In 1966 it was clear to everyone that this program
https://en.wikipedia.org/wiki/ELIZA
parasitically depends on the hunger for meaning that people have.
Recently GPT-3 was held back from the public on the pretense that it was "dangerous" but in reality it held back because it is too expensive to run and the public would quickly learn that it can answer any question at all... if you don't mind if the answer is right.
There is this page
https://nlp.stanford.edu/projects/glove/
under which "2. Linear Substructures" there are four projections of the 50-dimensional vector space that would project out just as well from a random matrix because, well, projecting 20 generic points in a 50-dimensional space to 2-dimensions you can make the points fall exactly where you want in 2 dimensions.
Nobody holds them to account over this.
The closest thing I see to the GPT-3 cult is that a Harvard professor said that this thing
https://en.wikipedia.org/wiki/%CA%BBOumuamua
was an alien spacecraft. It's sad and a little scary that people can get away with that, the media picks it up, and they don't face consequences. I am more afraid of that than I am afraid that GPT-99381387 will take over the world.
(e.g. growing up in the 1970s I could look to Einstein for inspiration that intelligence could understand the Universe. Somebody today might as well look forward to being a comic book writer like Stan Lee.)
There appear to be an awful lot of conversations in which people care about other things much, much more than what is objectively correct.
And any technology that can greatly amplify engagement in that kind of conversation probably is dangerous.
The parameter size is so large that it has in essence memorized its training data, so if the right answer was already present in the training data you'll get it, same if the answer is closely related to the training data in a way that lets the model predict it. If the wrong answer was present in the training data you may well get that.
If GPT-3 is known to have some sort of intelligence, I think it logically follows that one can differentiate that from the reflected intelligence of all the humans that produced all the data it ingested.
How do you see intelligence in GPT-3 and not see it in the data fed to it?
Well, no. Linear projections follow a bunch of rules that enforce conservation of various linear structures. You can't manipulate things arbitrarily.
For example, if three points are colinear in the original space, they will be colinear in every projection. Perhaps more relevant to the GloVe examples, if a-b = c-d in the original space then the same equality holds in every projection. Since the projections are Lipschitz continuous, we can also say that is a-b is close to c-d in the original space then they are also close in every projection.
Anyway, the point that these figures illustrate is easily confirmed by downloading the embeddings yourself, so insinuating that the authors are getting away with something that they should be held to account for is silly.
https://www.edge.org/response-detail/27150
Many GPT-3 cultists are educated in computer science so they should know better.
GPT-3's "one pass" processing means that a fixed amount of resources are always used. Thus it can't sort a list of items unless the fixed time it uses is humongous. You might boil the oceans that way but you won't attain AGI.
There are numerous arguments along the line of Turing's halting problem that restrict what that kind of thing can do. As it uses a finite amount of time it can't do anything that could require an unbounded time to complete or that could potentially not terminate.
GPT-3 has no model for dealing with ambiguity or uncertainty. (Other than shooting in the dark.) Practically this requires some ability to backtrack either automatically or as a result of user feedback. The current obscurantism is that you need to have 20 PhD students work for 2 years to write a paper that makes the model "explainable" in some narrow domain. With this insight you can spend another $30 million training a new model that might get the answer right.
A practical system needs to be told that "you did it wrong" and why and then be able to correct itself on the next pass if possible, otherwise in a few passes. Of course a system like that would be a real piece of engineering that people would become familiar with, not a outlet for their religious feelings that is kept on a pedestal.
Actual "understanding" means mapping language to something such as an action (I tell you to get me the plush bear and you get me the plush bear,) precise computer code, etc.
Thus I reject Wittgenstein.
https://arxiv.org/abs/2106.07890
outlines a promising approach to mapping the semantics encoded in these models. Understanding their limits could make your rejection of Wittgenstein all the more precise.
I have used a similar argument to show that the simulation hypothesis is wrong. If any algorithm used to simulate the world takes longer than o(N) time, then the most efficient possible computer for that is the universe which computes everything in O(n) time where n is time. In other words, you never get "lag" in reality no matter how complex the scene you're looking at is. Worse than that, some simulation algorithms are exponential time complexity!
If the containing universe has like 21 dimensions and otherwise have similar tech computers as we do today then you should be able to simulate it on a datacenter just fine as computation ability grows exponentially with number of dimensions. 3 dimensions you have 2 dimensions of computation surface, 21 dimensions and you have 20 dimensions of computation surface, so our current computation to the power of 10. GPT3 used more than a petaflop real time compute during training, so 10 to the power of 15. Using the same hardware in our fictive universe would give us 10 to the power of 150 flops. We estimate atoms in the universe to be about 10 to the power of 80, with this computer we would have 10 to the power of 70 flops of compute per atom, that should be enough even if entanglement gets a bit messy. We have around that much memory per atom as well, so can compute a lot of small boxes and sum over all of it etc, to emulate particle waves. We wouldn't be able to detect computational anomalies on that small scale, so we can't say that there isn't such a computer emulating us.
You see the same thing in all popular reporting about science and tech. Endless battery breakthroughs that will quadruple or 10x capacity become a couple percent improvement in practice. New gravity models mean we might have practical warp drives in 50 years. Fusion that's perpetually 20 years away. Flying cars and personal jetpacks. Moon bases, when we haven't been on the moon since the 70s.
AI reporting and hype is no different. Maybe slightly worse because it's touching on "intelligence", which we still have no clear definition of.
That doesn't seem correct. Intelligence came much later than when most of our brain evolved.
Planaria can move towards and away from things and even learn.
Bees work collectively to harvest nectar from flowers and build hives.
Mammals have a "theory of mind" and are very good at reasoning about what other beings think about what other beings think. For that matter birds are pretty smart in terms of ability to navigate 1000 miles and find the same nest.
People make tools, use language, play chess, bullshit each other and make cults around rationalism and GPT-3.
GPT-3 is obviously not the AI end goal, but we are on the path and the end might lead to aeroplanes than flapping machines.
This isn't actually clear; with things like this we are on a path but it may not lead anywhere that fundamental (at least when we are talking "AI", especially general AI).