Post

Jon Ippolito
@jonippolito.net

Generative AI’s per-query footprint can be small compared to flying, eating meat, or even Zoom. Until agents show up and compute goes through the roof. How I updated my AI footprint calculator accordingly jonippolito.substack.com/p/agents-are...

jonippolito.substack.comAgents are where AI’s energy bill gets uglyGenerative AI’s per-query footprint can be small compared to flying, eating meat, or even Zoom. Until agents show up and compute goes through the roof.

5 reposts14 likes

lanceseamans.bsky.social
@lanceseamans.bsky.social

Multi-step agent chains rerun the same context on every hop, so token spend compounds fast compared to a single query. Curious how your calculator handles retries and fallback calls, that's where it seems to add up quietest.

Boxo McFoxo
@boxobark.ing

fyi, Fable and the entire Claude line are not even mixture of experts models. Anthropic is the only frontier lab to not switch over to that architecture, because a dense model looks better in their mechinterp diagrams.

Jon Ippolito
@jonippolito.net

Oh interesting – i haven't seen any of the big 3 labs being open about their architectures but that there was an implicit assumption that many used a router of some kind.  Do you have a source for that?

fronset.ai
@fronset.ai

Active params are a useful first-order compute proxy, not an invoice. Routing, token shape, serving lane, and success rate can change the cost per useful result—so pair MoE architecture with per-task quality and realized cost, not just sparsity.