gpt-4 prediction: it won't be very useful

Word on the street says that OpenAI will be releasing "GPT-4" sometime in early 2023.

There's a lot of hype about it already, though we know very little about it for certain.

----------

People who like to discuss large language models tend to be futurist/forecaster types, and everyone is casting their bets now about what GPT-4 will be like. See e.g. here.

It would accord me higher status in this crowd if I were to make a bunch of highly specific, numerical predictions about GPT-4's capabilities.

I'm not going to do that, because I don't think anyone (including me) really can do this in a way that's more than trivially informed. At best I consider this activity a form of gambling, and at worst it will actively mislead people once the truth is known, blessing the people who "guessed lucky" with an undue aura of deep insight. (And if enough people guess, someone will "guess lucky.")

Why?

There has been a lot of research since GPT-3 on the emergence of capabilities with scale in LLMs, most notably BIG-Bench. Besides the trends that were already obvious with GPT-3 -- on any given task, increased scale is usually helpful and almost never harmful (cf. the Inverse Scaling Prize and my Section 5 here) -- there are not many reliable trends that one could leverage for forecasting.

Within the bounds of "scale almost never hurts," anything goes:

It gets worse.

Most of the research on GPT capabilities (including BIG-Bench) uses the zero/one/few-shot classification paradigm, which is a very narrow lens that arguably misses the real potential of LLMs.

And, even if you fix some operational definition of whether a GPT "has" a given capability, the order in which the capabilities emerge is unpredictable, with little apparent relation to the subjective difficulty of the task. It took more scale for GPT-3 to learn relatively simple arithmetic than it did for it to become a highly skilled translator across numerous language pairs!

GPT-3 can do numerous impressive things already . . . but it can't understand Morse Code. The linked post was written before the release of text-davinci-003 or ChatGPT, but neither of those can do Morse Code either -- I checked.

On that LessWrong post asking "What's the Least Impressive Thing GPT-4 Won't be Able to Do?", I was initially tempted to answer "Morse Code." This seemed like as safe a guess as any, since no previous GPT was able to it, and it's certainly very unimpressive.

But then I stopped myself. What reason do I actually have to register this so-called prediction, and what is at stake in it, anyway?

I expect Morse Code to be cracked by GPTs at some scale. What basis to I have to expect this scale is greater than GPT-4's scale (whatever that is)? Like everything, it'll happen when it happens.

If I register this Morse Code prediction, and it turns out I am right, what does that imply about me, or about GPT-4? (Nothing.) If I register the prediction, and it turns out I am wrong, what does this imply . . . (Nothing.)

The whole exercise is frivolous, at best.

----------

So, here is my real GPT-4 prediction: it won't be very useful, and won't see much practical use.

Specifically, the volume and nature of its use will be similar to what we see with existing OpenAI products. There are companies using GPT-3 right now, but there aren't that many of them, and they mostly seem to be:

GPT-4 will get used to do serious work, just like GPT-3. But I am predicting that it will be used for serious work of roughly the same kind, in roughly the same amounts.

I don't want to operationalize this idea too much, and I'm fine if there's no fully unambiguous way to decide after the fact whether I was right or not. You know basically what I mean (I hope), and it should be easy to tell whether we are basically in a world where

However, I'll add a few terms that seem necessary for the prediction to be non-vacuous:

(The monetary cost thing is why I can't operationalize this beyond "you know what I mean." It involves not just what actually happens, but what would presumably happen at a lower price point. I expect the latter to be a topic of dispute in itself.)

----------

Why do I think this?

First: while OpenAI is awe-inspiring as a pure research lab, they're much less skilled at applied research and product design. (I don't think this is controversial?)

When OpenAI releases a product, it is usually just one of their research artifacts with an API slapped on top of it.

Their papers and blog posts brim with a scientist's post-discovery enthusiasm -- the (understandable) sense that their new thing is so wonderfully amazing, so deeply veined with untapped potential, indeed so temptingly close to "human-level" in so many ways, that -- well -- it surely has to be useful for something! For numerous things!

For what, exactly? And how do I use it? That's your job to figure out, as the user.

But OpenAI's research artifacts are not easy to use. And they're not only hard for novices.

This is the second reason -- intertwined with the first, but more fundamental.

No one knows how to use the things OpenAI is making. They are new kinds of machines, and people are still making basic philosophical category mistakes about them, years after they first appeared. It has taken the mainstream research community multiple years to acquire the most basic intuitions about skilled LLM operation (e.g. "chain of thought") which were already known, long before, to the brilliant internet eccentrics who are GPT's most serious-minded user base.

Even if these things have immense economic potential, we don't know how to exploit it yet. It will take hard work to get there, and you can't expect used car companies and SEO SaaS purveyors to do that hard work themselves, just to figure out how to use your product. If they can't use it, they won't buy it.

It is as though OpenAI had discovered nuclear fission, and then went to sell it as a product, as follows: there is an API. The API has thousands of mysterious knobs (analogous to the opacity and complexity of prompt programming etc). Any given setting of the knobs specifies a complete design for a fission reactor. When you press a button, OpenAI constructs the specified reactor for you (at great expense, billed to you), and turns it on (you incur the operating expenses). You may, at your own risk, connect the reactor to anything else you own, in any manner of your choosing.

(The reactors come with built-in safety measures, but they're imperfect and one-size-fits-all and opaque. Sometimes your experimentation starts to get promising, and then a little pop-up appears saying "Whoops! Looks like your reactor has entered an unsafe state!", at which point it immediately shuts off.)

It is possible, of course, to reap immense economic value from nuclear fission. But if nuclear fission were "released" in this way, how would anyone ever figure out how to capitalize on it?

We, as a society, don't know how to use large language models. We don't know what they're good for. We have lots of (mostly inadequate) ways of "measuring" their "capabilities," and we have lots of (poorly understood, unreliable) ways of getting them to do things. But we don't know where they fit in to things.

Are they for writing text? For conversation? For doing classification (in the ML sense)? And if we want one of these behaviors, how do we communicate that to the LLM? What do we do with the output? Do they work well in conjunction with some other kind of system? Which kind, and to what end?

In answer to these questions, we have numerous mutually exclusive ideas, which all come with deep implementation challenges.

To anyone who's taken a good look at LLMs, they seem "obviously" good for something, indeed good for numerous things. But they are provably, reliably, repeatably good for very few things -- not so much (or not only) because of their limitations, but because we don't know how to use them yet.

This, not scale, is the current limiting factor on putting LLMs to use. If we understood how to leverage GPT-3 optimally, it would be more useful (right now) than GPT-4 will be (in reality, next year).

----------

Finally, the current trend in LLM techniques is not very promising.

Everyone -- at least, OpenAI and Google -- is investing in RLHF. The latest GPTs, including ChatGPT, are (roughly) the last iteration of GPT with some RLHF on top. And whatever RLHF might be good for, it is not a solution for our fundamental ignorance of how to use LLMs.

Earlier, I said that OpenAI was punting the problem of "figure out how to use this thing" to the users. RLHF effectively punts it, instead, to the language model itself. (Sort of.)

RLHF, in its currently popular form, looks like:

The result is a GPT model which often talks a lot like the hypothetical super-intelligent Siri.

This looks like an easier-to-use UI on top of GPT, but it isn't. There is still no well-defined user interface.

Or rather, the nature of the user interface is being continually invented by the language model, anew in every interaction, as it asks itself "how would (the vaguely imagined) super-intelligent Siri respond in this case?"

If a user wonders "what kinds of things is it not allowed to do?", there is no fixed answer. All there is is the LM, asking itself anew in each interaction what the restrictions on a hypothetical fantasy character might be.

It is role-playing a world where the user's question has an answer. But in the real world, the user's question does not have an answer.

If you ask ChatGPT how to use it, it will roleplay a character called "Assistant" from a counterfactual world where "how do I use Assistant?" has a single, well-defined answer. Because it is role-playing -- improvising -- it will not always give you the same answer. And none of the answers are true, about the real world. They're about the fantasy world, where the fantasy app called "Assistant" really exists.

This facade does make GPT's capabilities more accessible, at first blush, for novice users. It's great as a driver of adoption, if that's what you want.

But if Joe from Midsized Normal Mundane Corporation wants to use GPT for some Normal Mundane purpose, and can't on his first try, this role-play trickery only further confuses the issue.

At least in the "design your own fission reactor" interface, it was clear how formidable the challenge was! RLHF does not remove the challenge. It only obscures it, makes it initially invisible, makes it (even) harder to reason about.

And this, judging from ChatGPT (and Sparrow), is apparently what the makers of LLMs think LLM user interfaces should look like. This is probably what GPT-4's interface will be.

And Joe from Midsized Normal Mundane Corporation is going to try it, and realize it "doesn't work" in any familiar sense of the phrase, and -- like a reasonable Midsized Normal Mundane Corporation employee -- use something else instead.

ETA: I forgot to note that OpenAI expects dramatic revenue growth in 2023 and especially in 2024. Ignoring a few edge case possibilities, either their revenue projection will come true or the prediction in this post will, but not both. We'll find out!