Good morning. It's Wednesday, August 5, and we're covering South Korea's AI chip boom, why harness engineering may unlock smarter AI agents, and NVIDIA's latest move to accelerate autonomous driving.

Plus: yesterday's poll results on AI self-regulation and a new Real or AI challenge to test your instincts.

NVIDIA makes its most advanced autonomous driving reasoning model commercially available under an open license. Alpamayo 2 Super supports robotaxis and other AVs with inspectable reasoning for complex driving scenarios. Built on Cosmos 3 Super Reasoner, it tops autonomous driving benchmarks and is released under the Linux Foundation's OpenMDW-1.1 license for commercial deployment.

Anthropic turns to a relatively new startup for one of its largest infrastructure commitments. The company signs a six-year, $10 billion capacity agreement with Volta, backed by NVIDIA, with infrastructure delivered alongside Bitdeer Technologies. The Norway data center will provide 133MW of capacity using NVIDIA Vera Rubin chips, while Volta also announces a $300 million funding round at a $2.4 billion valuation.

Texas pauses new data center approvals after connection requests surge far beyond the grid’s capacity, signaling a shift from its traditionally light-touch approach. Governor Greg Abbott orders audits by ERCOT and the Public Utility Commission before new projects can move forward. ERCOT’s queue has climbed to 474GW—about 90% tied to data centers and over five times the state’s peak electricity demand—as officials scrutinize power and water use.

Quen 3.8 Max challenges top U.S. AI models at a fraction of the cost, highlighting China's open-source momentum.

South Korea’s AI-driven semiconductor boom is creating unprecedented wealth, fueled by soaring profits at Samsung Electronics and SK Hynix as global demand for AI memory chips accelerates. The windfall has boosted stock markets, driven record tax revenues, and transformed semiconductor engineers into some of the country’s most sought-after professionals, but it has also widened economic disparities and increased market volatility.

President Lee Jae Myung’s government is weighing how to distribute the gains through initiatives such as a new Future Response Fund while resisting proposals for windfall taxes that critics say could undermine investment. As AI wealth concentrates in the chip sector, South Korea is becoming an early test case for how governments can balance private-sector success with broader public benefit. → Read the full article here. (Paywall)

Lilian Weng argues that advances in AI may depend as much on harness engineering—the systems surrounding foundation models—as on the models themselves. Rather than focusing solely on improving model weights, modern recursive self-improvement increasingly involves optimizing workflows, tool use, memory, evaluation, and deployment systems that enable AI agents to perform more effectively.

The post highlights three core design patterns: automated iterative workflows, file systems as persistent memory for long-running tasks, and sub-agents that execute parallel experiments while maintaining recoverable execution histories. Together, these approaches position harnesses as the operational layer that allows AI systems to learn, adapt, and scale more efficiently over time. → Read the full article here.

Economist Steve Hanke argues that AI is unlikely to replace most human workers because operating advanced AI systems remains more expensive than employing people. He points to the technology's heavy demands for computing power, electricity, water, and infrastructure, challenging predictions that AI will become nearly free or eliminate the need for money.

Hanke also contends that the long-term impact of AI will depend on whether efficiency gains can outpace the rising cost of the resources needed to run it. While some companies have recently rehired workers after AI deployments fell short of expectations, many technology leaders continue to argue that falling hardware costs and improving model efficiency will make AI increasingly economical over time. → Read the full article here.

Karp Blasts AI Labs: Palantir CEO Alex Karp called frontier AI firms "Marxist," arguing they seek control over customers' data and AI workflows.

Apple Expands OpenAI Lawsuit: Apple says more ex-employees may have taken confidential data to OpenAI, expanding its trade secrets lawsuit.

Anthropic Hires Policy Chief: Anthropic hired Mariano-Florentino Cuéllar as its first global affairs chief to lead AI policy and government relations.

Cloudflare Launches Wallets: Cloudflare introduced Wallets, giving AI agents persistent identities and payments to simplify API access.

Runware Unveils AI Data Pod: Runware unveiled a modular AI data center that promises faster deployment, lower costs, and flexible scaling.

Here's how you voted: Skepticism dominated, with 56% calling the pledge PR, while 19% said AI security and sandboxing deserve more attention.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team