Creativity Isn't Real

A Problem I Would Love To Have

The most recent AI Skeptics podcast episode features Jacob Tsimerman, the current Fields Medal winner, who recently announced that he is joining OpenAI to work in their AI safety division.

I know I've mentioned the podcast several times before, and have referenced Cathy O'Neil's (one of the hosts of the podcast) book, Weapons of Math Destruction, so I won't spend too much time here explaining why you should listen to it.

Though you probably should.

One thing I'll say, however, is that I very much appreciate their willingness to engage with guests who may not hold completely congruent viewpoints. Sometimes there is overlap and commiserating, sure, but I also enjoy and learn from some of the discussions around the rough edges.

In the case of the Tsimerman episode, most of the episode is rough edges.

Here are some things that grated on me.

Language

In trying to explain his move from mathematics into the field of AI safety, Tsimerman attempts to explain how these mechanisms (i.e., LLM chatbots) seem to have moved into the realm of what he describes as "intuition."

(Aside: I tend to think the "intuition" label is trying to distance itself from the more problematic subject of "intelligence." I fail to see what distinction Tsimerman is making between the two concepts, or whether it's a willful euphemism meant to shed some of the historical baggage carried by the latter.)

From the time that AlphaGo defeated a human opponent in 2016 to when ChatGPT came out, his notion of what humans could do (within the realm of intuition) versus what computers can do now has been totally dismantled. He notes:

When [ChatGPT] started to talk, then I was forced to fully undo this notion of intuition versus [programming]... There's nothing humans can do now... I have no more intuition about a thing that humans can do and a reason that computers can't do that anymore. Because once you can talk, you can do anything... how should I say this? There is no task that humans are capable of... where I can introspect, and think, I am doing this in a fundamentally more complicated way than what I do when I speak. Speaking is as good as it gets... I'm only as smart as my ability to talk.

He goes on to conclude:

It became clear to me... that AI's will just become, if we're willing to make them such, robustly better than humans at everything.

If that is something you can believe, then the obvious thing to think about is the end of civilization, as he does in this paper titled A Taxonomy of Omnicidal Futures Involving Artificial Intelligence.

Smart

This is so bewildering to me.

Speaking is as good as it gets?

I'm only as smart as my ability to talk?

I am dumbfounded by this (to put it mildly).

Stochastic systems generate text. We know how this happens. It's complicated, but certainly not magical. Turning written language into mathematical vectors is not magic. Inferring which vector follows a sequence of vectors is merely a matter of computation.

(I emphasize "written" language here, purposefully glossing over things like image/stable diffusion. Mostly, my point is to illustrate how much is omitted when limiting language to what is contained within the written medium. When you consider body language, sarcasm, satire, double meaning, comedy, etc... there is a level of meta-language that the extruders can only mimic. And that's also discounting things like body language, inflection, tone, silence, emphasis, etc...)

The mimicry is alluring, but always false.

In order to imbue meaning into generated output, the reader must imagine a certain amount of intent.

I haven't seen an explication of this concept better than Emily Bender's, in which she compares LLM output to a Magic Eight ball. In an interview with Holly Baxter, Bender states:

Our human tendency to apply context to language and imagine that it’s coming from a personality primes us to believe in AI consciousness.

When you play with a Magic Eight ball, you shape your questions in ways that make the answers coherent. For example, you might ask "Will it rain today?" The ball might answer with "Without a doubt." But if you were to ask, "How much will it rain today?" then the answer "Without a doubt" makes less sense.

Imagine you took a Magic Eight ball and instead of a 16-sided die or whatever inside of it, you made it really big. So you could have a 256-sided die. And then you filled up a football field with those. Is that any closer to consciousness than the one little one that I have in my hand here?

It boggles my humanities-laden mind that a Fields Medal-winning mathematician is so utterly vexed by text output from a Magic Billion Parameter Inference ball.

"I'm only as smart as my ability to talk," Tsimerman says.

Where is the disconnect? This isn't even touching on whether the inference systems are conscious or not. It just seems like a total sidelining of an entire disciplines of thought, such as Philosophy or Linguistics.

This belief that Tsimerman has—that language (speaking?) is the culmination of what humans can do—this completely overwhelms me.

Language is hardly the culmination of human complexity.

The mere process of writing is extremely difficult for people specifically because language, as great as it is, cannot possibly capture the nuances of wisdom, love, passion, anxiety, fear...

Poets, philosophers, authors, and erstwhile bloggers such as myself wrestle with language, constantly trying to capture the evasive Truth that seems to always be within reach, but impossible to capture.

One could say, that is why Art exists to begin with...

Creativity

Which brings me to this other startling thing that Tsimerman says.

Is AI creative, or is it just doing x, y, or z?... I basically don't think creativity is real. In the sense that I don't think it's an atomic unit in the way we treat it...

Uh huh.

...the fact that it is so unquantifiable... means that it's very easy to fall into the shifting targets trap.

He says this in response to pushback on his assertion that language (syntax generation) is the end all of intelligence, creativity, wisdom, etc...

I can almost hear in his response (reading between the lines, if you will), I have prepared for this.

The focus on consciousness and the focus on our human conception of human wisdom and creativity is, I think, very harmful to any discourse trying to operationalize intelligence.

Uh huh.

He continues by speaking of intelligence as something you can break down into a set of tasks. Given a goal, can the system achieve that goal? (He cites some examples of how certain tasks, such as solving math proofs, have been successfully completed when previously thought impossible.)

He concludes this particular train of thought with this:

What task now can AI not do concretely, that we can actually judge, that humans can do, that you would declare to be genuine understanding, or creativity, or reasoning? And if there is no such task, then in terms of discussing the future, what value does that concept hold?

What task can AI not do concretely?

If there is no such task...

Hmm. How do you love concretely? How do you pass on wisdom to your children... concretely? How do you weigh the value of a human life in the midst of a societal and political dysfunctional system such as what we are living in today?

Good grief.

In some sense, I do believe he's right about creativity being unquantifiable. And I even agree with his sense that there's no way to operationalize non-deterministic software in a way that would satisfy some preconceived concept of creativity.

Creativity is fire. It goes against the grain. It provokes.

Non-deterministic software that produces output based on an amalgamation and synthesis based on semantic similarity can never do that.

And how about morality? How about law or ethics? What of justice or inequality?

Imagine these systems reflecting a form of morality or ethics that is based primarily on an operationalized system, which, I might add, reflects the biases prevalent in the training data.

That is not progress. It is regression (by definition).

Accountability

Which leads to this final topic.

Yes, things are hard to measure and hard to define, as Tsimerman asserts.

But there are things that are not—such as the very real and very easy to see harms that are being produced by detestable humans, concealed beneath the cover of a "booming" AI industry. (And by "booming" I mean "exploitative" and "overvalued".)

After Tsimerman broadly explains his role in "AI Safety," O'Neil deftly shifts to the topic of accountability.

"A lot of the AI systems are, basically, extensions of the values of their deployers and their owners, and they do not agree with the values of the people that are targeted by them or that otherwise are harmed by them," she says.

She goes on to explain that the problem is not of alignment between computers and humans, but rather of how powerful systems work against other humans.

She continues, "There's always a human accountable to every harm... I'm wondering, where that sort of lens lives in your perspective of AI safety. Who gets in trouble when something bad happens?"

His answer is a bit remarkable, for all the wrong reasons.

I think what you are pointing at is a super, real and important problem that I would love to have. In the sense that, I don't think we are at the point where we have to solve the very difficult problem of AI's are doing what their owners want and now we have like a social, political, enforcement problem of let's get the humans to cooperate so that their AI's cooperate.

His reasoning for this is because he claims that the bigger, more urgent problem is to contain the AI systems that have been unleashed and cannot be controlled. This is somehow worse than, you know, holding companies accountable for the harms that are already actively being performed.

I think it's moot at the moment because computers are going to pursue unaligned goals that nobody can actually specify and that naturally leads to them taking over or otherwise causing problems.

Sigh.

Trolley

I was intrigued when news broke about the Fields Medal winner who would be moving from academia to Silicon Valley, and working in "AI" no less.

After listening to the podcast, I can understand that he is indeed, a very intelligent person.

But also, I learned that he is also just another person with the capacity to be wrong.

Tsimerman has somehow imagined an adaptation of the trolley problem where on one side, there are vulnerable people who will soon be demolished. In fact, there is a trail of blood proving that this is already happening. You could even operationalize a set of tasks that proves this is so.

Instead, he chooses to focus on a divergent track that doesn't exist. Furthermore, he imagines that the calamity of going down that particular track (which doesn't exist) could be an existential threat to humanity.

The track doesn't exist.

And even if it did, it would be so, so, so very easy to avoid an apocalyptic, humanity-ending calamity at the six-fingered hands of rouge "AI" agents.

You know, just stop the train.

Doesn't take a Fields Medal to come up with that one.