TLDR AI 2026-01-08

OpenAI has introduced ChatGPT Health, a new feature designed to assist users with health and wellness by securely integrating their medical information into a personalized ChatGPT experience.

OpenAI set aside an employee stock grant pool equivalent to 10% of the company last fall, when it was valued at $500 billion. The startup has already given $80 billion in vested equity. It is in preliminary talks with some investors about raising funds at a valuation of around $750 billion.

Anthropic plans to raise $10 billion at a valuation of $350 billion. The funding round is expected to close in the coming weeks, and the total amount of the deal could change. AI companies collected a record $222 billion in funding last year. This year is slated to be a big year for public offerings.

Recursive Language Model (RLM) is a flexible solution to long-context limitations in LLM agents. RLMs avoid summarization by delegating memory management to scripts and sub-models.

Context graphs shed light on how work really gets done in the enterprise. AI agents are a major unlock with automation, but only if their reasoning is grounded in the right enterprise context. It's not enough to just understand the enterprise data, systems also need relationship knowledge. Context graphs enable agents to learn and automate distributed work by capturing processes and learning intent over time.

Chinese companies are making strong, open AI models that are applying increasing pressure on the US AI economy. Dethroning Qwen in adoption in 2026 looks impossible, but there are areas of opportunity. The US could very well have the smartest open models again in 2026, even if they're used far less across the ecosystem. This post contains 8 plots that show the impact of Qwen, DeepSeek, Llama, GPT-OSS, Nemotron, and other new entrants to the ecosystem.

Most companies aren't getting the AI results they want. This Forrester study shows how an overemphasis on individual productivity, instead of collaboration, could be the culprit. For a look at how collaboration-first AI lives up to the hype, join the Miro webinar on January 22nd. Register now.

NitroGen is a generalist gaming agent trained on internet gameplay videos capable of predicting controller actions from pixels and adapting to unseen games through post-training.

PDFs have been largely ignored by the open source AI community, despite being a source of high-quality data for pretraining. Hugging Face has created a new pretraining dataset of more than 3T tokens spanning over 1,000 languages extracted from PDFs. This post details how the dataset was created.

WebGym is a large-scale environment with nearly 300,000 real-world web tasks for training visual agents. It has a high-throughput RL system that boosts training speed and a fine-tuned Qwen-3-VL that outperforms GPT-4o and GPT-5-Thinking on unseen web navigation tasks.

Brendan Foody, the youngest unicorn founder and CEO of Mercor, is transforming AI training by hiring experts like poets and economists to create evaluation frameworks for frontier models. Mercor pays top talent to teach AI models that are rapidly improving in economically valuable tasks.

Screenless and always-on AI devices will fail, while mini-apps enable UGC personal software and a US consumer super-app. By 2030, consumer AI will split into a reliable task assistant and an AI companion, performance marketing will collapse, and AI-generated virtual creators will explode. The biggest win will go to whoever solves AI discovery beyond chat and search.

Making AI agents doesn't have to drag on. Read this eBook from Algolia to see how you can deploy branded conversational AI agents while meeting business-sensitive timelines. Get the PDF

AI agents and MCP servers enhance threat modeling by automating the synthesis of organizational context, enabling security teams to map threats, identify detection gaps, and target high-priority threats efficiently.

Google now lets educators create podcast-style audio lessons using Gemini in Google Classroom.

The breakneck speed at which AI tools are being integrated into existing systems and workflows without proper threat modeling, testing, and security analysis is dangerous.