ian_j_butler

Right, I don't think researchers are confused on this point.. the anthropic piece is good outreach / science comms. OTOH this thread has like 200 comments and no mention of faithful/faithless reasoning. The idea that "of course the models can reason and here is the proof/artifact" is probably closer to the general understanding. That's kinda the whole setup for TFA and all the rest of the thread.

But the nuance under discussion here is exactly the kind of stuff you people take for granted in the AGI or reasoning threads. If it's practically relevant for tools/workflows with claude code, it's a good angle, maybe people are more willing to pay more attention to the details.

It's well-known that the reasoning model output is not necessarily faithful to the content of the thinking scratch pad anyway, even if you had it unsummarized and available verbatim.

Setting aside coding agents.. we really need this information to even pretend to evaluate the claims of stuff like mathematical breakthroughs, which is exactly why we will never see it. Very embarrassing to get the right answer for the wrong reason. But to give the models some credit, you could argue that even paying too much attention to the thinking is misunderstanding how CoT works. The argument would be that thinking in LLMs isn't really thinking, that it's self-reinforcement and circling to to encourage stability around beneficial attractors instead of degenerate ones. Can't have it both ways though: either the thinking is thinking and so it should be correct. Or the thinking is NOT thinking, and it's NOT real justification for the outcome, and these systems are even more hopelessly opaque than we usually assume.

And that's the real tl;dr. Hybrids win whenever anyone actually checks. To really be scientific we still have to check, but.. why wouldn't they? Probabilistic AI brings intuition/learning but can't plan/search. Classical brings planning and search, but has no intuition or learning.

Like US politics now, AI talk is just a team sport where there's no room for nuance. If you want to talk about such things then you need to anticipate this and work around it from the jump, making clear the expectations about the scope of the discussion and ground somewhere concrete. For politics it's gotta be economics, data, specific policy. For AI, it's gotta be research, benchmarks, very specific workflows or use-cases.

Even establishing the scope of the discussion is tedious, and most people have no patience for that, so the practical choices are basically having a dumb rah-rah type of discussion, or none at all. I'm breaking my own best advice by even writing this down but the only answer is: Read More, Talk Less. If you're in school and talking with other students or if you're talking to colleagues with known background in the context of known practical problems they would like to solve.. it's different. But the forums? Read-only is not even effective as far as "checking in" on the "collective culture" in engineering to get an accurate picture. Actually engaging is even more effort for even less profit. I regret having opened the thread, so back to burying my head in more research and hands-on projects :)

> I think we’re all used to user-hostile software these days, but this is the first time I’ve experienced software that feels like it’s actively trying to be disrespectful.

Always wild to hear people say stuff like this. First, all user-hostility is clearly disrespectful by definition. Second, almost all software, even the free stuff, is insanely user-hostile. We are all so completely frog-boiled on this it's not even funny. Yes, even people in tech and maybe especially people in tech.

Everyone reading this has probably used 10 applications today that are completely ignoring instructions to disable updates/telemetry if they even bothered to lie to you that this is possible. IOS has years-old "bugs" where turning off voice control isn't actually possible, official docs are gas-lighting you, and the settings are just ignored.. so people just deal with paused music that inconsistently triggers on 1/5 of your sneezes or coughs and get used to it. Spotify performance/ux/sanity has been completely degraded for months now. Web-browsers routinely force updates to require multi-gig downloads of AI models, and before that, they had on-and-off regressions in basic stuff like copy/paste for multiple years. Your popup-blocker that helps you to stay sane feels fine about popping up some shit that tells you how many pop-ups it freaking blocked. This is just my last 10 minutes. You can dig into any one of these problems, lose 45m on some janky fix, and also know for sure that you'll need to spend the same effort on some related goddamned problem less than a week later.

Besides the "ads for paying customers" type of stuff, this drip-drip of millions and millions of little points of persistent friction never stops. You think you broke it or you are going crazy until you deep-dive the bug reports or the reddit threads and realize it's all gas-lighting, and someone has made a choice. If the choice wasn't about disrespectful surveillance, auto-updates, or profit-maxxing enshittification then it's a greenhorn developer refactoring something for devx or aesthetics over UX, and the breakage didn't even happen in service of a real feature.

You try to freeze the apps with snap or containers or whatever for some stability hoping to GTFO the fix-it-again treadmill. You assert proudly that "Computers work for me, I don't work for them!" It's smoother for a while but there's always something. A phone-home with a suddenly bad endpoint, a missing remote tag/version gets yanked, or the operating system itself will betray you with yet-another iteration of unnecessary path-changing nonsense that breaks everything anyway.

Although they are opposites in every other way, Linus and Bezos may be the last living bosses that valued stability, backwards compatibility, and not fucking up shit that works fine. When they are gone god help us all.

Challenging questions for short answer format! But there's a pattern with this stuff that's always mostly the same. Specialist logics work great on the happy path, but when you dig into it, they always work because of a bunch of simplifying assumptions that keep us within a zone that's sound/complete/decidable.

Roughly speaking GOFAI usually "solves" the frame problem only to the extent it accepts hard limits on the domain of discourse, which trade has bad consequences for expressiveness and/or the ability to update things like axioms and beliefs. We tolerate this because if we don't then we're off the happy path and basically doomed. Then time passes, we forget that we agreed to tradeoffs and float the idea of possible generalizing/taming/adapting ;)

Not expert enough on SL to really connect the dots, and it's complicated by the fact that the BI family is a weird place that is in some sense neither monotonic nor non-monotonic. Two main issues though. The first is the applicability thing, because lots of stuff just will not allow clean partitions anyway (back to causal and ramifications). Even in the restricted domain of program-logic, if we're talking side-effects and outside the world of theorem provers or something, the real domain is causal logic anyway, plus all the irreducible difficulties of distributed consensus, etc. The second issue is the partition itself, which is fine if you bring it with you, but it has to be discovered otherwise. For problems of unknown structure, I'm guessing that immediately puts you in the position of looking at the exploding space of all possible partitions. Adding an algebraic operator makes for very clean notation but in practice you can't just wave away how difficult that actually is, if it's even decidable in general (and I'm guessing it's not).

Turing/Godel/Tarski will all have their due.. so back to the argument for hybrid systems. Heuristics can prune a search space, but what can't be decided ultimately needs some kind of intuition. Which isn't to say that we don't need all kinds of sophisticated heuristics and backtracking and the rest, but just that we need everything we can get for hard problems. Another way of looking at it, GOFAI and new-wave AI are both pretty bad at generalizing but since this happens in ways that are sort of complementary, we can get a lot of leverage by combining them.

> resolved this well enough. Opinions?

Not at all resolved. If anything it is worse than before as we begin to understand it better, and now there are different versions of it that cover representation, relevance, epistemics. Pivoting "away" from logic just relocates it again. Arguably the whole challenge of neurosymbolics is (still) getting a persistent sidecar for logic bolted onto something like a language model. We actually have fairly decent autoformalizers (!!) and we still can't make that work very well in general.

From one perspective, the frame problem is pretty closely related to the "binding problem", causal reasoning and ramifications in general, and relevance is central to all. We have good pure formalisms for relevance, epistemics, and do-logic too. But we can't get language models to drive them very well, and language models alone are terrible at trying to do this sort of thing natively (see distractor sensitivity, mediated causality and multi-hop reasoning with implicit bridges).

Neurosymbolics probably is the key, but until there's more traffic between old-school and new we're facing the same old problems. When/if there's real progress.. I think we'd know. It may or may not be AGI-complete but the improvements for things like long-horizon and truly out-of-distribution planning would probably be immediate, obvious, and jaw dropping

> If you want to overcome this situation you need to replace capitalism globally.

Not true.. you just need to replace late-stage capitalism locally. There's a very reasonable concept of circular economy that's relevant here. It breaks down only in that sometimes you need things that you can't make and must go outside the circle. Especially if you intentionally work to mitigate the main stuff requiring you to go outside, it's not some law of nature that it must continue, or even a law of capitalism. Some protectionism / local-first is part of it, but the bigger part is just being rational.

The bad kind of globalism, enshittification, dead-economy theory, and basically ALL of the really ugly stuff we could talk about here are characteristic of late-stage capitalism and the associated short-term thinking, and it's totally consistent for a real capitalist to reject it. Why? Because getting as rich as possible isn't incompatible with sustainability.. particularly from the perspective of corporations/countries that plan to last longer than a single human CEO or exist beyond a single generation of shareholders/citizens.

> If we want to have a game where every weapon is a viable option, then we can’t bump RPS’s 3 to 4. We have to jump all the way to 5 choices with the game Rock – Paper – Scissors – Lizard – Spock. Each weapon beats two others and loses to two. The big downside of this game is complexity – I’m not a Trekkie and have no idea how Spock and Lizard are supposed to relate to the others. The upside is that ties are less common. For RPS 1/3 of all games will be a tie, while for RPSLS this drops to 1/5 – good if you want more decisive outcomes.

This is exactly like when Skippy is trying to work out the flavor profiles for the kinds of mush that monkeys will find most appealing. Yes, we know that "chocolate peanut butter bananas" is the true king of flavor latent space, but even a slight error in floating point precision could stick you with "cantor's peach butt nut abalone".

> something like Celery with all it's underlying infrastructure.

Unpopular opinion, but combining this with the other "no thanks" sentiments in this subthread is the right answer. Your app is so complicated you need async? Then it's complicated enough that you can benefit from infrastructure. I don't want to watch coworkers try to badly rebuild message queue or scheduling semantics in an application code base. Just use infrastructure that's made by people who know what they are doing. That was problematic in 2015, but in 2026 it's a bit of docker, and it's not just about web/microservices. Very easy for sufficiently complex apps to simply leverage a local sandbox of celery, redis, graphdb's and whatever. Stand-alone is overrated since we don't have to do it anymore.. app devs should get more comfortable working with ensembles like this so they have access to best-in-class solutions.

You don't like infrastructure AND have such a need for performance AND don't want threads or multiprocess? Consider using another language. Async is mostly a solution in search of a problem, and the enduring popularity of TFA goes to show this has been the right conclusion for ~10 years.

Why does someone go to the effort/expense of doing summaries without adding tags? The goal is less slop and not more, and this really just hurts usability after you've already spent money that could actually help? Isn't this the whole value proposition for literally everyone or am I nuts? I don't want a summary. I will switch immediately to the first alternative that allows tags/categories/filtering.