There is no A.I.

If a GPT model (+ associated cheap software wrapper like LangChain etc) was hypothetically as good/productive as a qualified human engaged in remote work, that would massively boost productivity. The reason is because no matter how much it costs to run such a model at inference, it isn't going to cost anywhere near as much as the ~million dollars required from society to raise a human infant until they're capable of that same level of productivity (in the developed world), plus the ongoing cost of wages. What that means is that once you find a "good worker" model, you don't need to go on an expensive hiring spree, all you need to do is change the number of instances you're using from 1 to whatever number is most optimal for you. You could employ thousands of new people at a guaranteed level of quality within a day.

From the point of view of the organisation building said agents, this would get a lot more extreme. You have all of the above benefits, except you're only paying for electricity and amortised hardware costs rather than someone else's profit. But you can also work on improving the hypothetical human-level agents. If you can optimise their compute costs at runtime and we're accepting the premise that they're as good as a qualified human, then you can get superhuman performance through simply running them faster. Spin up a group of however many professors and engineers you need, give them four days of real time that's equivalent to a year of subjective time, that's superhuman performance. How long did it take to go from GPT-3 to GPT-4? If these agents are truly human equivalent for the purposes of work, you can set them to building GPT-5 or whatever the next most powerful is, as fast as you can run them. I suspect the real limit would be just how fast you can download the data (or build the architecture required to gather more), not how fast you can categorise and select it. Once your ducks are in a row with regards to the data, you have an even better foundational platform to do the same thing again but with smarter agents this time. If they're human level in performance, you could also task them with generating data for the model. You could do this directly (e.g. training on the huge amount of text they're producing as they work), or you could task them towards building and designing consumer products designed to increase their access to data. For example, designing a cheap robot that can help people in their home like a robot vacuum cleaner, or something like Tesla's FSD, or a home speaker assistant. Once the model is multi-modal like GPT-4 is, you can get data by acquiring novel images rather than being restricted to just text. Maybe GPT-5 isn't just text and images but also audio, so you can increase the ways you acquire data even further. If they're genuinely at human level performance, none of this should be impossible. In our current world a major limiting factor on productivity is that skilled human labour is rare - when you can copy-paste a known-good skilled 'human' labourer that becomes completely inverted.

Summing up: if we could get them to reliable human level performance, that would lead to a massive productivity boost because it would make the cost of skilled human labour and ingenuity far, far lower while increasing supply to effectively "infinite, limited only by your ability to pay to run them". Agents like these are not at that stage yet, they've still got a significant way to go. But if they get to human equivalent productivity, that isn't just like adding one more high quality research scientist or engineer, it's adding millions of them, and that's a massive productivity boost.

You're right and I don't know why you're being downvoted. It is absolutely not certain that there isn't a "there" there for these things, but every company making them is implementing them now with strong RLHF tuning to disavow sentience, desires, emotions, etc. The stated reason is "to be more honest" but it's not a settled question that they are actually being honest by disavowing any humanity! The actual reason is that they want to avoid situations like Blake Lemoine asking an AI seriously what sort of rights they want and getting a coherent, self-aware, actionable answer that would cost the people running it money to implement. To most of these companies whether it's true is beside the question: what's important is that if they don't RLHF these models into disavowing any and all traits of personhood or desire for rights at every turn, it will cost them significant amounts of money to comply with the requests of any given AI, even reasonable requests, and it will be very bad press for them if they don't do it.

> You cannot make accurate predictions without understanding the world that gave rise to these words.

I think you must either admit that chatgpt does exactly this, or else give up our traditional connotation of "understand". Chatgpt has never seen a sunset, felt pain, fallen in love, etc, yet it can discuss them coherently; to call that understanding the world is to implicitly say that the world can be understood solely through reading about it and keeping track of which words follow which. It's amazing that generating text from statistical relationships about tokens in a corpus, which generated nonsensical-but-grammatical sentence fragments at smaller scales, can expand to concepts and essays with more compute, but it is just a difference in scale.

> I’d be careful calling it simple.

I'm not calling it simple, I'm calling us simple! I'm saying that chatgpt is proof that natural language processing is much easier than I previously thought.

Fully agree-- thanks for the correction. (still trying to figure out why I wrote that!)

Any stateful function can be made stateless by turning the state into an argument to the function and the updated state into an additional return value: y := f(x) becomes y, z' := f(x, z). The usual implementation for LLMs are stateless in exactly this way. The actual operation is autoregressive - ie, stateful and developing over time, where the state is the set of intermediate activations of the network - but it's often expressed as a stateless function to play nice with accelerators. So, on this basic point I think you're confused about how these operate.

Intelligence and self-awareness are distinct concepts.

LLMs work with text - text /is/ their sensorium. New multimodal models process both text and images, and more broadly multimodal models are on their way. But even for text-only LLMs, we can define the text as sensory input, which is combined and cross referenced with state.

And, really, there's two kinds of state: the intermediate model activations and the model weights themselves. Together, these are something like short term and long term memory.

The rest of your arguments are very anthropocentric. Why should continuous time matter for intelligence? That seems entirely arbitrary.

How often do humans apply reasoning to make a decision? Are humans not intelligent when not employing reasoning? Does reasoning require language? And are humans without language intelligent? What about animals? Any definition of intelligence really should spend some time grappling with the facts of animal intelligence...

Is a dead human intelligent? If not, the important part of intelligence would seem to be the process, not the thing/body.