Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push EP 299

The Human Control Over AI Act would restrict recursively self-improving models until federal guardrails exist and require shutdown controls and independent auditors, while Richard argues that enforcing such a ban could require surveillance of private AI use. The discussion also contrasts those concerns with a speaker’s preference for targeted safeguards, such as cybersecurity and enforcing existing research rules, rather than restricting the technology itself.

The mosquito executive order sets 2028 targets to reduce invasive mosquitoes in Washington, DC, by at least 90% and ticks by at least 50%, prioritizing methods such as sterile insects, genetic modifications, and beneficial bacteria over conventional pesticides. Speakers also describe genetic approaches as a way to prevent disease transmission while preserving carrier insects, against a backdrop Richard puts at roughly half a million malaria deaths annually.

Speakers describe drones as increasingly consequential in both border security and warfare: one estimates about 10,000 drones cross the Mexico–US border each month, while another says Russia and Ukraine will each make 10 million drones this year. The border estimate is used to question reliance on walls, and the production figures illustrate the scale of drone use in the Russia–Ukraine conflict.

Richard expects AI to connect fragmented scientific specialties using language models, digitized data, simulations, robotic process automation, and agent swarms, and he invokes the bitter lesson in arguing that data- and compute-intensive neural networks can outperform systems built around handcrafted expert ideas. He also describes virtual models as an important tool where outcomes can be simulated and verified.

Richard predicts that multiple simpler, potentially single-gene diseases could be cured within 12 months, while saying progress on more complex diseases will depend partly on compute investment and that FDA approval and long-term human trials can add roughly five to ten years. He points to Peril Bio’s FDA approval to skip animal trials after its stem-cell-derived lymph-node organoids showed drug and toxicity responses that were more predictive of human responses.

The discussion contrasts biology professors’ reported low day-to-day AI use with Richard’s view that larger biological datasets, including perturbation studies, can make neural networks increasingly useful. It also notes that some physics questions require data from large hadron colliders, making those studies expensive.

The Brain It system is described as reconstructing images viewed in an fMRI scanner and using predicted responses to unseen images to generate training data, but speakers say fMRI’s roughly one-second temporal and cubic-millimeter spatial resolution limits decoding. Separately, a speaker says learning how biological brains change synaptic weights could make neural-network training 10 to 1 million times more efficient than current gradient-descent training with large numbers of examples.

The host reports that 48% of participants mistook Tavis’s Griffin AI avatar for a real person in a live-video interaction, compared with 3% for previous systems. Speakers say very inexpensive interactive video could automate some face-, voice-, and real-time-interaction service jobs, while Richard warns that convincing video calls could enable executive impersonation and fraudulent wire transfers, citing a reported $20 million scam.