[georgehill]

> Given the breadth and depth of GPT-4’s capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.

I don't know why, but my brain refuses to accept GPT-4 as something close to AGI. Maybe I am wrong. It is hard to believe that our brain is just a bunch of attention layers and neural nets.

[Robotbeat]

Well it’s not quite that simple. Brains use spiking neural networks, not the kind used typically in artificial neural networks like those used by LLMs. The “weights” can be changed over time, new connections and even new neurons formed. And the number of connections (“weights”) is about 500-1000x more in our brain than GPT-3. The connection topography is a lot different.

But ultimately, our brains are still just made of neurons. As far as we know, there isn’t some sort of extreme molecular computing going on (ie memories directly stored in RNA or whatever) or any large scale quantum mechanics (temperature too high).

The differences between AI approaches like artificial neural networks and our animal meat brains could be just the difference between a propeller and flapping wings. Same base mechanics (airfoil producing lift as thrust), different substantiation.

[thro1]

Do you consider that every neuron in the brain has unique DNA and ancestorship ?

[thro1]

It seems no. Those are facts - whoever argue with facts (parent still downvoted?).. is an idiot.

https://www.scientificamerican.com/article/scientists-surpri...

https://www.science.org/doi/10.1126/science.aab1785 - Somatic mutation in single human neurons tracks developmental and transcriptional history

(good luck simulating that)

[atleastoptimal]

> It is hard to believe that our brain is just a bunch of attention layers and neural nets.

Our brain isn't, but I'd wager the architectural complexity of a physical, neuronal brain is not optimized for all useful mental tasks, and has perhaps a fair amount of local maxima that are near vestigial in overall positive impact on cognition. Just because the human brain model of cognition is the only way nature has been able to create GI doesn't mean it's the only way GI can be attained.

The best kind of machine is the simplest one needed to produce a desired outcome.

[georgehill]

I agree that GI can have a different implementation compared to our human brain, but one thing is for sure: as of right now, the human brain can become more creative with a fraction of the data consumed by GPT-4.

GPT-4 could be AGI, but it feels like cheating to achieve AGI by feeding the entire internet. If someone can build AGI with only the data that humans consume in their lifetime, then that, imho, is the real AGI.

[angusturner]

I guess the challenge here is that the human mind is not a blank slate, and has been optimized first by billions of years of evolution.

If it takes all the data on the internet (or more) to bootstrap AGI, but that system is then capable of leveraging its knowledge to solve new out-of-distribution tasks, that seems like a fair test to me.

I agree with the article that we see "sparks" of this generality with GPT4.

[singularity2001]

   become more creative with a fraction of the data consumed by GPT-4

not if you understand the input stream of vision as an equivalent input stream of semantic tokens as in multimodal models. under that definition people looking around for 10 years receive much more training data than large language models and thus perform a bit better at zero shot inference.

[scubakid]

Not sure I would call constant real-time perceptual stimuli since before birth "a fraction of the training data."

[Avicebron]

or better yet, chuck it in an open plain and see how long it takes to figure out how to attach a rock to a stick and fight a gazelle to refuel it's energy supply.

[jltsiren]

It's a well-established principle in computer science that the input/output behavior of a system may not capture all of its important properties. Take zero-knowledge proofs for example. Their entire point is that they are indistinguishable from randomly generated garbage from a specific distribution. The proofs only gain value if you make causal assumptions about the system that generated them.

I don't think systems like GPT-4 can ever be truly intelligent, because they simply output randomly generated garbage from a specific distribution. Their output may eventually be indistinguishable from that of a truly intelligent system, but the causal mechanism behind them is not intelligent.

On the other hand, most people lose their ability to think when they are under sufficient pressure (such as fighting for their lives). It's plausible that people are fundamentally no different from systems like GPT-4 in such situations. Then a language model could be a key part of an AGI, but true intelligence would also need higher-level causal mechanisms.

[adroniser]

Do you think anything digital could ever become conscious?

[mrandish]

Based on the wording of your question, I can't see a way today to prove it never could, therefore the answer currently must be "Yes, it may someday be possible."

Note: This assumes that "conscious" as defined in this context is specific enough for the question to ever be meaningfully answered "Yes." This is a non-trivial assumption because there are criteria by which some would judge AIs as already conscious. Alternatively, some philosophers of mind have criteria by which they assert humans aren't conscious.

[coldtea]

The differences between ChatGPT and a conscious human brain are not unsurmountable.

Let's consider a potential future conscious AGI created by advancing from something like ChatGPT.

The human brain is "always on". It's possible to have a digital system be always on, i.e. not just train once and then just respond, but constantly take new input.

The human brain has way more connections/layers than ChatGPT. It's possible to imagine the digital system getting the same number of connections.

The human brain gets real time sensory input. It's possible to add cameras, microphones, etc to that digital system so it gets a constant feed. Maybe even let it process what it saw during the day in a batch training/GC run (we could call that "sleep").

The human brain has a different topology. It's possible to alter the topology of a digital system neural network to mimick that, instead of using the ChatGPT topology. It's not like we're forever doomed to its simpler statistical model. But it's interesting that it already gets very significant emergent intelligence-like properties.

The human brain is self-conscious. This can very well be an emergent property of the above. I think all that's needed is the ability to have some form feedback mechanism.

[Mike_12345]

The question is whether consciousness is computable. Can a Turing machine be conscious? Probably not.

https://www.newscientist.com/article/mg25634130-100-roger-pe...

https://www.youtube.com/watch?v=hXgqik6HXc0

[coldtea]

Note that Penrose's answer is not the "consensus".

Also, Penrose doesn't conver if I recall correctly about modelling the quantum part too. It's just statistics after all.

[Mike_12345]

So the consensus is that consciousness is computable by a Turing machine?

[coldtea]

The consensus is that if it's not, it's not because of the reasons Penrose gives.

[Mike_12345]

Is there a consensus? I haven't been able to find much else via Google search. At that level of theorizing I wouldn't expect any consensus, only original ideas from a few elite researchers.

[steve_adams_86]

I have a feeling it's like the saying "Any sufficiently advanced technology is indistinguishable from magic". At a certain point they could become practically identical things.

[georgehill]

yes

[cowmix]

Conversely, after using ChatGPT-4 (and generally loving it) -- I'm at peace with this maybe fact.

[sdenton4]

There's no rule that agi has to have the same architecture as a human...

[robotresearcher]

There’s no agreement about what an AGI is or does, let alone how it should do it.

[carapace]

It seems clear to me that these systems think in a meaningful sense, but I don't think they are beings. In Cybernetics there is a result that says that any well-regulated system must contain a model of itself. This seems as good a definition as any of a "being", and by this definition these language models don't make the cut.

[Robotbeat]

The architecture for large language models is summarized in the training set for large language models. With fairly minimal modification such as via a plug-in, chatGPT and the like are Turing complete and can thus model themselves.

[carapace]

Hmm, then, from first principles, we should expect "ghosts" to arise in those systems. (These will not be the "virtual entities" that people talk to and call by name (Alexa, Cortana, Siri, etc., but more akin to fixed points in the flow of information.)

[hotpotamus]

"We are the meat in our heads" is the way I've heard it said that human intelligence is just a physical phenomenon created by our brains. And there was never any reason to believe that intelligence could not arise from other substrates.