Everyone forgot about Sora, the new hit of the week is Groq. The team created a custom ASIC for LLM, which allows generating ~500 tokens per second. For comparison, GPT averages 30 tokens/s.

VC @ cyber.fund

Generative AI
Decentralized identity and reputation 
Network state

- a chatbot: one request, one response
- a simple RAG makes 2-3 requests (search, reasoning, generating a response)
- complex chain of / tree of thoughts can makes 10 requests (clarifying the question, deciding on a response strategy, generating candidates, selecting the best)

Thanks to custom chips (don't forget about Sam's $7 trillion and Masayoshi Son's $100B raised for AI chips), this number can increase to hundreds or even thousands of logical steps, internal checks, fact clarifications — while keeping the response time under one second.

The ability to instantly and conditionally free read, analyze, and generate dozens of pages of text improves AI system performance. For clarity, think about the metric "LLM requests per user task":