pull down to refresh

Transformer based AI will not lead us to AGI/ASI and is just a hype machine

(i've decided to stay and use SN as a rant space for my ideas, please bear with my indecision)

AI will not lead to AGI, they're next token predictors and no, I'm not an internet schizo doomer I'm basing my claims off of facts.

1. We have never been able to embed common sense into artificial systems

This is one of the reasons why symbolic AI lost its momentum after the initial optimism of the 60s and 70s and had to be revived through the invention of back propagation.

2. LLMs fail at trivial tasks such as spatial reasoning

They absolutely suck at spatial tasks such as imagining object rotation in 3D space and so on, they also suck at vision tasks due to them not understanding distances and physics (images are converted into tokens which the transformer architecture.. predicts) ARC-AGI 2 is a good benchmark to note this failure in performance, most models fail to cross 10 percent whereas humans are at a solid 100 percent.

3. The solution is not more data and scaling

People like Scam Altman (scam because look at "worldc*in, the most dystopian shit i've read in a decade"), have repeatedly parroted the claims that more data and more compute will solve the issue of LLMs being inherently dumb, the way tokenization works (i.e. how data is represented) itself is flawed, leading to issues such as the LLMs not knowing how many r's in a strawberry.

4. LLMs learn patterns in their data, it's a weird case of overfitting

Ask an LLM to generate a random number between 1-6 without code or the internet and it will be 4 most of the times, this and the use of em dashes are all patterns it had learnt from its data, where dice rolls resulted mostly in 4.

5. LLMs simply do not understand math

As stated in the previous point, they simply have memorized the fact that 2+2=4 and not that adding two integers gives you another integer, there is no computation of that operation rather just statistics and probability at play (softmax and its consequences) Read Anthropics paper (Biology of an LLM) for more info on how LLMs are REALLY DUMB

All these are just a few points I have shared here, maybe someday I'll have a more detailed and technical blogpost about this. Now coming to the real issue, LLMs and AI API wrappers have redirected precious research funds towards these statistical parrots, and not only that a recent study by Stanford I think showed that people who over relied on ChatGPT showed signs of cognitive "decay".

We need more physics based and biologically grounded models, bio inspired and neurosymbolic models as Yann LeCun and Gary Marcus have said, instead all we have is a glorified markov chain. fml.

Such machines will never cure cancer, or come up with a solution for poverty and world hunger. It's a VC money burning shitshow, I would like to know your thoughts on this.

view all related items

225 sats \ 3 replies \ @optimism 2 Jul

I would like to know your thoughts on this.

I won't argue with you that LLMs aren't going to evolve into AGI/ASI; I'd find that highly unlikely. I am not even sure how desirable that outcome would be in the first place. So yes, it feels like it's a

VC money burning shitshow

but now I wonder: besides AGI/ASI, are there things that we can do with all this not-AGI stuff that can be useful? And I have a feeling we can (but I'm not sold.)

This past month I have solved 2 issues for myself that were annoying tf out of me for years:

Clickbait. I hate it. So now I auto-leech an article fed to me and let an LLM summarize it for me. I ignore the title and just read the summary. If I like the summary I will take the trouble to read more. If not, I will ignore it and move on. Saves me a lot of frustration. What I do need is a way to filter the really bad feeds better, like zerohedge because it kinda promotes scammers, which I will not leave to an LLM, but just pre-emptively filter out all promoted articles programatically.
Videos. I hate these even more. Because even at 2x speed where you can't understand what they are mumbling, this shit just takes too much time. So this past week I have transposed around 40 vids that were recommended here on SN using this process, minus the training part. Listened to the one recommended by @plebpoet/@ek, read some parts of others, discarded everything that smelled like bs. My NLP chip is anyway idle otherwise, might as well use it for something useful, like saving my time.

And that is in the end how I think that we (rational human beings that are disconnected from the VC's tit) should approach this. Silver bullets don't exist, and even if I were wrong about that, if there were one, it was probably not something like an LLM. Instead, let's just create value. For ourselves, and for each other. Without some effing subscription. Let them waste money on building new datacenters too. When the bubble bursts, we'll have ourselves extremely cheap compute that can be used for real solutions.

PS: If you want to present facts, link something that proves the fact, please.

100 sats \ 1 reply \ @cy OP 3 Jul

you are right to point out that LLMs have been useful, i should've worded my argument better. I'm disappointed in how LLMs are being advertised as something they're not. As for the PS, I have mentioned relevant papers and benchmarks I think

10 sats \ 0 replies \ @optimism 3 Jul

Yeah I share that disappointment.

Re: PS. Oh I just meant literal links. The internet is kinda rotten with tons of opinion that may get upper rank on search engines. Just makes it harder to reconstruct your thinking without them.

110 sats \ 0 replies \ @south_korea_ln 2 Jul

Pretty much agree with OP and this response. Matches my experience and understanding of what LLMs can and cannot do.

I discard any bold claims made by Altman. He needs to convince investors.

110 sats \ 3 replies \ @SwapMarket 2 Jul

AlphaFold is impressive though.

100 sats \ 2 replies \ @cy OP 3 Jul

true, areas like protein prediction and computer vision should be getting more attention than a freaking glorified autocomplete is what I'm saying

50 sats \ 1 reply \ @optimism 3 Jul

I wonder if this is what the whole "safety" dispute between Sam and Ilya was ultimately about. Not safety in the sense of there being SkyNet, but more as in complete misrepresentation of what it truly does. Safety as in "let's not gaslight".

I think that the results of my testing the cashu MCP with latest gen open source LLMs in #1021115 shows that use cases that are obviously outside of the instruct scope such as repeating a bolt11 invoice... bear terrible results.

100 sats \ 0 replies \ @SwapMarket 3 Jul

Nah, Ilya was a student of Geoffrey Hinton. They both are predicting skynet end game.

210 sats \ 0 replies \ @mtk4000 2 Jul

Brains have a bunch of functional specialization in different regions. Expecting an LLM to develop general intelligence is like expecting a human to function with a fraction of one lobe of one hemisphere of the brain.

10 sats \ 1 reply \ @cy OP 2 Jul

why did this get 1k sats the moment I posted it lol

112 sats \ 0 replies \ @SimpleStacker 2 Jul

coz someone liked the post and zapped you 1,000

5 sats \ 4 replies \ @SwapMarket 3 Jul

Letting AI evolve its own code and giving enough time and compute may get us there sooner than we think. Human brain has 500 trillion synapses vs 1.8 trillion parameters in ChatGPT-4. We similarly can do math much worse than a classical computer and suck at generating random numbers. Thought process in our heads is verbal and very similar to an LLM. This internal dialogue gives us self awareness. AI has huge advantage of replication, immortality and the speed of communication. I would not be so cocky.

5 sats \ 3 replies \ @cy OP 3 Jul

Letting AI evolve its own code and giving enough time and compute may get us there sooner than we think. Human brain has 500 trillion synapses vs 1.8 trillion parameters in ChatGPT-4. We similarly can do math much worse than a classical computer and suck at generating random numbers. Thought process in our heads is verbal and very similar to an LLM. This internal dialogue gives us self awareness. AI has huge advantage of replication, immortality and the speed of communication. I would not be so cocky.

I would agree if human synapses and neurons were similar to parameters of an LLM and if humans learnt through back propagation (which they don't btw), I am not talking about the quality of the math but how they do it, we know the rules of additions and apply it, LLMs memorize 2+2=4 and fail to generalize at 427382098402973498279382 + 9847293847203874892374923, it completely makes shit up. There is no "thought process" in LLMs at all, CoT is a sham, technically it's a prompting technique, pls read the paper I've mentioned in the original post.

I am not denying their usefulness, but do not bet on LLMs to come up with anything useful.

5 sats \ 2 replies \ @SwapMarket 3 Jul

427382098402973498279382 + 9847293847203874892374923

can you generalize this? the latest chatgpt I tried solved a-level math exam problems that I already could not. so it is smarter at math than I am. we don't know how humans learn exactly, but neural nets were modeled after our brains structure.

0 sats \ 1 reply \ @cy OP 3 Jul

yes I can generalize it, I know how to add 2 integers and how carry over works, AI does not. neural networks were modeled after the impulse flow/data flow of human neurons and are not based off of neurobiology (this is a common misconception).

you said chatgpt solved a math problem which you solved, it did so because the data was present in its dataset and has seen multiple problems like it, this is, as I said before a weird case of generalisation.

0 sats \ 0 replies \ @SwapMarket 3 Jul

Not really. Despite 250x more synapses, you would need a pen and paper to add these numbers. LLMs don't have to count, but use tools like calculators for that.

5 sats \ 0 replies \ @Scoresby 3 Jul

I pretty much agree. But I'm not sure that a next-token prediction machine based on everything available on the internet is doomed to unhelpfulness. So, they don't think -- what if that's fine? We don't need it to think, we need it to squeeze juice out of the lemon of humanity that we can't otherwise squeeze.

Here's a whimsical hypothetical: if there was a prediction market for next-token generation ("Here's a questions, bet on what the next letter in the answer will be"), and if lots ("lots" is doing a lot of work here) of bets were made every second, would it provide a similar experience to interacting with an LLM?

Probably, if the bets were all being made by real humans it would do substantially better. This may not be a super helpful insight, but maybe it points to some of the problem: there's no "survival of the fittest model" function here. models that make bad next-token predictions get some feedback, but they don't necessarily die.

We got human brains because humans with "bad" brains don't pass on their genes. Prediction markets work (if they work) because predictors who are wrong lose money. What happens when an LLM is wrong?