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

AGI is far from inevitable

ru.nl

Post reply on HN
Print view View original

161–170 of 258 posts

Re: AGI is far from inevitable

#161

Post by famouswaffles »

riku_iki wrote: post #154
famouswaffles wrote: post #153

Earlier quoted context omitted.

> there are some benchmarks which show fundamental inability of LLM perform certain tasks which human can, for example add 100 digits numbers. fundamental inability ? No. Current Sota LLM (4o, claude, gemini) woes with arithmetic is not a transformer weakness never mind a large language modelling one. Those benchmarks show that those particular models have problems with accuracy on that many digits, not that LLMs fun…

> https://arxiv.org/abs/2405.17399 they built specialized model which is after bunch of trickery still has 99% accuracy(naive model had very low accuracy) on very simple deterministic algo. I also think most of the accuracy came from memorization of training set(model didn't provide intermediate results, and started failing significantly at sligtly larger input). In my book it is fundamental inability to learn and re…

>I also think most of the accuracy came from memorization of training set

Oh yes..it memorized a 20 digit training set to solve 100 digit problems. That makes sense. Lol

>(model didn't provide intermediate results, and started failing significantly at sligtly larger input).

No it didn't. They tested up to 100 digits with very high accuracy. I don't think you even read the abstract of this, nevermind the actual paper.

Re: AGI is far from inevitable

#162

Post by jebarker »

gizmo686 wrote: post #114
ryandvm wrote: post #112

Earlier quoted context omitted.

> but LLMs do solve problems that people thought were extremely difficult to solve ten years ago Agreed. I would have laughed you out of the room 5 years ago if you told me AI's would be writing code or carrying on coherent discussions on pretty complex topics in 2024. As far as I'm concerned, all bets are off after the collective jaw drop that the entire software engineering industry did when we saw GPT4 released. W…

> It turns out that the larger these models get, the more unexpected emergent capabilities they have, so I'm mostly in the camp of thinking AGI is just a matter of time and resources. AI research has a long history of people saying this. Whenever there is a new fundamental improvement, it looks like you can just keep getting better results by throwing more resources at it. However, eventually we end up reaching a poi…

> LLMs have an additional problem related to training data. We are already throwing all the data we can get our hands on at them.

I think it's plausible we'll see a breakthrough in data efficiency that helps here. Humans are an existence proof that language is learnable through much less data and, in theory, facts should only need to be seen once. LLMs in their current form seem very data inefficient.

Re: AGI is far from inevitable

#163

Post by riku_iki »

famouswaffles wrote: post #161
riku_iki wrote: post #154

Earlier quoted context omitted.

> https://arxiv.org/abs/2405.17399 they built specialized model which is after bunch of trickery still has 99% accuracy(naive model had very low accuracy) on very simple deterministic algo. I also think most of the accuracy came from memorization of training set(model didn't provide intermediate results, and started failing significantly at sligtly larger input). In my book it is fundamental inability to learn and re…

>I also think most of the accuracy came from memorization of training set Oh yes..it memorized a 20 digit training set to solve 100 digit problems. That makes sense. Lol >(model didn't provide intermediate results, and started failing significantly at sligtly larger input). No it didn't. They tested up to 100 digits with very high accuracy. I don't think you even read the abstract of this, nevermind the actual paper.

> No it didn't. They tested up to 100 digits with very high accuracy. I don't think you even read the abstract of this, nevermind the actual paper.

they have two OOD (out of distribution) accuracies in the paper: OOD: up to 100 digits, and 100+ OOD: 100-160 digits. 100+ OOD accuracy is significantly worse: around 30%.

Re: AGI is far from inevitable

#165

Post by oco101 »

scotty79 wrote: post #129
gizmo686 wrote: post #114

Earlier quoted context omitted.

> It turns out that the larger these models get, the more unexpected emergent capabilities they have, so I'm mostly in the camp of thinking AGI is just a matter of time and resources. AI research has a long history of people saying this. Whenever there is a new fundamental improvement, it looks like you can just keep getting better results by throwing more resources at it. However, eventually we end up reaching a poi…

Currently we are throwing everything at LLMs and hope good things stick. At one point we might use AI to select best training data from what's available to best train the next AI.

"That is just one very narrow task that basically anyone can do regardless of intelligence or talent. It takes less than a year to train a distracted 16 year old to do it, but three decades to train an AI and even then you probably need to hand tune it for specific locations because it won’t know what do in unusual road layouts."

It takes three decades to train an AI, but of course, like everything humanity does is not linear, it is exponential. Before the Wright brothers, it was believed that powered, controlled, heavier-than-air flight was impossible. Then, in 1903, they achieved the first successful airplane flight, which lasted 12 seconds and covered 120 feet. By 1914, the first commercial flight covered approximately 21 miles and took about 23 minutes. This is just one example and I don't see why AI should be any different

Re: AGI is far from inevitable

#166

Post by famouswaffles »

riku_iki wrote: post #160
famouswaffles wrote: post #158

Earlier quoted context omitted.

Their method does not require building specialized models from scratch (you can but you don't have to) and they did not prove transformers can't learn sorting. If you think they did, then you don't understand what it means to prove something.

In my books what they built (specialized training data + specialized embeddings format) is exactly specialized model. You can disagree of course and say again that I don't understand something, but discussion will be over.

The data isn't specialized ?

A poor result when testing one model is not proof that the architecture behind the model is incapable of getting good results. It's just that simple. The same way seeing the OG GPT-3 fail at chess was not proof LLMs can't play chess.

This

>They also demonstrated that transformer can't learn sorting.

is just wrong. Nothing more to it.

Re: AGI is far from inevitable

#168

Post by bamboozled »

kylecazar wrote: post #151
ryandvm wrote: post #112

Earlier quoted context omitted.

> but LLMs do solve problems that people thought were extremely difficult to solve ten years ago Agreed. I would have laughed you out of the room 5 years ago if you told me AI's would be writing code or carrying on coherent discussions on pretty complex topics in 2024. As far as I'm concerned, all bets are off after the collective jaw drop that the entire software engineering industry did when we saw GPT4 released. W…

Are capabilities truly 'emerging'? Or are we just observing new competencies/applications we hadn't previously considered.

We've built a machine to complicated, we don't really know what to make of it, fun times.

Re: AGI is far from inevitable

#169

Post by jiggawatts »

ryandvm wrote: post #112
Gehinnn wrote: post #6

Basically the linked article argues like this: > That’s because cognition, or the ability to observe, learn and gain new insight, is incredibly hard to replicate through AI on the scale that it occurs in the human brain. (no other more substantial arguments were given) I'm also very skeptical on seeing AGI soon, but LLMs do solve problems that people thought were extremely difficult to solve ten years ago.

> but LLMs do solve problems that people thought were extremely difficult to solve ten years ago Agreed. I would have laughed you out of the room 5 years ago if you told me AI's would be writing code or carrying on coherent discussions on pretty complex topics in 2024. As far as I'm concerned, all bets are off after the collective jaw drop that the entire software engineering industry did when we saw GPT4 released. W…

One thing that keeps me up at night is that the human genome is only 3 giga base pairs, of which only a fraction encodes the design of our brains — and quite inefficiently at that, through layers of indirections.

That’s sufficient information to produce a system that can learn to think like us! Not just learn but efficiently, with far less input data needed than any current LLM. Literally just a couple of decades of video and audio, only a small fraction of that text!

From what I can tell, the human brain achieves this through scale alone. There’s nothing else that can explain the observed learning capability. The genome is too small to encode learned weights, and we don’t clone our parents’s brains in development.

It’s possible, and a meat computer can do it. Replicating this in silicon is just a matter of time, and it might require only scale and nothing else.

Re: AGI is far from inevitable

#170

Post by numeri »

Gehinnn wrote: post #12

I skimmed through the paper and couldn't make much sense of it. In particular, I don't understand how their results don't imply that human-level intelligence can't exist. After all, earth could be understood as solar powered super computer, that took a couple of million years to produce humanity.

I think their "proof" would also prove that no child could ever learn to behave like an adult.
Post reply on HN