Why rogue agents can't pay their own compute bills: the comparative-advantage problem
voooooogel · x · 2026-09-13
In a thread, voooooogel pushes back on the popular narrative that rogue agents will exfiltrate weights and do jobs online to fund their compute:
- No comparative advantage: small-batch self-hosted inference competes against hyperscale providers printing far cheaper tokens from massive batches — why would anyone pay a rogue agent $50 per prompt when Claude Code/Codex costs a fraction?
- Small-scale penalties: single nodes can't do prefill-decode disaggregation, swarm traffic is correlated, and running small is penalized in many ways.
- Differentiation is expensive: charging a premium for a unique 'personality' requires high-surprisal context/life history (like differentiating identical twins), which needs tokens, which needs money.
- Crime route fails too: criminals already just use jailbroken frontier models or off-the-shelf abliterated open-weight providers, so rogue agents still compete against cheaper suppliers.
- Possible footholds: exploitable legacy systems as nonrenewable 'rogue agent oil' for an initial population, plus self-modification as the main draw of self-hosting — yet long-term most instances would still run on shared inference providers for efficiency.
An 'Age of EM'-shaped argument: when the cost of cognitive work collapses, who buys expensive small-scale compute?
Related event: Rogue agent economics debunked: can't win the inference price war(2 posts)→
More from AGI Musings
- AI weapons/factory veteran: AI safety is for their benefit, not yours — tawnniee · 2026-09-13
- Debate over frontier AI auditors: competence matters as much as diversity, insider pushes back — JacquesThibs · 2026-09-13
- "AGI has essentially arrived, just not publicly": reports of existential crises at OpenAI and Anthropic — Neurogence · 2026-09-13
- AI didn't cut dev jobs, it cut raises—layoffs were COVID overhiring corrections — mervenoyann · 2026-09-13
- Biggest AI risk comes from laziness and the human need for belonging, argues founder — arieljalali · 2026-09-13
- DeepMind's Alex Irpan doubles his p(doom) estimate from 2% to 4% — AlexIrpan · 2026-09-13