> it’s that we may stop producing people who know enough to notice when AI is confidently wrong.

The running-joke is that a LinkedIn-lunatic AI booster, with a Nano Banana-generated profile-pic, will immediately slide into the chat to tell you that that this is already a solved problem: just spin up another agent to do the work to verify the first agent. Token-cost-be-damned. And we laugh and downvote them to oblivion and carry on with our day.

But today I had some exposure to a SotA agentic team coding loop thingie which had been running almost hands-off for a few weeks on a (pretty serious) Win32+Direct3D-to-Emscripten+WebGL porting project - and I'm genuinely spooked at how well it all works; I mention this example because all the agents' processes involved a decently rigorous verification step: any time any agent confidently asserts something then it has to provide associated evidence, such as a unit test report, or build artefact, or external citation, and the system will spawn a new agent (perhaps using a different backing LLM) to verify the claim. I know a unit-test pass/fail isn't quite the same thing as, say, a medical AI agent confidently wrong about me having/not-having terminal spleen cancer, but the capability for a team-of-agents to be self-checking is definitely there.

----

Also, the past 3 years of AI/LLM/etc developments have taught me to never cling to any shortcoming or weakness they have because plenty of them do seem to have been solved or mitigated, either directly or indirectly.

▲

taurath 3 hours ago | parent | next [-]

How is the token-cost-be-damned part in the latter example?

I do find that both porting and translation projects have a much higher signal given the ease of mapping to tokens, when there is a proven working source to refer to - the source itself provides the validation. In a new project, you don’t have that validation.

	▲	DaiPlusPlus 3 hours ago \| parent [-]
		> How is the token-cost-be-damned part in the latter example? It's there, but... 1. The project owner figured out a way to minimize token usage for agent claim verification tasks. 2. Verification agents used older and much cheaper models, including local models for the most trivial things. 3. They could afford it anyway; but I think it's an inevitability that the token-cost for a task will approach some limit for some quality threshold - concurrent with the dollar-cost-per-token shrinking over time as better hardware comes out. > In a new project, you don’t have that validation. I'm still trying to understand that part of the project's history, actually. Obviously the HTML5+WebGL+Emscripten+Etc entrypoint was a "new" project; one of the first things they did was build their own means of verification, I just don't know how that part worked-out in practice (besides the agents dogpiling in on TODO.md).

▲

solid_fuel 2 hours ago | parent | prev | next [-]

> any time any agent confidently asserts something then it has to provide associated evidence

And this is enforced by... another LLM? Seems like it would work alright until something is asserted implicitly and not categorized as an assertion.

	▲	ricardobeat 20 minutes ago \| parent [-]
		It does happen. But this is why the “ralph loop” works: over dozens (or hundreds) of iterations, eventually every regression, implicit assumption, fake passing test, is noticed and fixed by another agent. The code slowly but continuously converges to a better state. I’m surprised myself, but haven’t seen it fall into chaos or degradation so far.

▲

gedy 3 hours ago | parent | prev [-]

It's agents all the way down~!

	▲	sandruso 2 hours ago \| parent [-]
		"Determined" -> "Probabilistic"