The Thermodynamic AI Chip

Value of the Information & Strength of the Argument

The video provides valuable insights into the intersection of AI and hardware design, particularly the innovative approach of using AI agents for chip design and the emerging field of thermodynamic computing. Ahle’s arguments are well-reasoned, drawing on his extensive experience and specific examples like the Verilog simulator project. He presents a balanced view, acknowledging both the potential and the limitations of current AI capabilities. The discussion on ‘understanding debt’ and the distinction between passing tests and true correctness is particularly thought-provoking. However, some claims are anecdotal and the conversation is exploratory rather than presenting rigorous experimental evidence.

The title accurately reflects the main focus on thermodynamic computing chips, though the conversation covers broader topics like AI-driven chip design and formal verification.

8/10

High-quality discussion by an expert in theoretical computer science and chip design, with references to specific papers and projects. However, some claims are anecdotal and the discussion is partly speculative.

• 00:00Meet Thomas Ahle: the Lovable for chip design
• 03:41Why hardware needs formal verification
• 06:36Ten thousand dollars per core and a six-month agent run
• 07:40Rebuilding programs from tests: ProgramBench and zero percent
• 12:15Structure vs competence: can you learn a program from behavior?
• 15:27Continual learning, abstraction, and Claude as an ecosystem
• 23:17Autoformalization and the AlphaProof trick
• 29:31No single true representation: specs, Petri nets and TLA+
• 34:43Thermodynamic computing: when noise is the computation
• 37:32Bayesian uncertainty in the age of token streams
• 41:12Hybrid compute: vibe-coding loops, binaries and Stockfish
• 44:44Co-design, central-AI apps and API pricing
• 49:45Chain of thoughtlessness and the Chomsky hierarchy
• 53:40AI psychosis, slop and the broken social contract
• 57:34Typing it yourself, teamwork and performance vs competence

• Subsets and Supermajorities: Optimal Hashing-based Set Similarity Search — Ahle's earlier work on hashing algorithms.
• Clustering the Sketch: Dynamic Compression for Embedding Tables — Related to Ahle's research on embedding compression.
• ProgramBench: Can Language Models Rebuild Programs From Scratch? — Benchmark discussed for testing AI's ability to rebuild programs.
• Autoformalization with Large Language Models — Paper on autoformalization using LLMs.
• Autoformalizing Memory Device Specifications with Agents — Paper on autoformalizing DRAM specifications.
• Thermodynamic AI and the Fluctuation Frontier — Key paper on thermodynamic AI.
• Thermodynamic Computing System for AI Applications — Paper on thermodynamic computing systems.
• Thermodynamic Linear Algebra — Paper on thermodynamic linear algebra.
• An efficient probabilistic hardware architecture for diffusion-like models — Paper on probabilistic hardware for diffusion models.
• Chain of Thoughtlessness? An Analysis of CoT in Planning — Paper on chain-of-thought limitations.
• TLA+ — Specification language mentioned for formal verification.
• Stockfish — Chess engine used as example of hybrid compute.
• Building an Open-Source Verilog Simulator with AI: 580K Lines in 43 Days — Blog post about the Verilog simulator project.
• Normal Computing Announces Tape-Out of the World's First Thermodynamic Computing Chip (CN101) — Announcement of the CN101 chip.
• World's first thermo chip reaches tape out — News article about the CN101 chip.
• Math Does Not Represent — Video referenced in discussion about representation.
• DRAMBench: Autoformalizing DRAM Specifications with Timed Petri Nets — Blog post on autoformalizing DRAM specs.
• Extropic, Normal Computing, and D-Wave? — Comparison of thermodynamic computing companies.

• Thermodynamic AI and the Fluctuation Frontier — Supports the concept of thermodynamic computing.
• Thermodynamic Computing System for AI Applications — Provides technical details on thermodynamic computing systems.
• Autoformalization with Large Language Models — Supports the discussion on autoformalization.

• ProgramBench: Can Language Models Rebuild Programs From Scratch? — The video discusses the benchmark's results, but the paper may present a different perspective on the success rates.

• app.rescript.info
• normalcomputing.com
• thomasahle.com