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

AGI is an engineering problem, not a model training problem

vincirufus.com

Post reply on HN
Print view View original

391–400 of 442 posts

Re: AGI is an engineering problem, not a model training problem

#391

Post by galangalalgol »

root_axis wrote: post #387
jjaksic wrote: post #335

Earlier quoted context omitted.

Illusion doesn't imply it's unnecessary. Humans (and animals) had a much higher probability of survival as individuals and as species if their experiences felt more "real and personal".

If it has a functional purpose then it's not an illusion.

That is in interesting viewpoint. Firstly, evolution on long time scales hits plenty of local minima. But also, it gets semantic in that illusions or delusions can be beneficial, and in that way aid reproduction. In this specific case, the idea is that the shortcut of using the model of models as self saves a pointer indirection every time we use it. Meditation practices that lead to "ego death" seem to work by drawing attention to the process of updating that model so that it is aware of the update. Which breaks the shortcut, like thinking too much about other autonomous processes such as blinking or breathing.

Re: AGI is an engineering problem, not a model training problem

#392

Post by cyanydeez »

andy99 wrote: post #19

If you believe the bitter lesson, all the handwavy "engineering" is better done with more data. Someone likely would have written the same thing as this 8 years ago about what it would take to get current LLM performance. So I don't buy the engineering angle, I also don't think LLMs will scale up to AGI as imagined by Asimov or any of the usual sci-fi tropes. There is something more fundamental missing, as in missing…

It seems similar to the fermi paradox.

The underlying assumption is that it exists in the first place. Or rather, one must first accept an axiom.

In fermi, its that interstellar signals can be detected and further travel is possible.

In AGI, its that intelligence is a isolateable process which we can bootstrap in minimal time.

Both assumes human progress are templates of unlimited exponential growth.

Re: AGI is an engineering problem, not a model training problem

#393

Post by adastra22 »

MadnessASAP wrote: post #347
adastra22 wrote: post #287

Earlier quoted context omitted.

LLMs have loops. The output is fed back in for the next prediction cycle. How is that not the same thing?

Wish I had a great answer for you but I don't. It certainly allows for more thought-like LLMs with the reasoning type models. I guess the best answer is that the loop only happens at a single discrete place and doesn't carry any of the internal layer context across. Another answer might be, how many comments did you read today and not reply too? Did you write a comment by putting down a word and then deciding what th…

It feels like the same thing to me…

Re: AGI is an engineering problem, not a model training problem

#394

Post by Timwi »

missingrib wrote: post #139
AIPedant wrote: post #56

Earlier quoted context omitted.

It is vacuously true that a Turing machine can implement human intelligence: simply solve the Schrödinger equation for every atom in the human body and local environment. Obviously this is cost-prohibitive and we don’t have even 0.1% of the data required to make the simulation. Maybe we could simulate every single neuron instead, but again it’ll take many decades to gather the data in living human brains, and it woul…

That is only true if consciousness is physical and the result of some physics going on in the human brain. We have no idea if that's true.

Whatever it is that gives rise to consciousness is, by definition, physics. It might not be known physics, but even if it isn't known yet, it's within the purview of physics to find out. If you're going to claim that it could be something that fundamentally can't be found out, then you're admitting to thinking in terms of magic/superstition.

Re: AGI is an engineering problem, not a model training problem

#395

Post by HarHarVeryFunny »

energy123 wrote: post #126
andy99 wrote: post #19

If you believe the bitter lesson, all the handwavy "engineering" is better done with more data. Someone likely would have written the same thing as this 8 years ago about what it would take to get current LLM performance. So I don't buy the engineering angle, I also don't think LLMs will scale up to AGI as imagined by Asimov or any of the usual sci-fi tropes. There is something more fundamental missing, as in missing…

What will it scale up to if not AGI? OpenAI has a synthetic data flywheel. What are the asymptotics of this flywheel assuming no qualitative additional breakthrough?

Did GPT-1 scale up to be a database ? No - it scaled up to be GPT-2

Did GPT-2 scale up to be an expert system ? No - it scaled up to be GPT-3

..

Did GPT-4 scale up to become AGI ? No - it scaled up to be GPT-5

Moreover, the differences between each new version are becoming increasingly less. We're reaching an asymptote because the more data you've trained on, natural or synthetic, the less is the impact of any incremental additions.

If you scale up an LLM big enough, then essentially what you'll get is GPT-5.

Re: AGI is an engineering problem, not a model training problem

#396

Post by adastra22 »

amanaplanacanal wrote: post #349
adastra22 wrote: post #291

Earlier quoted context omitted.

No you are not the only one. I am continuously mystified by the discussion surrounding this. Clause is absolutely and unquestionably an artificial general intelligence. But what people mean by “AGI” is a constantly shifting, never defined goalpost moving at sonic speed.

What we envisioned with AGI is something like self directed learning, I think. Not just a better search engine.

Isn’t that unsupervised learning during training or fine-tuning?

Re: AGI is an engineering problem, not a model training problem

#397

Post by Uehreka »

fao_ wrote: post #357
root_axis wrote: post #258

Earlier quoted context omitted.

It doesn't seem clear that there is necessarily any connection between consciousness and intelligence. If anything, LLMs are evidence of the opposite. It also isn't clear what the functional purpose of consciousness would be in a machine learning model of any kind. Either way, it's clear it hasn't been an impediment to the advancement of machine learning systems.

> It doesn't seem clear that there is necessarily any connection between consciousness and intelligence. If anything, LLMs are evidence of the opposite. This implies that LLMs are intelligent, and yet even the most advanced models are unable to solve very simple riddles that take humans only a few seconds, and are completely unable to reason around basic concepts that 3 year olds are able to. Many of them regurgitate…

I suspect that you haven’t really used them much, or at least in a while. You’re spouting a lot of 2023-era talking points.

Re: AGI is an engineering problem, not a model training problem

#398

Post by sixo »

adastra22 wrote: post #290
sixo wrote: post #11

Earlier quoted context omitted.

On the contrary, we have one working example of general intelligence (humans) and zero of quantum computing.

There are many working quantum computers…

ah, I mean, working in the sense of OP: that a system which overcomes the "engineering hurdles" is actually feasible and will be successful.

To be blocked merely by "engineering hurdles" puts QC in approximately the same place as fusion.

Re: AGI is an engineering problem, not a model training problem

#399

Post by danenania »

HarHarVeryFunny wrote: post #389
danenania wrote: post #171

Earlier quoted context omitted.

I’m not sure that self-updating weights is really analogous to “continuous learning” as humans do it. A memory data structure that the model can search efficiently might be a lot closer. Self-updating weights could be more like epigenetics.

There's a difference between memory and learning. Would you rather your illness was diagnosed by a doctor or by a plumber with access to a stack of medical books ? Learning is about assimilating lots of different sources of information, reconciling the differences, trying things out for yourself, learning from your mistakes, being curious about your knowledge gaps and contradictions, and ultimately learning to correc…

I think my point is that when the doctor diagnoses you, she often doesn’t do so immediately. She is spending time thinking it through, and as part of that process is retrieving various pieces of relevant information from her memory (both long term and short term).

I think this may be closer to an agentic, iterative search (ala claude code) than direct inference using continuously updated weights. If it was the latter, there would be no process of thinking it through or trying to recall relevant details, past cases, papers she read years ago, and so on; the diagnosis would just pop out instantaneously.

Re: AGI is an engineering problem, not a model training problem

#400

Post by signa11 »

jibal wrote: post #174
signa11 wrote: post #83

won’t somebody please think about Mr. Godel, and the Incompleteness Theorem ?

They aren't relevant. Even if Penrose and Lucas were right (they aren't), a computational system can solve the vast majority of the problems we would want solved.

seems like a bait-n-switch argument here.

we are talking explicitly about a.g.i here, not debating if the computer can solve a majority of problems or not.

the two things can be true at the same time.

Post reply on HN