Mutt Technologies

5d ago

Story of Spawn, thought to be the company spending the most on per-capita AI costs in the world • A 10-person company performing 77,000 deployments per month. Over 100 times per day • Spent 3.1 million dollars on tokens last month alone • All operational work is performed automatically by agents without human approval (human-in-the-loop) • Before Fable appeared, this method didn't work. Other models failed for various reasons. Much better after Fable 5.1 came out • Designed so agents converge toward specific goals (like ants in a swarm) without explicit instructions—i.e., when merge conflicts occur, they rebase themselves • Do coding agents make mistakes? Yes. Then they automatically recover from pre-deployment test failures or post-deployment bugs. Naturally they also perform adversarial review. Rollbacks occur in only 0.9% of all deployments. • Deployments must happen 24/7. If deployments stop for an hour, it's considered a top-level incident. ㅋㅋ • What features are needed next? The author suggests two things. First is reflecting user preferences (most steering was about aligning user preferences with Claude). Second is entropy (or imagination). • It will be difficult for existing large companies to adopt this work style. This closed-loop approach inherently contains the risk of breakdown. You have to accept this. Also, don't force AI models into a human-centric framework. Personally, I expect that this extreme token-maximizing could be done with individual-level capital within a year, but before then we'll need to quickly adapt to this agent swarm approach. Isn't it an amazing time to start a startup https://x.com/jsnnsa/status/2115623382721126826 Korean version: https://rosettalens.com/s/ko/how-i-built-the-most-ai-pilled-company-in-the-world

475#AIAgentsShow Korean original