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

The changing goalposts of AGI and timelines

mlumiste.com

Post reply on HN
Print view View original

401–410 of 411 posts

Re: The changing goalposts of AGI and timelines

#401

Post by nomel »

daxfohl wrote: post #314
godelski wrote: post #301

Earlier quoted context omitted.

> there's no reason we'd need to have humans working jobs that only involve typing stuff into a computer and going to meetings all day I'm not sure I understand, and want to check. That really applies to a lot of jobs. That's all admins, accountants, programmers, probably includes lawyers, and probably includes all C-suite execs. It's harder for me to think of jobs that don't fit under this umbrella. I can think of s…

Actually it occurs to me that even if we did have AGI, or even if ASI, heck if ASI even moreso, we'd still need desk jobs to maintain the guardrails. Intelligence is one thing, being able to figure out how get a task done (say). But understanding that no, I don't want you to exploit a backdoor or blackmail my teammate or launch a warhead even though that might expedite the task. Or why some task is more important tha…

> we'd still need desk jobs to maintain the guardrails.

Simply put, an economy and society of managers. That's terrifying.

Re: The changing goalposts of AGI and timelines

#402

Post by froggit »

"if a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project"

The tried and true technique of speeding up completion of an established project by adding more people; well known for the tendency to fail successfully.

Re: The changing goalposts of AGI and timelines

#403

Post by marcus_holmes »

dragonwriter wrote: post #397
marcus_holmes wrote: post #390

Earlier quoted context omitted.

Good question. I don't know enough about neuroscience to really answer your question in depth. My opinion, uninformed as it is, is basically around the intuitive reasoning that something cannot be "highly autonomous" if it has to be kicked every second ;) Autonomous is defined as not needing to be controlled externally. And coupling that part with something as simple as cron job doesn't solve that in any meaningful w…

It seems to me that you are trying to define "autonomy" as a structural property rather than a behavioral one, and then adopting arbitrary rules as to what structures do not count as autonomous whether or not they produce the same behavior as structures which do. I guess that's fine, autonomy has lots of definitions (some in overlapping domains) and I guess one more doesn't hurt, but I'm pretty sure the intended use…

yeah, if we're just arguing semantics then I'm happy to let it go ;)

Re: The changing goalposts of AGI and timelines

#404

Post by godelski »

nomel wrote: post #400
godelski wrote: post #392

Earlier quoted context omitted.

> To me, a general intelligence is one that is not just specific Which is why a calculator is a great example. > A calculator is a very specific, very inflexible, type of intelligence, so it's not, by definition, general Depends what kind of calculator and what you mean. I think they are far more flexible than you give them credit for. > I'm in hardware engineering, so maybe I have a different perspective, but goals…

> And, I'm not talking about the indirect applications of a calculator or a specific intelligence. This was an attempt to prevent this exact chain of response. I calculator can only be used indirectly to solve a practical problem. A more general intelligence is required to know a calculator is needed, and how to break down the problem in a way that a calculator can be used effectively. For example, you can't solve an…

I can definitely write programs to solve real world problems. I think you're just so set on your answer that you're not recognizing the flexibility that exists. Your argument isn't so different from all the ones I hear that argue that science is useless and that it's engineers who do everything. It has the exact same form and smell albeit with different words. But as generalists we understand abstraction, right?

Re: The changing goalposts of AGI and timelines

#405

Post by nomel »

godelski wrote: post #404
nomel wrote: post #400

Earlier quoted context omitted.

> And, I'm not talking about the indirect applications of a calculator or a specific intelligence. This was an attempt to prevent this exact chain of response. I calculator can only be used indirectly to solve a practical problem. A more general intelligence is required to know a calculator is needed, and how to break down the problem in a way that a calculator can be used effectively. For example, you can't solve an…

I can definitely write programs to solve real world problems. I think you're just so set on your answer that you're not recognizing the flexibility that exists. Your argument isn't so different from all the ones I hear that argue that science is useless and that it's engineers who do everything. It has the exact same form and smell albeit with different words. But as generalists we understand abstraction, right?

yes you, a general intelligence, write a specific expert for a specific problem.

> If you want to argue that we don't need the concept of AGI, because something like specific experts could be enough to drastically change the economy, then sure! That would be true. But I think that's a slightly different, complimentary, conversation. Even then, say we have all these experts, then a system to intelligently dispatch problems to them...maybe that's a specific implementation of AGI that would work.

was to prevent this comment chain.

I think we'll have to give up at this point, with whatever smells you may be attempting to communicate with me. Cheers!

Re: The changing goalposts of AGI and timelines

#406

Post by nomel »

irishcoffee wrote: post #399
nomel wrote: post #386

Earlier quoted context omitted.

I think a charitable interpretation of their comment is that it hasn't stopped us from practical, fruitful, use. We didn't need an understanding of how it worked, or even a word for intelligence , let alone a definition of it, to get good practical results.

It’s a statistical gradient descent prediction machine. I never said it wasn’t useful. It absolutely isn’t smart.

What is "it" you're referring to here?

Nobody is talking about current LLM. And, we don't know if gradient decent will be involved, since the systems don't exist yet. Maybe it all be some runtime gradient decent. If there's an optimization problem involved in any part of the process, and numbers are involved, there's a near guarantee it will include gradient decent.

Re: The changing goalposts of AGI and timelines

#407

Post by fwip »

fwipsy wrote: post #396
fwip wrote: post #121

Earlier quoted context omitted.

> How many "I deleted the prod database" stories have you seen? Humans do this too. Humans generally do it on accident. They don't preface it with "Let me delete the production database," which LLMs do.

Hey, you're the guy who got the username I wanted. Hell of a coincidence

:)

Re: The changing goalposts of AGI and timelines

#408

Post by wise_blood »

DiscourseFan wrote: post #337
wise_blood wrote: post #330

Earlier quoted context omitted.

the ARC definition is the one I like the best, something like: "it is AGI when we can no longer come up with tasks easy for humans to solve but hard for computers"

We are very very far from that point

If we include tasks in the physical world, yes we are.

for abstract tasks (like ARC-AGI), who knows?

Re: The changing goalposts of AGI and timelines

#409

Post by Otterly99 »

pu_pe wrote: post #377
Otterly99 wrote: post #361

Earlier quoted context omitted.

But chess models aren't trained the same way LLMs are trained. If I am not mistaken, they are trained directly from chess moves using pure reinforcement learning, and it's definitely not trivial as for instance AlphaZero took 64 TPUs to train.

I meant trivial in the sense it's a solved problem, I'm sure it still costs a non-negligible amount of money to train it. See for example the chess transformer built by DeepMind a couple of years ago which I referred to in a sibling comment [1]. [1] https://arxiv.org/abs/2402.04494

Thank you for the link.

I admit, my knowledge of reinforcement learning is a bit outdated so it seemed to me that it was unattainable for a non-specialized model to train efficiently on something like chess, which has a huge state space.

Re: The changing goalposts of AGI and timelines

#410

Post by DiscourseFan »

wise_blood wrote: post #408
DiscourseFan wrote: post #337

Earlier quoted context omitted.

We are very very far from that point

If we include tasks in the physical world, yes we are. for abstract tasks (like ARC-AGI), who knows?

We are far from that point for abstract tasks
Post reply on HN