Unverified50% confidenceFactExact time
The experiment used expert iteration, with about 440 decisions per round, running overnight on a rented GPU at a total cost of about 40 US dollars.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe model's key assumptions include: reviewer calls are 2000 input + 300 output tokens; system-under-test calls are 1500 input + 500 output tokens; a mid-tier model ($2/$10 per million tokens) is used by default; offline runs use the batch API for a 50% discount; online review uses standard pricing; human review is 4 minutes per case at $45 per hour; tooling costs follow Langfuse list prices; and the loaded cost of US full-time labor is $16188 per month.68% similarUnverifiedFine-tuning a 7-billion-parameter model can be done on a consumer-grade GPU, potentially at a cost of only a few hundred dollars68% similarUnverifiedCalculated with 500 agents, 2 LLM calls per step, and 10,000 steps, a total of 10 million LLM API calls would be required, and the total cost estimated at GPT-4 pricing could reach tens of thousands of dollars.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1050042API
curl https://kongchang.com/api/v1/knowledge/claims/1050042MCP
get_claim(id=1050042)