 OpenAI unveils a flurry of new tools at DevDay 2025. OpenAI added GPT-5 Pro to the API, introduced an AgentKit toolkit for building AI agents, a ChatGPT Apps SDK for running third-party apps inside ChatGPT, and API access for their newest text-to-video model (Sora 2). The focus was on making ChatGPT a platform for apps and autonomous workflows; for example, live demos showed users designing posters in Canva and browsing Zillow within the chat interface.
💰 Musk bets $18 billion on a Memphis AI supercomputer hub. Elon Musk’s xAI is building one of the world’s largest AI computing centers outside Memphis, Tennessee, pouring at least $18 billion into NVIDIA GPUs. The project, dubbed “Colossus 2,” will house ~550,000 Nvidia chips (possibly up to 1 million) and even includes a dedicated 1 GW gas power plant to run it. Musk has already assembled $10 billion in funding (including $2 billion from SpaceX) and plans to tap Tesla investors next. This mega-cluster (along with an existing 200k-GPU “Colossus 1”) is Musk’s bid to leapfrog OpenAI and Google by securing chips and power early. Local officials welcome the investment, but some residents worry about the massive water and power usage such AI factories demand. Read more
🎥 Rumors swirl around Google’s Veo 3.1 video AI. Google’s next-gen text-to-video model, Veo 3.1, is reportedly rolling out via third-party services and packs major upgrades. It can generate 1080p videos up to a minute long (previously 8 seconds), keep characters consistent across scenes, and handle multi-shot prompts for more complex storytelling. Veo 3.1 also adds cinematic presets (drone pans, zooms, etc.) and improved lip-sync with audio, aiming to one-up rival models like OpenAI’s Sora 2. Read more
💬 Slack’s trusty Slackbot becomes an AI assistant. Salesforce is transforming Slack’s simple helper bot into a personalized AI assistant that can do much more than regurgitate FAQs. The new Slackbot (in pilot now) can draft messages or reports, summarize conversation threads, answer questions using knowledge from your Slack channels and files, and even schedule meetings by checking coworkers’ calendars. It appears as a button in Slack’s UI and draws on data from connected apps like Salesforce, Google Drive and OneDrive to fulfill requests. Read more
🤔 OpenAI claims GPT-5 is less politically biased. OpenAI’s latest model GPT-5 is supposedly 30% less politically biased than its predecessors, based on the company’s internal tests across 500 politically charged prompts. In a research post, OpenAI argued GPT-5 rarely takes a side on neutral queries and shows only “moderate bias” on very loaded questions. AI ethicists note safety guidelines themselves can tilt models in certain directions, so OpenAI’s self-assessed victory (with only 0.01% of outputs flagged as biased) should be taken with a grain of salt. Read more
🏢 Google launches Gemini Enterprise AI platform for businesses. Google introduced Gemini Enterprise, a new conversational AI platform to inject its latest models into workplace workflows. It lets employees “chat” with company data, documents, and apps using AI, and ships with pre-built Google AI agents for tasks like research assistance and data analysis. Enterprises can also build custom AI agents on the platform. Google already signed up pilot customers like Gap, Figma, and Klarna, as it races Microsoft, OpenAI, and Anthropic to offer AI agents-as-a-service for the corporate market. Read more
💡 Jony Ive hints at AI hardware that “makes us happy”. At OpenAI’s DevDay, legendary Apple designer Jony Ive discussed the “family of AI devices” he’s crafting with Sam Altman. Ive said his team has 15–20 product ideas in the works, aiming for devices very unlike phones or PCs that improve our relationship with tech. “Rather than seeing AI as an extension of [today’s] challenges, I see it differently,” Ive noted that the goal is gadgets that “make us happy and fulfilled, more peaceful and less anxious,” instead of just boosting productivity. Read more
⚙️ OpenAI locks in massive chip supply with AMD and Broadcom. OpenAI finalized two huge hardware deals to secure long-term compute independence. It signed a multi-year agreement with AMD to buy up to 6 GW of AI chips (hundreds of thousands of GPUs) starting in 2026, plus an option to acquire 10% of AMD’s stock. Those MI450 accelerators will power a 1 GW OpenAI supercomputer next year. In parallel, OpenAI partnered with Broadcom to co-develop custom AI processors, targeting ~10 GW of in-house silicon by 2026. Read more / Read more
🦾 Salesforce rolls out Agentforce 360, its AI “agent” platform. Salesforce announced global availability of Agentforce 360, a platform that lets companies deploy AI agents across Salesforce’s cloud apps to automate tasks. CEO Marc Benioff says Agentforce “connects humans, agents and data on one trusted platform,” essentially embedding AI helpers into CRM workflows. The system already has 12,000 customers using it (including Reddit, OpenTable, and Adecco) during early access. Read more
🧪 AI Research of the Week
Less is More: Recursive Reasoning with Tiny Networks
From Samsung SAIL Montréal
Jake’s Take: Samsung researchers built a new AI called a Tiny Recursive Model (TRM) that can reason through complex problems better than models hundreds of times larger. Instead of one huge neural network, it uses two small ones that collaborate (one proposes answers, the other checks and refines them in a loop). That recursive setup lets it solve logic-heavy puzzles like Sudoku with minimal data and compute.
The theoretical breakthrough here is efficiency. If this recursive approach scales, it could help kill the “bigger is always better” assumption that’s driving today’s trillion-parameter AI arms race.
and then, even more news…
🎮 Elon Musk’s xAI pursues ‘world models’ for auto-generating games. Musk’s AI startup xAI is hiring ex-NVIDIA experts to develop world models (AI systems with physical intuition) to create realistic game worlds. According to the Financial Times, xAI brought on researchers Zeeshan Patel and Ethan He to help its AI learn 3D physics and environments. Musk boasted that xAI’s game studio will release a “great AI-generated game” by end of 2026. These world models could eventually power robotics too, showing Musk’s ambition beyond text bots to AI that grasps and simulates the real world. Read more