Startups competing with OpenAI's GPT-3 all need to solve the same problems Text-generating language models are difficult to control. These systems have no sense of morality: they can spew hate speech and misinformation. Despite this, numerous companies believe this kind of software is good enough to sell. OpenAI launched its powerful GPT-3 to the masses in 2020; it also has an exclusive licensing …
There is a fundamental flaw in GPT-3 (and which Google always had).
It is that languages provide base sets but these are not culture or sub-culture differentiated in AI training. In other words, within English, for example, there are many users in distinct cultures each having their own specialized meanings and even separate words. The AI language engine must have a means of maintaining separate dictionaries but joining them when suitable to do so. For example, there should be a general population dictionary but then also maybe a doctor's, medical, dictionary, and a lawyer's, and a chef's, and a gangster's etc. GPT-3 can't do that. Therefore there IS a need for someone(s) to step up and do augmented trained vocabularies. However there is no need for training for words like 'shiznit', potrzebie, or sploodge, though they might be used in banking and finance, in my nightmares.
That, I think, is part of a greater problem: in order to "train" these models, you need a huge amount of raw material. There likely isn't enough if you subdivide it into domain-specific chunks. And equally, owing to the volume, it's infeasible to filter out the "unsuitable" material on the input side that result in the need for low-pass (in the quality sense) profanity filters on the output side.
My uninformed guess is perhaps that the systems might need to know a little bit more about languages so they can ingest their training data more efficiently as neither approach on its own seems to be without a fundamental limitation.
On a different level, it's just a hypocritical statement. They say that people shouldn't develop their own model because "it would be massively environmentally costly, compute costly, and we should be trying to share resources as much as possible", but they can develop their own model despite those costs. Basically, what their argument boils down to is "We can build our model, and you should pay us for ours instead of doing one of your own because you will be harming the environment".
This shallow attempt to put down competition might bother me except the product's so obviously useless and I really don't want people to waste resources making more of it.
I strongly doubt that those problems can be solved, within the current paradigms.
A language model just knows what bits of text tend to go together. The reason it can answer correctly "how many teeth humans have" is just that those words tend to go together with "32" in the training set.
The same is not true for "how many teeth math teachers have". The "math teacher" bit there is not connected to "teeth" at all. It would have to go through "humans" in order to arrive at "32", but "humans" doesn't appear in the question at all, and it's not a word that's so strongly connect to either "math teacher" or "teeth" that the model will give it a high weight all by itself.
The reason "math teacher" is not strongly connected to "human" is that that connection, for people, comes from non-textual experience in a very obvious fashion, so we don't write it very much at all. But a language model only has text, so that connection is very weak.
So: I don't think those problems can be solved within the paradigm of language models. You'd need something that's also trained on real-world input somehow.
Yes, exactly.
Ultimately, we don't even have a decent definition of what "true intelligence" is, nevermind how it works. Who knows - it might turn out to be fundamentally statistics too! If it isn't, then the current wave of AI research is fundamentally wrong and won't ever work really well.
But even if it is, it's still mostly not statistics on text. Most of our intellect works properly before we learn to read, and people who can't read are usually not mentally deficient either.
So, I feel fairly confident in saying that we are never going to get "common sense" or reliable correctness out of text models alone, no matter how many teraflops we throw at them, or how many clever tweaks we apply.
How do you know that that's not how you think? Maybe you're statistics all the way down.
Realistically, we're simulating intelligence the best way we know how, but if it's related to the way WE think is rather tangential because a reasonably accurate simulation of something is good enough to be perceived as that thing and once you get to that point it doesn't matter much. We haven't gotten there yet, but it doesn't make the approach bad.
I mostly agree with you. However, I take exception to this part: "Realistically, we're simulating intelligence the best way we know how".
No, we're not. Collecting terabytes of text someone else wrote and writing a search algorithm isn't simulating intelligence. We have that already. It's called the "I'm feeling lucky" button on Google, and how many people use that? It's obvious what a text prediction model isn't doing, and if you need to simulate intelligence, you'll need to do at least some of those things. A child can make the connection that teachers are usually human, if they teach mathematics, they're still teachers, and therefore understand questions even if they don't have the answer. This model can't do that. There's a way to check whether it understands the question. Ask a five-year-old how many teeth a math teacher has, and they will likely say "I don't know" because they're unaware of the 32 number. Or maybe they'll guess, but they will use intonation to indicate that they're doing that. GPT won't do that.
Which seems to indicate a lack of context is all. Just more data to feed into the system to create those associations that teacher = human.
Try something no AI could figure out that a human can do easily, even as a child in the sticks.
There's a "feeling of making a right argument" and a "feeling of making something up" in humans. So we can exert pressure via this mechanism and notice that we're talking nonsense. But I don't know that the system underneath that, the system that generates the broad strokes of "Well, a math teacher is a ... " " ... human, so they would ... " "... have 32 teeth" in humans is fundamentally different from a text predictor.
I've noticed myself saying things that are utter nonsense, just because they're words that were historically associated. *Usually* I catch myself before actually vocalizing them, or at least notice in hindsight. But GPT has no module that could notice that. That said, there are systems like that - generative-adversarial networks. It seems possible that a transformer set up like a GAN, with a babbler and a nonsense-noticer, could approach the human tier or even surpass it.
(Why surpass it? For "just the part of my brain that makes up things that I could possibly say", GPT-3 is extremely well informed. It would not surprise me if it already has human-level "overhang".)
On a more serious note, machine prose generation is already widely used commercially, for example in sports and financial reporting, and in the niche non-fiction market. Philip Parker's system is the best-known example of the last; his Icon Group International claims to have published over a million titles, either in electronic form or print-on-demand.
The first documented commercially-published machine-generated novel seems to have been True Love from SPb, in 2008. It wasn't very novel, actually, being a style-transformed pastiche of Anna Karenina,1 but it was produced by software. There appear to be a number of commercial projects in this area so it's likely there are quite a few more commercially-published machine-generated novels by now.
And of course there is extensive academic and commercial research in the field of machine-generated prose, and related fields such as machine-assisted prose generation.2 Doyle's efforts are not representative of the state of the art. A few years back CACM had a cover story on computational journalism, for example.
(Note, too, that in some other fields computer-generated "creative" work has been around longer and had greater success. Computer-generated classical music was winning praise from critics in the 1990s.)
1If you're going to steal, steal from the best.
2I've done some work in computational rhetoric, for example. Computational narratology, computational adaptations of formalist and structuralist literary analysis, computational folklore, computational psychology ... there are many cognate fields of research.
It seems to me that if you scrape Reddit & Twatter for your training text the "AI" will be generating an awful lot of 'interesting' language.
Project Gutenburg might be more useful, but the generated patterns/word sequences would resemble turn of last century (1900s) language....
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
