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

191–200 of 442 posts

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

#191

Post by aorloff »

tshaddox wrote: post #58
aorloff wrote: post #32

Earlier quoted context omitted.

A question which will be trivial to answer once you properly define what you mean by "brain" Presumably "brains" do not do many of the things that you will measure AGI by, and your brain is having trouble understanding the idea that "brain" is not well understood by brains. Does it make it any easier if we simplify the problem to: what is the human doing that makes (him) intelligent ? If you know your historical cont…

> Does it make it any easier if we simplify the problem to: what is the human doing that makes (him) intelligent ? Sure, it doesn’t have to be literally just the brain, but my point is you’d need very new physics to answer the question “how does a biological human have general intelligence?”

Suppose dogs invent their own idea of intelligence but they say only dogs have it.

Do we think new physics would be required to validate dog intelligence ?

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

#193

Post by danenania »

taormina wrote: post #136
dcre wrote: post #40

The first premise of the argument is that LLMs are plateauing in capability and this is obvious from using them. It is not obvious to me.

Just ancedata, but they keep releasing new versions and it keeps not being better. What would you describe this as if not plateauing? Worsening?

I see a lot of people saying things like this, and I’m not really sure which planet you all are living on. I use LLMs nearly every day, and they clearly keep getting better.

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

#194

Post by cakealert »

There is a reason why LLM's are architected the way they are and why thinking is bolted on.

The architecture has to allow for gradient descent to be a viable training strategy, this means no branching (routing is bolted on).

And the training data has to exist, you can't find millions of pages depicting every thought a person went through before writing something. And such data can't exist because most thoughts aren't even language.

Reinforcement learning may seem like the answer here: bruteforce thinking to happen. But it's grossly sample-inefficient with gradient descent and therefore only used for finetuning.

LLM's are regressive models and the configuration that was chosen where every token can only look back allows for very sample-efficient training (one sentence can be dozens of samples).

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

#195

Post by tomrod »

galangalalgol wrote: post #173
Uehreka wrote: post #116

Earlier quoted context omitted.

Nah. The real philosophical headache is that we still haven’t solved the hard problem of consciousness, and we’re disappointed because we hoped in our hearts (if not out loud) that building AI would give us some shred of insight into the rich and mysterious experience of life we somehow incontrovertibly perceive but can’t explain. Instead we got a machine that can outwardly present as human, can do tasks we had thoug…

I think Metzinger nailed it, we aren't conscious at all. We confuse the map for the territory in thinking the model we build to predict our other models is us. We are a collection of models a few of which create the illusion of consciousness. Someone is going to connect a handful of already existing models in a way that gives an AI the same illusion sooner rather than later. That will be an interesting day.

I mean, I'm conscious to a degree, and can alter that state through a number of activities. I can't speak for you or Metzinger ;).

But seriously, I get why free will is troubleaome, but the fact people can choose a thing, work at the thing, and effectuate the change against a set of options they had never considered before an initial moment of choice is strong and sufficient evidence against anti free will claims. It is literally what free will is.

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

#196

Post by starchild3001 »

I think this essay lands on a useful framing, even if you don’t buy its every prescription. If we zoom out, history shows two things happening in parallel: (1) brute-force scaling driving surprising leaps, and (2) system-level engineering figuring out how to harness those leaps reliably. GPUs themselves are a good analogy: Moore’s Law gave us the raw FLOPs, but CUDA, memory hierarchies, and driver stacks are what made them usable at scale.

Right now, LLMs feel like they’re at the same stage as raw FLOPs; impressive, but unwieldy. You can already see the beginnings of "systems thinking" in products like Claude Code, tool-augmented agents, and memory-augmented frameworks. They’re crude, but they point toward a future where orchestration matters as much as parameter count.

I don’t think the "bitter lesson" and the "engineering problem" thesis are mutually exclusive. The bitter lesson tells us that compute + general methods win out over handcrafted rules. The engineering thesis is about how to wrap those general methods in scaffolding that gives them persistence, reliability, and composability. Without that scaffolding, we’ll keep getting flashy demos that break when you push them past a few turns of reasoning.

So maybe the real path forward is not "bigger vs. smarter," but bigger + engineered smarter. Scaling gives you raw capability; engineering decides whether that capability can be used in a way that looks like general intelligence instead of memoryless autocomplete.

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

#197

Post by tomrod »

fuckaj wrote: post #110
ants_everywhere wrote: post #90

Earlier quoted context omitted.

> there is missing philosophy I doubt it. Human intelligence evolved from organisms much less intelligent than LLMs and no philosophy was needed. Just trial and error and competition.

The physical universe has much higher throughput and lower latency than our computer emulating a digital world.

Wouldn't it be nice if LLMs emulated the real world!

They predict next likely text token. That we can do so much with that is an absolute testament to the brilliance of researchers, engineers, and product builders.

We are not yet creating a god in any sense.

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

#198

Post by ComplexSystems »

I would like to see what happens if some company devoted your resources to just training a model that is a total beast at math. Feed it a ridiculous amount of functional analysis and machine learning papers, and just make the best model possible for this one task. Then instead of trying to make it cheap so everyone can use it, just set it on the task of figuring out something better than the current architecture and literally have it do nothing else but that and make something based on whatever it figures out. Will it come up with something better than AdamW for optimization? Than transformers for approximating a distribution from a random sample? I don't know, but: what is the point of training any other model?

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

#200

Post by aorloff »

mdp2021 wrote: post #73
aorloff wrote: post #41

I keep asking it, and nobody wants to answer because it doesn't fit within the paradigm of "making AGI" What if intelligence requires agency ?

Agency within the world model is sufficient.

Just a matter of upping the model resolution then ?
Post reply on HN