I would like to think that pg_hint_plan is designed in such a way that any hint it accepts must be a valid plan for the query. I’m quite confident that schemes with this property that can also express high quality plans are possible and not even excessively complicated.
This is not to say that it’s possible to genetically verify that a proposed algorithm does what you want it to — that would be undecidable or NP-hard or co-NP-hard depending on how you formulate the question.
I wouldn't be very excited about adding a 4B param model to my database deployment, but using this kind of approach while testing an app to identify query plans where Postgres is leaving performance on the table seems valuable without much risk.
You're misunderstanding the setup here. The LLM doesn't modify the query, just some details about how to choose between different ways to break the query into basic operations on the tables. The SQL doesn't change. It's still up to postgres to guarantee that the results match the query. If the proposed plan were nonsense that didn't amount to carrying out the query, postgres would ignore it.
Because P!=NP (very probably IMHO) there's a huge class of problem for which LLMs are useful on the expensive and heuristic-y generation side because the verification is relatively inexpensive.
You could set the CSS zoom property to 1/window.devicePixelRatio with JavaScript but I think you can't do it with CSS only. If you do that then px is physical px. This doesn't include browser zoom though or mobile pinch to zoom. You could also get those with JavaScript but then people would be really confused by the behavior of your website.
You definitely can't get pure integers though it's always a float even if you use whole numbers. It does eventually get rounded before becoming pixels though.
That is sort of where I ended up, Not that it matters, Pixels should probably never actually be used. It just offends me that a unit that is sort of messy and tied to physical hardware but does provide measurement that can not be found elsewhere. gets neutered and we end up with another useless absolute unit, inch, point, millimeter, pica, the Q(did we really need the Q?) and now the pixel.
I think there were some who really cared but I also think a lot of the 4chan types that found him egged him on and tried to trigger his schizophrenia because they treated it like it was all a big joke.
Absolutely. A lot of the 4chan types thought he was "based", less due to any of his actual talents, and more due to his outbursts and language choices (e.g. his "glow-in-the-dark" comments). It was a spectacle more than anything. He was just a lolcow to them.
The entire point of writing proofs is for advancing human understanding. A giant dump of symbols that passes the lean compiler is meaningless besides human beings understanding it.
> The entire point of writing proofs is for advancing human understanding.
Proofs also enable AIs to direct search and generate knowledge. Verifiability is immensely useful for keeping AI grounded.
One might imagine AI generating enormous numbers of hypotheses and then trying to prove or disprove them, and then mine that data for new abstractions and heuristics.
But what does it mean? The theorems are just symbols in lean. The conjectures humans chose are carefully selected to be the questions that are interesting and relevant to our intuition about the real world.
Math often doesn't have applications for hundreds of years and that application is only possible because people deeply understand it and how it applies to the real world.
Generating an endless list of true statements doesn't really do anything, those things are already true regardless of whether someone has written a lean program to model them.
Lean terms and programs have defined meanings, just like anything mathematical does.
It sounds like you're asking something nebulous, like does it have a soul.
Mass generation of conjectures and proofs/disproofs could AI to discover objectively mathematically useful things. For example, it might discover shortcuts, lemmas, even abstractions that are useful in the proofs of these things -- and judge that utility by how much they improve the ability of the AI to prove things in this mass of problems. It wouldn't say whether the things are useful for non-mathematical human problems, but then human mathematicians, as you say, can't really judge that either.
Sometimes the purpose of the proof is simply to demonstrate that some construct is a safe assumption for other more interesting work-- and could still serve that purpose even if it was entirely a black box.
Is it the AI's fault we can't understand? If the GUT is beyond human comprehension does it matter less? We don't apply this reasoning to other animals or even to less capable humans. Besides, the robots may want to ponder maths for their pleasure.
Qwen 27B is runnable for that price and it can produce patches that work using opencode as a harness. It's obviously not as capable as frontier models but it can do things if you provide the judgement.
It seems that the "open access" is more marketing than something reality based they actually want to do.
https://dl.acm.org/openaccess ---> So how to access the content? Do I have to register or what? It the "open access" only for academic org's people or for everyone in the world?
https://dl.acm.org/ ---> Okay, nice simple search field without loggin in, but when you try to search something, you get thousands of results, even if you search specific author and the exact name of paper, you will get hundreds of results and the thing you want is buried somewhere on page 247. Filters of authors, years etc. are for Premium subscription. But just googling the thing finds the link to ACM... And want to get the actual PDF? Hope it says "free access"...
Google scholar has become the way to search papers (which is somewhat worrisome). What people want from there is a download for all the stuff that does not list an author copy. Open access IMHO is just a reaction to the fact that mostly you would not need a subscription anyhow. Now the authors are paying upfront or universities are paying flat for all their researchers. The problem is now the incentives are not increasing the number of subscriptions but increasing the number of papers published.
I think the difference is in math the problem is fully specified and easily verifiable and in programming it's not. I don't agree that we always know we can solve the problem.
For example, create a DFA for a regex, not too bad just use Thompson's algorithm and then NFA->DFA. But now we have to care about efficiency, user API, maintainability of definitions etc.
Yes exactly. The author’s design decisions only make sense if this is supposed to be a toy language. On using pthreads rather than fibers:
> I decided not to use one. I wanted something dead simple — an approach I could explain in a paragraph, using tools every C programmer already knows. The trade-off is that you lose some performance with fine-grained blocking, but in many real-world situations, pthreads work fine if you use a worker pool.
Sure. You can take a large production Go app and measure how many user space threads are launched; it’s decidedly a lot more than the typical number of threads if you were using pthreads.
And the author didn’t really justify why select isn’t implemented other than implementation difficulty.
reply