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

A new AI winter is coming?

taranis.ie

Post reply on HN
Print view View original

271–280 of 301 posts

Re: A new AI winter is coming?

#271

Post by claytongulick »

Uehreka wrote: post #41
gishh wrote: post #21

Earlier quoted context omitted.

> something fundamental has changed that enables a computer to pretty effectively understand natural language. You understand how the tech works right? It's statistics and tokens. The computer understands nothing. Creating "understanding" would be a breakthrough. Edit: I wasn't trying to be a jerk. I sincerely wasn't. I don't "understand" how LLMs "understand" anything. I'd be super pumped to learn that bit. I don't…

“You understand how the brain works right? It’s neurons and electrical charges. The brain understands nothing.” I’m always struck by how confidently people assert stuff like this, as if the fact that we can easily comprehend the low-level structure somehow invalidates the reality of the higher-level structures. As if we know concretely that the human mind is something other than emergent complexity arising from simpl…

You understand how reality works right?

It's all just atoms clinging to each other.

Simple.

Heh.

Re: A new AI winter is coming?

#272

Post by encyclopedism »

stanfordkid wrote: post #16

This article uses the computational complexity hammer way too hard, discounts huge progress in every field of AI outside of the hot trend of transformers and LLMs. Nobody is saying the future of AI is autoregressive and this article pretty much ignores any of the research that has been posted here around diffusion based text generation or how it can be combined with autoregressive methods… discounts multi-modal model…

> The fact is: something fundamental has changed that enables a computer to pretty effectively understand natural language.

I have commented elsewhere but this bears repeating

LLM's do not think, understand, reason, reflect, comprehend and they never shall.

If you had enough paper and ink and the patience to go through it, you could take all the training data and manually step through and train the same model. Then once you have trained the model you could use even more pen and paper to step through the correct prompts to arrive at the answer. All of this would be a completely mechanical process. This really does bear thinking about. It's amazing the results that LLM's are able to acheive. But let's not kid ourselves and start throwing about terms like AGI or emergence just yet. It makes a mechanical process seem magical (as do computers in general).

I should add it also makes sense as to why it would, just look at the volume of human knowledge (the training data). It's the training data with the mass quite literally of mankind's knowledge, genius, logic, inferences, language and intellect that does the heavy lifting.

Re: A new AI winter is coming?

#273

Post by krupan »

ako wrote: post #205
krupan wrote: post #191

Earlier quoted context omitted.

TFA's while point is that there is no easy way to tell if LLM output is correct or not. Driving mistakes provide instant feedback if the output of whatever AI is driving is correct or not. Bad comparison.

Many of the things that LLMs will output can be validated in a feedback loop, e.g., programming. It's easy to validate the generated code with a compiler, unit tests, etc. LLMs will excel in processes that can provide a validating feedback loop.

I love how everyone thinks software is easy to validate now. Like seriously, do you have any awareness at all about how much is invested in testing software by the likes of Microsoft, the game studios, and any other serious producers of software? It's a lot, and they still release buggy code.

Re: A new AI winter is coming?

#274

Post by encyclopedism »

aydyn wrote: post #183
leptons wrote: post #145

Earlier quoted context omitted.

If the information it gives is wrong, but is grammatically correct, then the "AI" has fulfilled its purpose. So it isn't really "wrong output" because that is what the system was designed to do. The problem is when people use "AI" and expect it will produce truthful responses - it was never designed to do that.

You are preaching to the choir. But the point is that everyone uses the phrase "hallucinations" and language is just how people use it. In this forum at least, I expect everyone to understand that it is simply the result of next token generation and not an edge case failure mode.

I would have thought to assume that, but given how many on HN throw about how LLM's can think, reason, understand I think it does bear clearly defining some of the terms used.

Re: A new AI winter is coming?

#275

Post by encyclopedism »

frotaur wrote: post #35
gishh wrote: post #21

Earlier quoted context omitted.

> something fundamental has changed that enables a computer to pretty effectively understand natural language. You understand how the tech works right? It's statistics and tokens. The computer understands nothing. Creating "understanding" would be a breakthrough. Edit: I wasn't trying to be a jerk. I sincerely wasn't. I don't "understand" how LLMs "understand" anything. I'd be super pumped to learn that bit. I don't…

It astonishes me how people can make categorical judgements on things as hard to define as 'understanding'. I would say that, except for the observable and testable performance, what else can you say about understanding? It is a fact that LLMs are getting better at many tasks. From their performance, they seem to have an understanding of say python. The mechanistic way this understanding arises is different than huma…

You can make categorical judgements on many things that are hard to define. Often by taking the opposite approach. It's hard to define what consciousness itself is yet I can define what it isn't and that teases a lot out.

So I can categorically say LLM's do not understand by quite literally understanding what NOT understanding is.

We know what LLM's are and what they are NOT.

Please see my earlier comment above:

> LLM's do not think, understand, reason, reflect, comprehend and they never shall.

Re: A new AI winter is coming?

#276

Post by encyclopedism »

fzeroracer wrote: post #247
LatencyKills wrote: post #48

Earlier quoted context omitted.

As someone who was an engineer on the original Copilot team, yes I understand how tech works. You don’t know how your own mind “understands” something. No one on the planet can even describe how human understanding works. Yes, LLMs are vast statistical engines but that doesn’t mean something interesting isn’t going on. At this point I’d argue that humans “hallucinate” and/or provide wrong answers far more often than…

> At this point I’d argue that humans “hallucinate” and/or provide wrong answers far more often than SOTA LLMs. Humans are remarkably consistent in their behavior in trained environments. That's why we trust humans to perform dangerous, precise and high stakes tasks. Humans have the meta-cognitive abilities to understand when their abilities are insufficient or when they need to reinforce their own understanding, to…

There is a qualitative difference between humans and LLM's 'hallucinating' (if we can even apply this terminology to humans which I contend is in-appropriate).

I'd add a simple thought experiment. A poor student doing a multi-choice exam paper and subsequently achieving a poor mark lets say 30% and now a child of 10 years attempting the same paper and achieving say 50%. Looked quantitively a perspective arises attributing understanding to a child who has chanced 50% on a multi-choice paper when compared to a student having studied the subject at hand.

Qualitatively however and we know this intuitively it is certainly NOT the case that the child of 10 comprehends or understands more than the poor student.

Re: A new AI winter is coming?

#277

Post by Eddy_Viscosity2 »

emp17344 wrote: post #141
Eddy_Viscosity2 wrote: post #45

Earlier quoted context omitted.

It could very well be that statistics and tokens is how our brains work at the computational level too. Just that our algorithms have slightly better heuristics due to all those millennia of A/B testing of our ancestors.

Except we know for a fact that the brain doesn’t work that way. You’re ignoring the entire history of neuroscience.

I didn't know that neuroscience even claimed to a full enough understanding of how people think to conclusively disprove this.

Re: A new AI winter is coming?

#278

Post by thr0waw3y »

hsuduebc2 wrote: post #115
dangus wrote: post #57

Earlier quoted context omitted.

I just want to point out a random anecdote. Literally yesterday ChatGPT hallucinated an entire feature of a mod for a video game I am playing including making up a fake console command. It just straight up doesn’t exist, it just seemed like a relatively plausible thing to exist. This is still happening. It never stopped happening. I don’t even see a real slowdown in how often it happens. It sometimes feels like the o…

When asking question I use chatgpt only as turbo search engine. Having it double check it's sources and citations helped tremendously.

I find it incredibly useful for information retrieval from dense, archival-like text knowledge. I research cellular networks, and everything on Google/DDG is just fluffy SEO spam, but I find Gemini can reliably hone into the precise subsection out of tens of thousands of dense standards to tell me what 5G should do in a given scenario

Re: A new AI winter is coming?

#279

Post by lambdaone »

encyclopedism wrote: post #272
stanfordkid wrote: post #16

This article uses the computational complexity hammer way too hard, discounts huge progress in every field of AI outside of the hot trend of transformers and LLMs. Nobody is saying the future of AI is autoregressive and this article pretty much ignores any of the research that has been posted here around diffusion based text generation or how it can be combined with autoregressive methods… discounts multi-modal model…

> The fact is: something fundamental has changed that enables a computer to pretty effectively understand natural language. I have commented elsewhere but this bears repeating LLM's do not think, understand, reason, reflect, comprehend and they never shall. If you had enough paper and ink and the patience to go through it, you could take all the training data and manually step through and train the same model. Then o…

You could use a similar argument to argue that "brains do not think, understand, reason, reflect, comprehend and they never shall." After all, there's nothing in there but neurons, synapses and other biological gunk. If you look at it that way.

Which you shouldn't.

Re: A new AI winter is coming?

#280

Post by geekymartian »

marknutter wrote: post #153
beachtaxidriver wrote: post #93

Earlier quoted context omitted.

It's true, but do you really trust the AI generated + Nurse Review output more than Organic Nurse generated? In my experience, management types use the fact that AI generated + Nurse Review is faster to push a higher quota of forms generated per hour. Eventually, from fatigue or boredom, the human in the loop just ends up being a rubber stamper. Would you trust this with your own or your children's life? The human in…

These same questions could be asked about self driving cars, but they've been shown to be consistently safer drivers than humans. If this guy is getting consistently better results from ai+human than it is from just humans, what would it matter if the former results in errors given the latter results in more and costs more?

You can't compare an LLM output with a self driven car. That's the flaw of using the term AI for everything, it brings two completely different technologies to an artificial level ground.
Post reply on HN