Hardware

Hardware Markets

Chips

  • What chips are being used at the frontier? Are SRAM machines hype? → Frontier Chips
  • Why does NVIDIA sell CPUs?
  • What innovations does the Vera Ruben bring?

Inference

  • If everyone is to have their own model weights, how will these be served? How hard is switching models for inference? Do you need to rent GPUs separately? → Serving Models
  • Why is cost to serve models decreasing so much? GPU prices don’t seem to be dumping. What are the other factors? → Inference Economics
  • Most of the inference efficiency improvements include accuracy loss, how do people know they are getting served the right model? → Inference Economics
  • What do inference libraries like vLLM do, really? → Serving Models
  • At a technical level, how do you serve a model? → Serving Models

Inference Markets

  • What actually drives the per-model price sawtooth? Needs a full record of each model’s providers + prices over time, not just the minimum.
  • How big of a market is OpenRouter inference? HuggingFace? → OpenRouter and Hugging Face
  • Who are OpenRouter and HuggingFace users? Why would you use these models? → OpenRouter and Hugging Face

Verified Inference

  • How does TEE work, really?
  • How big of a market is verified inference? decentralized? Volume / Revenue
    • Who are the participants?
  • What is the space of tests you an run to verify a model? Attestations? What’s actually detectable?
    • How degraded are these centralized provider’s models? Are they selling lemons?

Agents / RL / Fine-Tuning

  • What’s the difference between what Prime Intellect and Thinking Machines offers?
  • What are other products for RL/Fine-Tuning as a service
  • How expensive is fine-tuning? How much data do you need to steer the model?