Good morning. It's Wednesday, August 12, and we're covering a billion-dollar bet on open-source AI, Anthropic’s latest push into mathematical research, and xAI’s new agentic bots.

Plus: yesterday’s poll results on free AI, a new poll on AI-generated watermarks, and a Real or AI challenge.

Grok Bot gives AI agents their own cloud computers, letting them work across apps and websites without clean APIs. xAI launches the beta as always-on AI teammates that handle tasks end to end and seek approval only when needed. Bots can coordinate in parallel, learn workflows, and retain context. Access starts with SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium.

NVIDIA’s latest AI model can run on a single GPU, putting advanced capabilities within reach of laptops and desktops. The company releases Nemotron 3.5 Lightning as its first open-source model since CEO Jensen Huang began publicly backing open AI. Tested by CrowdStrike and others, it targets autonomous agents and is available through Hugging Face and NVIDIA.

Manus is preparing to resume operations as an independent company, requiring some users to back up and restore their data. Affected users must back up before 7:59 a.m. SGT on August 23, with restoration opening at 8:00 a.m. August 25. Manus says affected users will not be charged during the transition and will receive welcome-back bonuses.

Anthropic is adding watermarks to AI-generated text, with the marker surviving some copy-and-paste and editing. The change covers Claude models released after August 2 to comply with the EU AI Act’s transparency rules. Anthropic also plans to support older models and watermark files using C2PA. The system applies across Claude, Claude Code, Cowork, and its API.

Anthropic is adding invisible watermarks to Claude-generated content following new EU rules. Should every AI company be required to do the same?

Mark Zuckerberg just called out Dario over his centralized AI approach, arguing that open access is safer and more empowering.

Igor Babuschkin, a former xAI co-founder and researcher at OpenAI and Google, has launched River AI with $1 billion in funding to develop open-source artificial intelligence people can own, modify and run locally. The Palo Alto startup plans to build home servers that let individuals and businesses keep their AI interactions and private data on their own hardware, rather than relying on centralized systems.

Babuschkin also wants users to “retrain” AI through simple commands, changing how it behaves for their needs. River AI, which has about 20 employees, is still months from releasing its first open-source technologies. The effort comes as Chinese startups gain ground with open models while U.S. AI companies largely keep their most advanced systems closed. → Read the full article here. (Paywall)

An unreleased Anthropic model made significant progress on the Riemann hypothesis, a 150-year-old mathematical problem tied to the distribution of prime numbers and backed by a $1 million prize for a complete proof. The model tested 650 approaches across 60 subagents over roughly 36 hours, generating 31 million output tokens; two agents developed the key ideas, which Anthropic’s mathematicians then confirmed.

The work does not solve the hypothesis, but it substantially raises the range for which the conjecture has been verified. The result adds to a growing list of high-level mathematical advances by large language models and intensifies debate over how AI should change mathematical research and authorship. → Read the full article here.

NVIDIA is investing heavily in Nemotron 4, a new open-source AI model that employees expect could contain at least 1 trillion parameters and compete with the strongest open models. The strategy could expand demand for NVIDIA GPUs by making capable, relatively low-cost AI systems available to more companies, but it also puts NVIDIA in competition with customers and startups it backs, including OpenAI, Reflection AI and Thinking Machines Lab.

NVIDIA has raised multiyear cloud commitments to $28 billion through 2031, roughly triple the year-earlier figure, to support its growing compute needs. The release timing remains uncertain, with employees estimating late fall 2026 or later after a final training run that could take months. → Read the full article here. (Paywall)

Brad Lightcap Leaves OpenAI: The longtime executive is departing after eight years, adding to a string of exits as OpenAI reshapes its top ranks.

Google Tests AI-First Homepage: A redesign highlights image creation, file questions and brainstorming while pushing traditional search back.

OpenAI Buys Back $7B in Shares: The tender offer values OpenAI at $852B, giving employees liquidity while the company weighs a potential IPO.

River AI Raises $1.1B: The xAI co-founder’s startup will expand tools that let enterprises customize AI models using their own data.

Target Names First AI Chief: Chandhu Nair will lead a coordinated AI strategy spanning inventory, customer shopping and faster business decisions.

Here’s how you voted: The bottleneck won, with 47% saying AI’s real constraint will shift elsewhere, while 24% expect us to use 10× more AI.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team