jonahx

18 ms·

271.
by jonahx 1y ago
> But tutoring isn't obviously superior to practice. Good tutoring will essentially be practice and worked problems with instant feedback -- not an individual "lecture". While there is value to being in the forest entirely
272.
by jonahx 1y ago
> but I'm unsure of its message My takeaway: Modern software stacks are usually cloud-dependent, and much bigger and more complex than they need to be, especially for offline, low-bandwidth, or low computing power use cases. Small,
273.
by jonahx 1y ago
This is the "confused deputy problem". [0] And capabilities [1] is the long-known, and sadly rarely implemented, solution. Using the trifecta framing, we can't take away the untrusted user input. The system then should not h
274.
by jonahx 1y ago
With a proctored in-person exam, we're talking about the difference between gaming the SAT, say, and gaming a take home English essay. And rates of cheating of "well under 1%" vs "well over 50%". Even if we allow fo
275.
by jonahx 1y ago
> Mind you, I don't really have any alternative suggestions. This is thing. If this choice is between: 1. A gameable system that will be gamed by most students. 2. An ungameable system that will unfairly punish those bad under press
276.
by jonahx 1y ago
It looks like a pug: https://i.imgur.com/omF3Abn.jpeg
277.
by jonahx 1y ago
> learn more about the Israeli/Palestinians > to be quite good at math and programming Since LLMs are essentially summarizing relevant content, this makes sense. In "objective" fields like math and CS, the vast majorit
278.
by jonahx 1y ago
Oh I understand the standard arguments. I just don't agree with them. > The point of separation of concerns isnt to keep the simple layer separate from the complex one. It's to simplify the whole thing by only addressing one c
279.
by jonahx 1y ago
> The idea that tests should be split into a specification layer and execution layer is a good one that should have taken off by now. There is a fundamental reason it hasn't: An actual specification layer isn't any simpler th
280.
by jonahx 1y ago
As a side question, do you think using tools like Lean will become a staple of these "deep reasoning" LLM flavors? It seems that LLMs excel (relative to other paradigms) in the kind of "loose" creative thinking humans do
281.
by jonahx 1y ago
Those videos are possibly my favorite thing on YouTube. I can't think of another use of my tax dollars that I get as much direct pleasure from.
282.
by jonahx 1y ago
Dead on again. What this and your other comment make clear is that once you start talking about the probability that X is true, especially in the context of hypothesis testing, you've moved (usually unwittingly) into a Bayesian framing
283.
by jonahx 1y ago
This is the key point.
284.
by jonahx 1y ago
Was the graphic design created from prompts too? It's surprisingly nice, especially considering you spent 45 minutes on it.
285.
by jonahx 1y ago
It looks similar to Obsidian. Are there significant advantages?
286.
by jonahx 1y ago
I have it! But I've always wondered if a more systematic approach based on Victor Allis's paper might be achievable. Can you play perfectly?
287.
by jonahx 1y ago
Beautiful. Unrelated, but what would be your advice/study plan for a human to reach near perfect game play in Connect 4, and how difficult would it be?
288.
by jonahx 1y ago
> The intuition that such a thing is "too powerful" is actually a pretty bad intuition here. I still disagree. Just focus on theorem proving and not the examples that are too simple. If the halting problem could be solved, w
289.
by jonahx 1y ago
This is a fair point, and I conceded it above in one of my first replies.
290.
by jonahx 1y ago
But it won't finish if there is no proof! A halting oracle will finish either way.
291.
by jonahx 1y ago
> for proofs of finite length in general, not just some fixed finite length. For a brute force proof finder, for your program to be guaranteed to finish in theory, you have to pick a length. So it is fixed. Ofc you can choose whatever
292.
by jonahx 1y ago
Yeah, I thought I already addressed that above. Can we do the same for a theorem prover? For proofs of some fixed finite length, I think the answer is yes, but without that constraint the answer is no. Whereas with a halting detector we co
293.
by jonahx 1y ago
"Free" in the sense of thought, cleverness, insight. You have an "everything" calculator. That is the sense in which it might be intuitive that it couldn't exist. > So the blog post's claims that a hypothe
294.
by jonahx 1y ago
What do you not buy? He's saying it could be used to do anything for free. He gave four examples. The first two don't have general solutions, and never will. You don't like the last 2 examples because we have algorithms f
295.
by jonahx 1y ago
I am talking about size of cognitive load, complexity, attack surface area, etc.
296.
by jonahx 1y ago
This is a good point, but there is a countervailing one: Sometimes the already invented, tested, and documented "standard" thing is much bigger than what you need. Perhaps you need one feature out of ten. It can make sense to re
297.
by jonahx 1y ago
This is very cool. The thing I really want is this, combined with some automated local background LLM training / rag (not sure what the right approach is) process. So that, at the end of the day, everything I bookmark get saved loca
298.
by jonahx 1y ago
What are you disagreeing with? Even if you do this (and it's good practice!), it is, empirically, not done in the vast majority of codebases. And even if you succeed with the utmost diligence, a vastly greater number of decisions (
299.
by jonahx 1y ago
So is the famous "Attention is all you need" wrong?
300.
by jonahx 1y ago
Love it. If you plan on adding to it, would be cool see (maybe via heatmap) where the births/deaths are happening.

More ›