Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A bear case: My predictions regarding AI progress

lesswrong.com

Post reply on HN
Print view View original

201–210 of 220 posts

Re: A bear case: My predictions regarding AI progress

#201

Post by arrakark »

hangonhn wrote: post #145
mark_l_watson wrote: post #106

I have used neural networks for engineering problems since the 1980s. I say this as context for my opinion: I cringe at most applications of LLMs that attempt mostly autonomous behavior, but I love using LLMs as ‘side kicks’ as I work. If I have a bug in my code, I will add a few printout statements where I think my misunderstanding of my code is, show an LLM my code and output, explain the error: I very often get us…

I'll add this in case it's helpful to anyone else: LLMs are really good at regex and undoing various encodings/escaping, especially nested ones. I would go so far to say that it's better than a human at the latter. I once spend over an hour trying to unescape JSON containing UTF8 values that's been escaped prior to being written to AWS's Cloudwatch Logs for MySQL audit logs. It was a horrific level of pain until I ju…

> LLMs are really good at regex

Maybe that's changed recently, but I have struggled to get all but the most basic regex working from GPT-4o-mini

Re: A bear case: My predictions regarding AI progress

#202

Post by arrakark »

gymbeaux wrote: post #27

Yeah agree 100%. LLMs are overrated. I describe them as the “Jack of all, master of none” of AI. LLMs are that jackass guy we all know who has to chime in to every topic like he knows everything, but in reality he’s a fraud with low self-esteem. I’ve known a guy since college who now has a PhD in something niche, supposedly pulls a $200k/yr salary. One of our first conversations (in college, circa 2014) was how he ha…

But original MC ran on JVM, which can run on ARM...

Re: A bear case: My predictions regarding AI progress

#203

Post by kragen »

usaar333 wrote: post #37
demaga wrote: post #33

Earlier quoted context omitted.

I would expect similar doom predictions in the era of nuclear weapon invention, but we've survived so far. Why do people assume AGI will be orders of magnitude more dangerous than what we already have?

More ability to kill everyone. That's harder to do with nukes. That said, the actual forecast odds on metaculus are pretty similar for nuclear and AI catastrophies: https://possibleworldstree.com/

Prediction markets should not be expected to provide useful results for existential risks, because there is no incentive for human players to bet on human extinction; if they happen to be right, they won't be able to collect their winnings, because they'll personally be too dead.

Re: A bear case: My predictions regarding AI progress

#204

Post by nyrikki »

zero_bias wrote: post #157
nyrikki wrote: post #150

Earlier quoted context omitted.

We can't even define what intelligence is. We know or have strong hints at the limits of math/computation related to LLMs + CoT Note how PARITY and MEDIAN is hard here: https://arxiv.org/abs/2502.02393 We also know HALT == open frame == symbol grounding == system identification problems. The definition of AGI is also not well defined, but given the following: > Strong AI, also called artificial general intelligence,…

Humans can’t solve NP-hard problems either, so definition of intelligence shouldn’t lie here, and these particular limits shouldn’t matter too

NP is interesting because it is about the cost of computation, and LLMs, are computation. A DTM can simulate a NTM, just not in poly time.

It is invoked because LLM+CoT requires a polynomial amount of scratch space to represent P, which is in NP.

I didn't suggest that it was a definition of Intelligence.

The Church–Turing thesis states that any algorithmic function can be computed by a Turing machine.

That includes a human with a piece of paper.

But NP is better though of the set of decision problems verifiable by a TM in polynomial time. Any TM or equivalently lambda calculus or algorithm can solve the Entscheidungsproblem, which was used by Turing to define Halt.

PAC Learning depends on set shattering, at some point it has to 'decide' if an input is a member of a set, no matter how complicated the parts are on top of that set, it is still a binary 'decison'

We know that is not how biological neurons work exclusively. They have many features like spike trains, spike retiming, dendritic compartmentalization etc...

Those are not compatible with the fundamental limits of computation we understand today.

HALT generalizes to Rice's theorm, which says all non-trivial symantic properties of programs are undecidable.

Once again, as NP is the set of decision problems verifiable by a DTM in poly time, that is why NP is important.

Unfortunately the above is also a barrier to formal definition of the class of AI-complete.

While it may not be sufficient to prove anything about the vague concept of intelligence, understanding the limits of computation is important.

We do know enough to say that the belief that AGI being obtainable without major discoveries is blind hope.

But that due to the generalization concept, which is a fundamental limit of computation.

Re: A bear case: My predictions regarding AI progress

#205

Post by disgruntledphd2 »

fragmede wrote: post #186
hattmall wrote: post #180

Earlier quoted context omitted.

I don't think I would use it if I were paying the real costs, and not a ~90% VC funded mark down.

yeah, but why does the fact that it's vc subsidized matter to you? the price is the price. I don't go to the store and look at eggs and lettuce and consider how much of my tax money goes into subsiding farmers before buying their products. maybe the prices will go up, maybe they'll go down due to competition. Thai doesn't stop me from using them though.

Because if they're not covering their costs now, then eventually they will which either means service degradation (cough ads cough) or price increases.

I applaud the GP for thinking about this before it becomes an issue.

Re: A bear case: My predictions regarding AI progress

#206

Post by aetherson »

lackoftactics wrote: post #99

>GPT-5 will be even less of an improvement on GPT-4.5 than GPT-4.5 was on GPT-4. The pattern will continue for GPT-5.5 and GPT-6, the ~1000x and 10000x models they may train by 2029 (if they still have the money by then). Subtle quality-of-life improvements and meaningless benchmark jumps, but nothing paradigm-shifting. It's easy to spot people who secretly hate LLMs and feel threatened by them these days. GPT-5 will…

What the author is referring to there as GPT-5, GPT-5.5, and GPT-6 are, respectively, "The models that have a pre-training size 10x greater than, 100x greater than, and 1,000x greater than GPT-4.5." He's aware that what OpenAI is going to actually brand as GPT-5 is the router model that will just choose between which other models to actually use, but regards that as a sign that OpenAI agrees that "the model that is 10x the pre-training size of GPT-4.5" won't be that impressive.

It's slightly confusing terminology, but in fairness there is no agreed upon name for the next three orders of magnitude size-ups of pretraining. In any case, it's not the case that the author is confused about what OpenAI intends to brand GPT-5.

Re: A bear case: My predictions regarding AI progress

#207

Post by fragmede »

disgruntledphd2 wrote: post #205
fragmede wrote: post #186

Earlier quoted context omitted.

yeah, but why does the fact that it's vc subsidized matter to you? the price is the price. I don't go to the store and look at eggs and lettuce and consider how much of my tax money goes into subsiding farmers before buying their products. maybe the prices will go up, maybe they'll go down due to competition. Thai doesn't stop me from using them though.

Because if they're not covering their costs now, then eventually they will which either means service degradation ( cough ads cough ) or price increases. I applaud the GP for thinking about this before it becomes an issue.

Which, as we know, is what killed YouTube and no one uses that anymore.

Re: A bear case: My predictions regarding AI progress

#208

Post by fragmede »

namaria wrote: post #199
fragmede wrote: post #134

Earlier quoted context omitted.

> it's just not there Build the much maligned Todo app with Aider and Claude for yourself. give it one sentence and have it spit out working, if imperfect code. iterate. add a graph for completion or something and watch it pick and find a library without you having to know the details of that library. fine, sure, it's just a Todo app, and it'll never work for a "real" codebase, whatever that means, but holy shit, jus…

How are you surprise that getting an LLM to spit out a clone of a very common starter project is evidence of it being able to generate non trivial and valuable code - as in not a clone of overabundant codebases - on demand?

because in actually doing the exercise, and not just talking about it, you'd come up with your own tweak on the Todo app that couldn't be directly be in the training data. you, as a smart human, could come up with a creative feature for your Todo app to have, that no one else would make, showing that these things can compose between the things in their training data and produce a unique combination that didn't exist before. copying example-todo.app to my-todo.app isn't what's impressive, having it able to add features that aren't in the example app is what is. If it only has a box of Lego and can only build things from them, and can't invent new Lego blocks, there's still a large amount of things it can be told to build. That it can assemble those blocks together into a new model that isn't in the instruction manual might not be the most surprising thing in the world, but when that's what most software development is, is the fact that it can't invent new blocks really going to hold it back that much?

Re: A bear case: My predictions regarding AI progress

#209

Post by nasmorn »

sjsdaiuasgdia wrote: post #118
rlnvlc wrote: post #88

Earlier quoted context omitted.

Were you not able to order food before Uber/Grab?

Before the proliferation of Uber Eats, Doordash, GrubHub, etc, most of the places I've lived had 2 choices for delivered food: pizza and Chinese. It has absolutely massively expanded the kinds of food I can get delivered living in a suburban bordering on rural area. It might be a different experience in cities where the population size made delivery reasonable for many restaurants to offer on their own.

Now if anyone solves the problem that for most cuisines ordered food is vastly inferior to freshly served meals. That would be revolutionary. Crisp fries and pizza. Noodles perfectly Al dente and risotto that has not started to thicken.

Re: A bear case: My predictions regarding AI progress

#210

Post by sjsdaiuasgdia »

nasmorn wrote: post #209
sjsdaiuasgdia wrote: post #118

Earlier quoted context omitted.

Before the proliferation of Uber Eats, Doordash, GrubHub, etc, most of the places I've lived had 2 choices for delivered food: pizza and Chinese. It has absolutely massively expanded the kinds of food I can get delivered living in a suburban bordering on rural area. It might be a different experience in cities where the population size made delivery reasonable for many restaurants to offer on their own.

Now if anyone solves the problem that for most cuisines ordered food is vastly inferior to freshly served meals. That would be revolutionary. Crisp fries and pizza. Noodles perfectly Al dente and risotto that has not started to thicken.

It's far from a perfect solution, but I applaud businesses that have tried to improve the situation through packaging changes. IHOP is a stand-out here, in my experience. Their packaging is very sturdy and isolates each component in its own space. I've occasionally been surprised at how hot the food is.

Hah. I went to find a picture and apparently they gave an award to the company that designed the packaging: https://www.directpackinc.com/2018-ihop-vendor-partner-year/

Post reply on HN