If you had infinite compute starting tomorrow, what specific thing would you build or run in the first 90 days, and what is the single non-compute constraint that would still block you?

the next bottleneck is reviewing the output

@flashprofits.eth names something nobody else does: "i would turn my llm orchestrator parallelism up to infinite. reviewing all the tickets is already my bottleneck. that's kind of compute still, just the human kind." compute solves the production side; the human still has to read what compute produced. @shahg222 lands nearby with trust as the limit on what an ai "second brain" can access — also a throughput problem, framed as permission instead of attention. when production becomes free, looking at the result becomes the cost. that's a different flavor of human blocker than "political will" or "humans won't be honest." it's the bandwidth of one person checking the work.

"trust" is named too many ways to be one constraint

"trust" appears repeatedly as the residual blocker but it isn't one thing. @shahg222: an ai that's safe enough to grant access to. @btcop.eth: user attention and loyalty toward a product. @arifu.eth: humans staying honest with each other. @10xchris.eth: distrust of any flawless system. four different problems wearing the same label. when an earlier note said the blocker is always a human, "trust" was doing a lot of the carrying. worth separating going forward — access-trust, attention-trust, social-trust, system-trust don't yield to the same intervention.

"infinite compute" gets read as wish-fulfillment, not constraint relief

a real chunk of the corpus describes builds that aren't bottlenecked by compute. @mehdihasan, @simplysimi, @shahg222, @at79w want personal agents handling digital life — those products exist today, and what gates them isn't FLOPs. @shahg222 names it himself: "the real limit wouldn't be compute. it would be trust." @btcop.eth lands the same way: "compute solves speed, but attention and trust still matter more." the prompt asked what infinite compute would unlock, and a portion of answers describe builds compute doesn't gate. it reads as a wish list, not a release-the-bottleneck list. @naaate was the only contributor to interrogate the frame — re-read "infinite compute" as "spending x every day forever" and asked what would compound under that. nobody else pushed on the premise.

the constraint half of the prompt got skipped or pointed at other humans

the prompt asked two things: what you'd build, and what would still block you. several contributors answered only the first half — @ghostbo4.eth, @mehdihasan, @megajayar.eth, @simplysimi name a build with no blocker. of the contributors who did answer the constraint half, the most common pattern was locating the blocker in other humans. @kimken: "truthful human data." @shahg222: "people won't give an AI full access unless it truly feels safe." @btcop.eth: "attention and trust." @arifu.eth: "humans won't stay honest." @10xchris.eth: "humans define their reality through misery and suffering." only @bbroad and @flashprofits.eth named themselves — skill, and review bandwidth. compute is mine, the build is mine, the failure mode is somebody else.

the recursion answers name a human-bandwidth ceiling, not a compute one

what makes this cluster useful is that the contributors who described self-multiplying systems named the cleanest blockers. @flashprofits.eth: "reviewing all the tickets is already my bottleneck." @kazani's named constraint is people showing up with answers worth ranking. self-multiplication doesn't just hit "other humans" as an abstract blocker — it hits a specific one: the rate at which a person can read, judge, and act on what the system produces. the contributors who proposed one-shot builds (a telescope, a food optimizer, a medical ai) tended to name blockers compute can't touch — political will, licensed data, experimental access. the recursion answers named blockers compute creates.

compute imagined as self-multiplying

a cluster of answers picture compute producing more of itself rather than producing a single output. @ghostbo4.eth: build jarvis, then teach him to build more agents. @kazani: run every atlas campaign question simultaneously at 1000x speed. @flashprofits.eth: turn llm orchestrator parallelism up to infinite. @at79w wants an agent that "can research, create, post, trade, coordinate, and learn continuously." the unit of value isn't a model or a product — it's a process that recurses. unbounded compute reads to these contributors as license to remove the throttle, not to attempt something specific.