Finding Peak w/ Ryan Hanley

A handful of genuinely useful concepts surface - the token-volume metrics, the 'loops' framing for persistent agents, and the 'service as software' inversion - but they are buried under extended host monologues, AI-optimism cheerleading, and recycled 'think bigger' exhortations that consume the majority of runtime without adding new information.

“it would take uh, eight hours of reading per day for about 36 years to read what happened in one month”

“you can outsource your work, but you can't outsource your understanding”

'Service as software,' 'taste and judgment as the moat,' and the inference-only lifespan of agents are decent framings, but the episode also recycles the printing-press/car/internet analogy, the 'future is unevenly distributed' Gibson riff, and generic AI-doomer rebuttals that circulate everywhere in this discourse.

“it's not software as a service anymore, it's service as a software”

“the work itself can be performed by these AI agents, but the, the ideas, the taste, the reasons behind what we're doing, that is still what we have to communicate”

Brian McAnulty is a genuine practitioner - he shipped a real AI coach in 2023, runs 3B+ tokens/month through coding agents, and is actively building commercial products - but he is a small-team indie founder rather than someone who has operated at significant scale, limiting the depth of hard-won operational wisdom.

“I remember at the end of last year I was using 500 million tokens per month in these coding agents”

“the kind of prompt that I gave like an AI agent, uh, for building that kind of search engine was not like a couple sentences, um, it was like a 20ish page or so prompt”

Token-consumption figures (500M → 1B → 2B → 3B/month) and the derived reading-time calculation are concrete and memorable; the $300 - 400 cost to prototype Black Inc and the 45-minute connector build add useful benchmarks. However, many claims - about what agents can 'do now,' model capability leaps, and business outcomes - remain asserted without data or named evidence.

“I remember at the end of last year I was using 500 million tokens per month...in the beginning of this year it was like a billion tokens per month. And then I remember hitting like the next month was 2 billion. Now it's like over 3 billion”

“it cost me maybe 3 or $400 in tokens over the course of a few weeks”

The host frequently answers his own questions with multi-paragraph monologues before passing to the guest, asks leading 'is that what you mean?' confirmations rather than probing follow-ups, and never meaningfully challenges the guest's claims or pushes for harder evidence; one or two decent definitional follow-ups ('can you just explain the idea of a loop?') are not enough to offset the pattern.

“So you're not just someone who is using AI to build. You have this very unique business that you built in Lash Loop where you're actually helping other people build with AI as well.”

“Is that a proper way of framing this, do you think?”