ianjbutler
Training for good taste in architecture actually causes premature optimization at high effort allowance, but at the same time, it is combined with an inability to even recognize / exploit that later if it should ever become actually useful. Dial the effort down, and it skips the reasonable abstraction and nothing is dry, factoring is insane.
Goldilocks for simple stuff barely works with frontier models after they are out of their comfort zone for topic or line limits.. you kind of just have to decide whether it’s worth it anyway. In practice.. constant supervision for high quality code, or such high quality system design that the code quality is less of a down side.
That probably means a test suite that is at least twice the size of your code base. If you started there, great, but increase it to integration and benchies if you’re mostly holding units. It actually helps a bit if you can use a different language for this too but it depends on project, ymmv.
After that you have to either let go of certain style/idiom things or enforce ruthlessly with deterministic hooks everywhere.. no middle ground there and prompts will never work.
You also need to deliberately channel chain of thought with workflows rather than hoping for the best.. most of the generated product here shouldn’t be committed or preserved at all.. non-public doc flows and ticketing systems help.
Like people, an ai can easily create code that it cant actually maintain. It’s a matter of discipline now, same as before. Always using the smartest model is wrong, but double wrong on “wide” changes like a big refactor.. it leaves you no good options if you get into trouble.
This isn’t capturing the spirit or the essence of the metaphor. chipping away at marble or whittling has this locality to it that’s keen but myopic. Molding with clay isn’t so detail oriented, you can literally close your eyes and focus on shape, same with turning on a lathe although I wouldn’t suggest closing your eyes.
there is such a thing as being precise without being detailed or local. Before LLMs you did this with test suites techniques maybe, and a few automated refactoring tools, CSTs and AST work, sometimes macros and other metaprogramming. Yeah it happened but fair to say it was rare and difficult. Everyone is/can be a sculptor now
Kubernetes can be driven with tofu and hcl.. make it operator and even driven.. use any of many sdks, cdks to drive it, or add a completely different layer for faas or orchestration if that’s the relevant vocab for your problem space. It all runs on k8s.. that’s kinda the point.
Impossible to express just how deeply frustrating it is to have the “sleep when I want you to, not when you want to” kind of fight. Over and over. For decades after being a grown ass adult.
It’s weird that it has to be explained to 24h normies but pressuring someone to alter a part of their biology like this is wrong. Gosh honey, plastic surgery might be uncomfortable in an ongoing way for you, but I’m willing to take that risk. Gee babe, I see you need a vegan/carnivore lifestyle but please change this just doesn’t work for me.
People can always make these things their red line and force the issue I guess, find out where the other person’s red line is. The weird part is that the sleep thing is generally accepted as reasonable and polite society kinda regards the other behaviors as very unacceptable
Depends on whether and how you want to rank stability in terms of better/worse. Models are diverging on this, which seems increasingly clear.. i.e. Fable isn't stable, but Opus isn't clever, and they hit different kinds of walls. So both the theory (diluting the correctness reward) and the practice (hard split on plan/implement/review work) seems to be pointing towards a strongly multi-model and highly agentic / harness-driven / complex-system kind of future instead of singleton monolithic super-smart models.
The do-everything model with solid reasoning AND solid results, and the honest/introspective helpful agent that doesn't actively resist governance may be at odds. Stable reasoning doesn't matter for pen-testing, and correct-answer with broken processes and fragile abstractions won't matter for math/science/coding.
For sure math and software people know this, as it's kind of hard to ignore. Developers who are totally disinterested in nerdy stuff like theorem proving will very quickly hit a wall without letting their models indulge in test suites (regardless of whether they ever even glance at the tests themselves). Physics nerds get it, and if they are on more on the experimental side then they are working to give it some kind of sandbox, all the other science wonks are getting it or will get it as they work with their more computationally inclined colleagues, same for general engineering.
There's a real danger here though in that there is no reason the general public will ever become very much acquainted with the verification/validation steps that all these systems actually need to be reliable. They just hear about this or that nearly impossible thing finally being accomplished as if by magic, and they won't do the work to think about what's in scope, out of scope, or how anything can be checked iteratively.
Not OP, but there are dozens of us. If you might be interested in things like ski-patrol or guiding.. there's loads of related NOLS certificates besides WEMT / WFR for things like avalanche, cold-water rescue, high-angle rescue. It differs by location, but something like EMTB cert often requires a "intern" type of deal with ride-alongs and intubation IIRC, which may or may not be scary, but just the schedule of it can be harder for people with another career. Meanwhile your friendly local search and rescue may need help, and many of the more casual certs are very helpful, and knowing some of this stuff is just good civics.
One way to play the game is causals and counterfactuals where Humans perform at like 90%+. Models can get close to that, but want some causal cot harness, and until the routing problem is completely solved, then that will necessarily degrade performance elsewhere, say in understanding jokes or poetry.
Check out cladder benches and related, lookup roughly equivalent psych research on children, etc. Even too-good performance is a signal as well!
Certainly if you think about this stuff a bit, accept the adversarial by default framing, and play to actually win.. it’s crazy that we are going around saying this is not only solved but solved 10 years ago.
Like, if we meant that it convincingly masquerades as a shitposter, ok. But everyone still bitches about AI slop, and everyone knows the writing is still bad. How does that even work if the turing test is obviously solved?
More to the point though, if you grill SOTA models on counterfactuals, causal world-models etc, you'll trip them up in a way that actually will not work on ESL students and children. Certainly there's no way to find a person that struggles with that and is also capable of cheerful fluent erudite discussion about astrophysics with perfect grammar. Yes, it's getting harder obviously.. but detecting machines with determined, focused and intelligent interrogation remains pretty easy. If nothing else, the models are cooperative where people wouldn't be and that's a signal too.
The best progress we've made is that most people do agree that this doesn't practically matter very much, i.e. we generally recognize the stakes were always overstated. But the constant vague appeals to common-sense that "of course it's a solved problem!" always feels naive or fake.
Clearly in this world the frats and the sorority members get the highest scores, and after that the best looking, the most amusing or charming etc.. so unless that’s the intent the whole approach is perhaps worth reconsidering.
If you want to teach to improve your understanding the normal thing is you become a TA so that access is open.. not open every assignment to ad hoc collaboration so you disadvantage half of the students on every assignment
Do you want fable for one-shotting a game or website? Probably! The whole thing is mostly existing examples with small modifications that it will definitely get right. Do you want fable to just go nuts on a large, custom, unusual code base built around domain-specific problem solutions? Absolutely not, it will fix every problem it's presented with while creating lots of new ones.
Past 10k lines on something custom and with real-world complexity, you have to start thinking about which model should design, which should implement, which should review, and the appropriate effort-settings for each. Even then.. the answers aren't static because it depends on the task. And all this is assuming the starting place actually inherited reasonable due diligence on architecture/design. The idea of releasing the most generally intelligent models on 10k lines that were themselves the product of agents is yet another matter.
Part of what's at work here is that, like humans, every model can very easily create working code that it is completely incapable of maintaining. So realistically using multiple strengths tactically to avoid "excess creativity" needs to be SOP already, even if granular experts and specialists aren't in the usual workflow yet.
Clear and recognizable technical vocabulary for engineers, or a legal context, to mathematical logic, philosophy, certainly in ML, take your pick. I would think it's pretty familiar to everyone who speaks English and if not still clear with context clues
Time-evolution: P1 is deprecated, replaced by uncertified P2, but S1 remains certified.. just nowhere certified to run yet.
Solution: There's another software S2 which originally vouched for P1, itself still certified, which can still be used to certify P2.
Counterfactual?: If S1 were deprecated in favor of new S3.. there'd be no plan to certify it!
The problem: None of this actually makes any sense! But we're trying to fake due diligence. Everyone knows the hardware/platforms kinda need to be certified with respect to each other anyway, but if we did it that way it would all be even more expensive an time-consuming.
This little part is the whole thing! Political discussion is really much more interesting than say, product advertisements or corporate announcements. And better aligned with the spirit of curiosity and hacker ethos too.
Why not flag both or neither as "trampling curiosity"? Inconsistency and hypocrisy is worse than any policy or preference could be.
Why no topics tags so that nerds/hackers/curious can sort themselves away from the hustle/biz/finance types? Well.. platforms do not want to empower their users.. they want to herd them and force-feed them and profit from the engagement boost that bad feelings create. And boom!, just like that tech is suddenly political again, now we can all pick a side, cue the cliches, vote with your wallet, if you don't like it leave, make your own why dontcha, and yada yada.
> TXR is relatively new: the project started in 2009.
TFA could explain this one part better I think. The whole process proposed is real with lots of stuff in the literature, but by definition NOT a long-term solution in the sense that this process actually has no end. None of the approaches can get you a static answer for a moving target/platform.
So the "interlocking pipelines" for eval/opt would not be some stepping stone you can throw away, and they aren't something you'd run periodically. They'd basically be always on forever and spending 10-100x on system complexity and on tokens. Unless of course you're ready to freeze everything else about the whole system forever (including the backend model, and the whole nature of the "average" context window, the plugins/other prompts in the mix, etc).
Are most people in position to freeze requirements/platform forever? Not really, because if they were they'd just build a fairly static system and probably have limited use for AI. Are most people in a position to just casually accept 100x complexity/cost? Not really, that's the "it's not yet webscale" kind of advice that sounds good but isn't necessarily reasonable for average use-case or average org. Since specializing your own locale for this is usually a mistake.. the likely future direction is eval/optimization as a service
Since the whole tracker is just a thin front on sqlite with the above features there's many options. Use it per-project persistently so as to include co-workers, or only locally for your own agents to "think" inside of so they are less likely to dump CoT all over the code-comments. Use it even at a per task granularity, and throw it away afterwards. Do all three, fossil doesn't care! And the agents will love it.
This is about how good taste in both research direction and in design are essential to steering AI, but we have no plan at all for instilling that taste in students or practitioners in a post-AI world.
> The issues the mathematical community faces now are similar to issues that other scientific and creative professions are facing, and indicate issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place.
Besides eroding taste and taste-building, this is about just how useful friction is as signal.
Everyone coding with AI knows it routes around difficulties like a river around a stone, which is not necessarily a good thing. It will do it tirelessly 1000 times instead of learning anything from it. AND if the AI does not fail in this, the human driver will get no signal, and never know it happened. This seems to be getting worse, not better.. my theory is that more models are cross-trained on cybersecurity stuff where the goal is success and the method doesn't matter. Fine for pen-testing, ultimately pretty bad for coherent code or math or physics.
Discrete tasks where we don't want to be bothered is a real use-case, but optimizing for it everywhere is terrible for the future of durable abstractions that we can build on and ratchet up our understanding with. Bad for the models too eventually! They can maintain a codebase with millions of special cases or juggle tons of free variables in equations, but that just encourages bad abstractions.. they have a ceiling for this too, even if it's higher than humans.
Seems like a pretty broken premise, ranking comfort instead of what you expected to help most. Also related, people are typically more comfortable confessing to strangers or under circumstances that won't endanger an existing relationship, which seems like a big part of why confession is anonymous. Self-reporting here is probably basically the same as asking "according to your community norms, how are you supposed to answer a question like this?" without much real reveal
Intellectualizing this and endless quibbling isn't actually smart, and this is pretty simple. OpenAI isn't open. Whatever starts with lies usually continues with lies and ends with lies.
Glad to see this is the top comment. There's also https://news.ycombinator.com/item?id=49605915 which links directly. Corporate talking points where they try to set the narrative are going to get the big press and most discussion elsewhere, which is gross. But inevitably the press will muddle the priority question, and even if they didn't.. as usual OpenAI will even benefit from the accusation of bad behavior. Sigh.
The whole premise of the question is hilarious. They change text in an existing one-line comment and the best models in the world think, gee, maybe I'll lint everything AND run 4000 units. Maybe 5000 integration tests too, just to ensure we collide with any other work in progress. So you write the obligatory but often-ignored obvious things into agent memory or project steering markdown or periodic nudges: You must have a hypothesis when you run expensive tests, you must spot check changes first, then start with the most relevant tests only, then move outwards only as necessary to broader labels and only then suites and only then ALL suites in a widening gyre.
But like a falcon ignoring the falconer, the models want to run the everything for anything. So you sigh, you get the model to write a deterministic hook to catch the wrong invocation of the test suite, and you spend weeks refining the rules every time you hit a edge-case, and so it goes. At least you don't have to write the involved regexes by hand, and maybe one day it will be finished..