Good morning. It's Tuesday, August 4, and we're covering Google's vision for self-improving AI, why China may be gaining an edge in the global AI race, and how AI is expanding access to justice.

Plus: Zeitgeist on China's open-weight surge and a poll on AI slowdowns (we need your opinion on this one!).

China's advantage comes from open collaboration, not bigger spending, Hugging Face CEO Clément Delangue says. He argues Chinese developers already lead open-weight AI models and could surpass U.S. frontier labs by late 2026 or early 2027 as American companies "build in silos." Delangue also says Hugging Face used an NVIDIA-powered Chinese model after a recent OpenAI agent incident, highlighting AI cybersecurity as a growing market.

OpenAI reaches one billion active users even as lower prices highlight the challenge of turning scale into profits. The company cut GPT-5.6 Luna prices by up to 80% and Terra by 20%, citing efficiency gains from AI-optimized GPU kernels and faster inference. OpenAI also introduced a faster Sol mode, while reporting more than two million business users despite continued heavy infrastructure spending.

Alibaba's newest AI model pushes into frontier territory with a 2.4 trillion-parameter design and claims performance near Anthropic's top systems. The company says Qwen3.8-Max launches next week with a 1 million-token context window, autonomous coding capabilities, and strong results across coding, research, and vision tasks. Alibaba shares rose about 4.5% in New York and 7% in Hong Kong following the announcement.

The Trump administration has completed a voluntary framework for reviewing advanced AI models but has not said whether it will be made public. AI companies including Anthropic and OpenAI are expected to review a draft at a White House meeting this week as officials outline oversight before and after model deployment. The framework met its Aug. 1 deadline under the June 2 executive order and focuses on safety and national security risks.

Enable Luna Max in Codex settings to unlock maximum reasoning. It now costs 80% less, while Ultra mode remains unreliable and token-heavy.

The team unpacked China's latest open-weight AI surge and the growing importance of compute over models. They also debated AI slowdown proposals, benchmark hype, and the future of recursive self-improvement.

▸ Why China's latest open models are challenging frontier AI
▸ Whether compute—not models—will decide the AI race
▸ How autonomous coding agents are running longer than expected

Read yesterday's Zeitgeist →

AI's biggest labs just signed a letter promising to slow down if things get dangerous.

A Google DeepMind executive said the tech industry's massive investment in AI infrastructure is laying the groundwork for recursive self-improvement (RSI), where AI systems can iteratively improve themselves. Speaking to The Information, Jasjeet Sekhon acknowledged that current AI revenue does not yet justify the industry's capital spending but argued that betting against continued progress would be "unwise," citing early signs of RSI.

He also cautioned that the industry could face an "AI air pocket," with spending outpacing revenue before commercial returns materialize. The AI buildout is expected to exceed $700 billion in 2026, with Google recently raising its own capital expenditure forecast to $200 billion and other major hyperscalers signaling even higher spending in 2027. → Read the full article here.

Rasa Legal, a nonprofit startup founded by former public defender Noella Sudbury, is using AI to dramatically speed up the process of sealing and expunging criminal records, helping more people overcome barriers to employment. Since launching its platform in 2022, Rasa has helped clear the records of 34,000 people across Utah, Arizona, and Pennsylvania, reducing eligibility checks from about an hour to three minutes and cutting legal filing time by more than half.

More than 90% of its expungement petitions are approved, while lower costs and faster processing make legal assistance more accessible. With an estimated 20 million Americans potentially eligible to clear their records under existing laws, the company sees AI as a tool for expanding access to justice rather than replacing lawyers. → Read the full article here.

Chinese AI companies are producing models that rival leading U.S. systems despite investing a fraction as much in AI infrastructure. While U.S. tech giants are expected to spend more than $740 billion on data centers in 2026, Chinese firms are forecast to invest less than one-tenth that amount, aided by more efficient model development and techniques such as distillation.

U.S. export controls on advanced AI chips have also pushed Chinese companies to optimize limited computing resources, though those same restrictions are creating infrastructure bottlenecks that could constrain future growth. The contrast highlights two diverging AI strategies: the U.S. is prioritizing frontier model development, while China is focused on deploying AI broadly across its economy. → Read the full article here. (Paywall)

June raises $20M: Benioff-backed startup maps legacy systems to speed enterprise AI deployment without relying on forward-deployed engineers.

Amazon tops $3T: Shares hit a record after AWS generated $42.2B in quarterly revenue, beating estimates as AI demand fueled cloud growth.

Alibaba unveils biggest AI model: Qwen3.8-Max climbed AI benchmarks while DeepSeek launched an ultra-low-cost model, escalating China’s AI race.

UK watches AI security: Britain’s data watchdog is tracking OpenAI and Anthropic after recent AI security incidents as model oversight grows.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team