What Extropic is building
A startup called Extropic claims to be building novel analog hardware that harnesses physical randomness and thermodynamic effects—using superconducting Josephson-junction chips and transistor-based variants—to accelerate probabilistic models for AI more efficiently than conventional digital GPUs. Commenters are sharply divided: some see the approach as a plausible extension of energy-based models and analog computing research, while many others view the jargon-heavy messaging, lack of concrete benchmarks, and founder’s social-media persona as red flags suggesting hype or even snake oil. Overall, people agree the underlying idea of specialized, noise-tolerant hardware for AI is intriguing, but doubt whether Extropic’s specific implementation or timelines will prove commercially meaningful.
What Extropic Claims to Build
- Many interpret the “thermodynamic AI” pitch as:
- A hardware platform that uses physical noise/fluctuations for probabilistic computation.
- Energy-based models or probabilistic graphical models implemented directly as stochastic analog circuits.
- Initial superconducting Josephson-junction chips at low temperature, plus a future transistor-based, room‑temperature line.
How Commenters Understand the Technical Idea
- Several see it as analog computing tailored for AI, possibly doing matmuls / sampling in analog.
- Others frame it as “baking probabilistic models into hardware” to accelerate sampling and inference.
- A minority misread it as mainly a better random-number generator; others correct that this is too narrow.
- Links are shared to related “thermodynamic computing” and probabilistic hardware work, suggesting some prior art.
Skepticism and Hype Concerns
- Many call it buzzword-heavy, opaque, or “new-agey,” noting the blog post is hard to parse.
- Doubts that RNG or sampling is a real bottleneck for mainstream AI, compared to matrix multiply and memory bandwidth.
- Several compare the rhetoric to long-running quantum computing hype and worry about another “just around the corner” story.
- Concerns about superconducting hardware practicality and system-level energy efficiency (cooling costs).
Questions About Practical Impact
- Unclear whether such probabilistic hardware can match or beat current LLMs and diffusion models in quality or cost.
- Some ask for simulations of their models on conventional hardware and concrete benchmarks (e.g., speedup or cost vs GPUs).
- Others note potential niches: MCMC, probabilistic finance, PGMs, neuromorphic/thermodynamic learning.
Communication, Branding, and Social Media Persona
- Several argue the marketing style (dense jargon, “full-stack,” Spotify widgets, meme-heavy founder persona) undermines credibility.
- Others counter that the founders are technically strong, analog ML is a reasonable bet, and ambitious hardware efforts deserve patience.
- Overall tone: intrigued but heavily skeptical, with most commenters wanting clear demos and simpler explanations before believing the claims.