[cubefox]

He (and most people on HN) are still underestimating the impact of this.

Everything points at AI soon becoming far more intelligent than any human. That's not science fiction about the remote future in a hundred years. We are talking about something in the order of ten years. We are actually at the brink of the "technological singularity". It's happening.

It's happening.

[nly]

I would wager in 10 years time we're all looking back in dismay about how AI has just been used to produce an endless stream of fake news, empty-feeling generated content, spam, ads and other forms of monetisation that make our lives miserable.

Some menial administrative jobs that have virtually no value but aren't yet fully automated will be wiped out by conglomerates who offer their white label AI solutions at scale...but that's it.

Think about Google search. Has it really improved much in the last 20 years? Some HN frequents will swear it has gotten worse.

What about your desktop computer? Sure it's about 10x more powerful than 2005, but has having one really changed your life much since then?

Diminishing marginal returns.

[latchkey]

> I would wager in 10 years time we're all looking back in dismay about how AI has just been used to produce an endless stream of fake news, empty-feeling generated content, spam, ads and other forms of monetisation that make our lives miserable.

I don't think we will have wait 10 months for that, let alone 10 years.

[Filligree]

The size of my chat log with ChatGPT speaks against that. The vast majority of the conversations were productive.

Of the remainder, most were highly entertaining—-which is a form of productivity in itself.

[latchkey]

I don't know what you were responding to. We're not talking about personal experiences, we are talking about people using this stuff to annoy others.

[Filligree]

Misunderstanding, maybe? You were responding to an absolute. I’m pointing out there’s plenty of use cases besides that.

[cubefox]

AI progress has strongly sped up in the last years. We have come a long way since the simple digit recognizers. And there is no end in sight, no AI winter. There are no diminishing returns in intelligence. Strong AI is the last invention.

[somewhereoutth]

Nothing points to this I'm afraid. All we have with LLMs is a probabilistic word salad generator that is realistic enough to get people who should / should not know better very excited.

It is better than Google search because it puts the results into a realistic sounding piece of text. It is worse than Google search because it does not provide sources or context. Both will steal the information you feed in.

Exactly no progress has been made in understanding or creating real intelligence in any meaningful sense, and we should have no expectation of any progress anytime soon while everyone is so focused on purely syntactical learning environments. Nobody learns anything by feeding them a dictionary.

[cubefox]

The leading neuroscientific theory of the brain says that, at it's core, it predicts experiences. It's called predictive coding. LLM's don't predict real time sensory data, they predict just text, and not a real time stream. But the conceptual similarity is still remarkable, both are at their core predictors.

[dwaltrip]

> probabilistic word salad generator

> Exactly no progress has been made in understanding or creating real intelligence in any meaningful sense

Strong words…

The thing has understanding. It may be alien and imperfect, with strange failure modes and quirks. But it has significant understanding of the real world. At this point, there are countless compelling examples demonstrating this.

The only way the model could so effectively predict the next token of such an immense and diverse collection of texts is by acquiring some level of understanding of the circumstances that gave rise to those texts.

The more I use gpt-4, this clearer this is to me, despite its limitations.

Edit: Removed unnecessary snark.

[somewhereoutth]

But it is being given only those texts, not the circumstances. All that is can do is associate and surface patterns in the texts - very complex though those patterns may be. It should not be surprising that those patterns are recognizable, and even that they appear 'real', as they are derived from text about our world. However there cannot be any understanding beyond that, just as with the shadows on the wall of Plato's cave.

[dwaltrip]

Are you arguing the humans are not merely looking at the shadows on the cave wall but LLMs are? Or that the shadows that the LLMs sees are fundamentally different / worse than ours?

> All that is can do is associate and surface patterns in the texts - very complex though those patterns may be

One of the more interesting ideas I’ve heard is that text contains embedded world models that are far richer that we previously imagined. We missed this because it is something we take entirely for granted and couldn’t imagine it any other way, like a fish who doesn’t realize they are swimming in “water”. This seems like a very plausible explanation for the performance of gpt-4 and the like.

[somewhereoutth]

We humans see and digest much much more than just texts we have been given to read. Indeed, and texts are exactly the shadows on the cave wall that we create and share to describe and remember the world around us - but without direct experience of that world around us they can only be shadows.

Yes quite plausible that extraordinarily complex models are embedded in the text corpus, but these can only relate to the texts themselves, as that is all that exists as far as the LLM is concerned. Again, it should not be a great surprise that they correlate to our real world to some extent, but they cannot go beyond those input texts.

[ThrowawayTestr]

> It is worse than Google search because it does not provide sources or context.

Try Bing chat. It provides sources for it's responses.

[ilaksh]

Way less than ten years. You have to also consider the speed of input, thought (output), and communication. GPT-4 is basically human equivalent IQ but probably faster than human. PaLM 2 is not quite human level IQ but is something like 20 times faster than human output.

The input rate of the Claude model is incredible, digesting a book in 20 seconds. Still not quite human level IQ but pretty close.

There is relatively low-hanging fruit for optimizing the models and software and also the hardware for this specific application. We can expect at least another 10 times performance improvement within a couple of years.

And of course as far as IQ that we can measure we can assume it will probably come close to top marks within a few years. But you should scale that based on the incredible speed and other advantages of being digital.

[cubefox]

Pure LLMs alone can probably not go vastly beyond human intelligence because they rely on imitating humans (human text). I think there is still one or two breakthroughs missing to get robotics solved.

[aoeusnth1]

PaLM aalready includes some LLM-generated training (consensus of different approaches), and these kinds of synthetic self-driven training metrics will only get more sophisticated and effective at improving the capabilities. It’s conceivable that we will start seeing AlphaZero-like improvement curves in reasoning.

[cubefox]

I'm not sure how consensus would get you significantly above human baseline. Doesn't that just get you some sort of average?

The basic problem with synthetic self-training is that we need some reward function which tells us whether a given synthetic example is good. In case of AlphaGo Zero, this was a synthetic strategy which won the game, or scored a lot. Which can be automatically detected. But how do we automatically recognize that synthetic text has "high quality"?

One case where it might work is proofs in a formal proof language which can be checked automatically via software. So if a language model is tasked to generate synthetic conjecture/proof pairs, it is possible to automatically recognize the correct ones, and use that for self-training data (unsupervised, supervised, reinforcement, I'm not sure), enabling it to recursively create more complex synthetic proofs.

A very similar approach (with some sort of unit tests instead of proofs) is described here in more detail: https://arxiv.org/abs/2207.14502 It was a while that I read it, so my description above is kinda fuzzy. It might involve some adversarial step that I missed.

One problem is to get this process off the ground (bootstrapping), which is difficult, since we need some baseline capability first to create any successful synthetic examples, and there aren't a lot of human created formal proofs which can be used as bootstrapping training data.

Another problem is that, even if it worked, this system would just be good at generating proofs. Maybe there is some amount of transfer to natural language intelligence, but I'm not sure about that.

If you have a different idea for creating a reward signal, I would be interested how it could be done.

[datathrow0007]

Yes, but that would require academics to grow a practical bone in their body and realize "understanding" language is just one piece of the "intelligence puzzle."

The hype will die down on LLMs (slowly... all those ML researchers need to justify the sunk-cost of specializing into a very niche, albeit sales-friendly, field).

If it's any consolation, I doubt we'll see progress towards a "technological singularity" until the current crop of career scientists retire into the dirt -- or there is a fundamental change in resource allocation allowing other, more creative types to start experimenting with building out a probabilistic model for:

Human language in -> determination of "what to do"/workflow to run -> run

We're seeing a few startups in this area, but I haven't seen anyone create any useful agent.