AI:AM Highlights: Clean Neurons, Black October, the GPU Credit Crunch, and Buying the Supply Chain
Cognitive RevolutionAI Summary
→ WHAT IT COVERS Nathan Lebens and Prakash Narayanan host an 89-minute AI:AM highlights episode featuring Zvi Mowshowitz on recursive self-improvement timelines, Evan Conrad on GPU credit markets, and Lanus Hamali on physical supply chain investing, plus analysis of neural network interpretability advances and a leaked Anthropic federal AI bill. → KEY INSIGHTS - **RSI Timeline:** Zvi Mowshowitz argues that recursive self-improvement has already begun in slow-takeoff form, with broad consensus among people at OpenAI, Anthropic, and Google DeepMind that 2027 will be "completely crazy" and 2028 "unrecognizable." Nvidia data showing dramatically faster CUDA efficiency doubling rates in 2026 serves as concrete third-party evidence that AI-assisted technical work is accelerating well beyond prior projections. - **GPU Credit Bottleneck:** The primary constraint on AI compute buildout is not power or chips—it is credit. Lenders cannot underwrite GPU clusters as collateral because they cannot value or operate a supercomputer if a borrower defaults. This forces AI startups to raise astronomical funding rounds upfront purely to pay cash for multi-year compute contracts, creating survivorship bias where only massively funded labs can exist. - **Atoms Over Bits Investment Thesis:** Physical AI supply chain companies—silicon wafer makers, memory fabs, capacitor manufacturers—trade at three to five times sales while S&P averages 22 times earnings and Walmart trades at 40 times. Labs will never vertically integrate into wafer production the way they already compete with software design firms like Nvidia and Broadcom, making physical suppliers structurally protected with decades of irreplaceable tacit manufacturing knowledge. - **Neural Network Interpretability Progress:** The "clean neurons" paper by Goodfire's Nick Camerata suggests individual neurons are more monosemantic than previously believed, making interpretability roughly ten times more tractable. However, weak activations, attention mechanisms, and the KV cache remain unexplained. Critically, the highlighted "Colorado neuron" example was found to also respond equally strongly to tornado systems—an unrelated concept—undermining the paper's central demonstration. - **Memory Supply Constraint:** No credible expert projects DRAM bit supply can more than 2.5x by end of decade, while AI demand trajectories suggest 3x or greater growth. Memory companies currently trade at five to six times earnings, meaning even without price increases, a rerating to ten to twelve times earnings represents significant upside. Memory compression improvements from GPT-3 to DeepSeek v4 saw per-token cost drop 500x while DRAM bit prices rose six to seven times simultaneously. - **Anthropic AI Bill Critique:** A leaked Anthropic-drafted federal AI bill would immediately nullify state AI regulations while delaying federal implementation by six or more months, creating a regulatory vacuum. The bill's emergency intervention provisions require judicial approval with high evidentiary burdens, making them practically unusable in real emergencies. Zvi argues this reveals Anthropic prioritizing preemption of state laws over substantive safety provisions, weakening rather than strengthening the overall regulatory framework. → NOTABLE MOMENT Evan Conrad reveals that SF Compute applies a Harberger tax mechanism to GPU capacity—owners must keep a portion perpetually available for purchase by others at a self-declared price. This forces market liquidity and prevents hoarding, resulting in SF Compute remaining the only major GPU cloud not currently sold out amid industry-wide capacity shortages. 💼 SPONSORS [{"name": "OutSystems", "url": "https://outsystems.com/tcr"}, {"name": "Parallel", "url": "https://parallel.ai/tcr"}, {"name": "Tasklet", "url": "https://tasklet.ai"}] 🏷️ Recursive Self-Improvement, GPU Compute Markets, AI Supply Chain Investing, Neural Network Interpretability, AI Regulation, Memory Semiconductors
