1. OpenAI's Audited Books Leaked, Showing a $38.5 Billion Loss for 2025 Ed Zitron saw the documents first. The audited financial statements he describes put OpenAI's 2025 net loss attributable to the company at $38.53 billion.

2. An AI Chemist Improved a Hard Drug-Synthesis Reaction, Running With Little Human Help OpenAI and Molecule.one say a near-autonomous AI chemist, built on GPT-5.4, improved a difficult reaction used in medicinal chemistry.

3. The IPO-Summer Boom Is Running on 16 Percent Public Approval Capital is pouring into artificial intelligence faster than the public will vouch for it.

In Brief

• US delays DeepSeek blacklist while flagging 100-plus Chinese firms The US held off adding DeepSeek to its export blacklist even as it designated more than 100 other Chinese companies as security risks. The decision leaves the AI lab in regulatory limbo rather than cutting it off outright.

• Odyssey raised funding at a $1.45B valuation with Amazon backing World model startup Odyssey secured a $1.45 billion valuation in a round including Amazon. The company builds models that simulate environments rather than generate text, positioning it in the category labs are funding beyond LLMs.

• Anthropic opened a Seoul office and signed Korean partners Anthropic opened an office in Seoul and announced partnerships across Korea's AI sector. The move follows surging Claude usage in the region and gives Anthropic local sales and support presence.

• Google launched a $99.99 Gemini smart speaker Google released a $99.99 Home Speaker that swaps fixed Assistant commands for Gemini conversation. The hardware tests whether generative AI can revive a smart speaker category that stalled under rigid voice commands.

• Pinterest launched an experimental AI shopping app, Ask Pinterest Pinterest released Ask Pinterest, a standalone app that returns product recommendations through a chat interface. The app keeps the experiment separate from the main Pinterest product while it tests conversational shopping.

• DeepL acquired Mixhalo and opened a San Francisco office DeepL bought Mixhalo to add live-event audio streaming and real-time translation to its products. The translation company is opening a San Francisco office to grow its US business.

• Pramaana Labs raised $27M to apply formal verification to AI Pramaana Labs closed a $27 million seed round from Khosla Ventures to build formal verification for AI outputs. The company targets law, drug discovery, and tax preparation, where wrong answers carry direct cost.

• Nvidia used AI coding agents to direct robot training Nvidia ran a self-improvement program where teams of AI coding agents directed robots to install GPUs and cut zip ties. The agents wrote and adjusted the training routines with limited human input.

• AI labs are paying XDOF to collect robot training data Some AI labs now pay XDOF to gather physical-world data for training robots. The work addresses a data shortage that blocks physical AI from matching the progress of text models.

• TNO is building GPT-NL, a state-backed Dutch language model TNO, SURF, and the Netherlands Forensic Institute are developing GPT-NL, a Dutch language model with disclosed training data and governance. The project aims to give the Netherlands a model it controls rather than relying on foreign providers.

• GameCraft-Bench tests whether agents can build playable games end-to-end Researchers released GameCraft-Bench, a benchmark measuring whether coding agents can turn natural-language specs into working games inside a real game engine. It evaluates the full chain of scripts, scenes, assets, and runtime interaction.

Read the full edition →