Last poster before threadlock gets a cookie (cookie thread (Part 7)) (Part 11)

once the dot com bubble collapsed all those slop websites just stopped existing and it became no longer possible to make any more. it’s true

7 Likes

AI-generated press releases will remain but they’ll probably be far less “WELCOME TO THE FUTURE OF INTELLIGENCE !”

1 Like

its going to be nice. if we ignore the massive economic collapse that, of course, the working class will eat the brunt of

2 Likes

my personal prediction is a fair bit higher than that (the mathematicians I’ve seen discussing it online seem to think it checks out, + (less confidently) if this were solely just rushed out the door to boost their IPO I expect they would have dropped it just before the IPO rather than just after, call it like 95/5 in favor of the core proof being legit (+ much higher likelihood of smaller issues that don’t undermine the actual proof) (I thought about writing 90/10 since that Feels less Overconfident but I don’t think I sincerely believe that there’s a 10% chance it’s wrong)), but I sincerely appreciate you being willing to commit to a prediction here

3 Likes

sshhh i didn’t say there’d be none

1 Like

just Less. like let’s not pretend that AI-generated youtube shorts are this super long-term content that’s going to keep coming back in the algorithm after a hypothetical massive reduction in public accessibility of AI tools

2 Likes

look i may be way less of a hardline anti-AI person than I used to be but I’m still extremely skeptical of basically everything these companies say

4 Likes

I would proooobably personally bet that AI-generated YouTube shorts are unfortunately here to stay

4 Likes

all the math people basically already thought riemann was true. idk nearly enough about the actual techniques but my basic reaction is “checks out”

1 Like

the surprising part is that it actually got proven but it was basically suggested to me with pretty high confidence when ai was finding counterexamples left and right but it didn’t find one for riemann

1 Like

i’ve been using it at work and my experience has been that you cannot do anything that’s too broad in scope because it will always do weird hack job nonsense. sometimes it’s like “yeah okay sure i guess i could see why someone would do that” and other times it made an odd decision because i found out it was effectively covering up a problem that required more nuanced decision making. like “oh you did this famously bad programming thing because doing it this other way created a narrow case that made something ambiguous and the solutions to making it less ambiguous required decision making outside of the problem scope.” it’s not the path of least resistance at all because it’ll constantly overengineer things in confounding ways that make no sense for the scope of the problem at hand (it seems to be tuned to make assumptions that work if you’re doing Big Tech numbers). i would say it’s probably more accurate to say that its output is optimized to be sycophantic and appeal to very specific kinds of people, and seem to be correct. it’s still very clear to me that is still extremely far away from being incapable of anything approaching actual thought.

they also rapidly degrade in quality as the number of tokens increases. context tends to compact about 1 million tokens for all frontier models still and both my coworkers have mentioned to me that if you get to a second context compaction you just need to stop what you’re doing and open a new chat (i’m not sure i’ve ever had it compact even once, but that’s probably influenced by me just taking that to heart).

but this probably wasn’t a “single agent” working alongside a mathematician and was a swarm of them effectively brute-forcing chunks of it. exactly how the solution was arrived at would affect the odds i would give it of being legit. the more hands off they were and the more it was similar to a brute-force swam, the higher odds i’d give it of being something like a subtle Lean exploit than an actual solution. it also depends on how many iterations of verification were being done because i have deadass told Astra to essentially check its own code and its found bugs in it, and then i did it again on its “fixed” draft and it then again found more bugs. then again maybe i just happen to work with all the moron models with the stupid dial turned to max. can’t really say

4 Likes

oh yeah no im in touch enough with mathematics to say that like. nobody is surprsied by the notion that Riemann is true, the problem is and ahs always been proving it

2 Likes

tbc my odds were intended to be on this proof being correct, not on the Riemann hypothesis as a whole

4 Likes

i am honestly so glad that i “work” in a field (creative writing) where AI is threatening but not actually, like, remotely good at its’ job because I genuinely cannot stand the idea of trying to corral the temperamental and Bad models that are actually accessible to most people and businesses into doing something useful

3 Likes

lets not forget how much more heavy-duty the models that researchers directly connected to the industry have access to compared to the stuff they’re selling to businesses and consumers

4 Likes

the metaphor i was given and one i’ve found has been true is using an LLM to code is like supervising an idiot that can type 5,000 words per minute. and even then i don’t actually think it saves time overall because it’s constantly doing something that needs to be corrected or ignored or thrown out, and it takes time to read and parse what it spit out. it’s useful when i feel stuck because it making bad decisions about what to do is at least a decision, and it’s pretty easy to recognize the bad parts of something and the process of fixing it usually helps me arrive at something useful. and usually about 2 working days later i throw out what it wrote because it becomes more apparent all the ways it wasn’t good, but again it did Make A Decision

3 Likes

the models the cutting edge research mathematicians are working with might be slightly better than the ones you use, I think

plus if Ai is good at one thing it is totally mathematics

2 Likes

i understood that, to be clear! and I do think it’ll be a monumental achievement no matter how we get to it - though I suspect the proof itself won’t be useful until us humans digest it and understand it and can direct the machines through the next bit of research. it’s not like high-level mathematics was this Serene and Computer-Free field before AI came along, we’ll be fine.

1 Like

i did understand that, to be clear

1 Like

We have invented a new way to Prove Everything and it is totally lame

3 Likes