Two reflections on AI

I

[While you’re here, please sign the petition to save my friend.]

Recent developments in AI, for example, language models are an event- an outside element that disrupts equilibrium. No one expected these developments ten years ago. No one ever thought that we’d have machines that could do this:

---

Describe a war between sapient cows and vicious, semi-sapient pizzas

Title: The Battle of Moos and Crusts

Background:

In a distant universe, two unusual species evolved into prominence. On the rolling hills of the planet Lactopia, the Bovinae sapiens, commonly known as the "Sapient Cows," developed advanced civilizations, founded on their agricultural prowess and ability to harness the power of the sun for energy. Meanwhile, on the fiery planet Crustacea, semi-sapient pizzas evolved. These pizzas, with molten cheese and toppings that had developed into aggressive weaponry, thrived in extreme temperatures and had a predatory nature.

The Conflict Begins:

For millennia, the two species remained unaware of each other's existence. However, as their technology advanced, so did their exploratory ambitions. A scouting ship from Crustacea crash-landed on Lactopia, igniting the initial conflict. The pizzas, driven by their insatiable hunger and predatory instincts, began attacking the Sapient Cows' settlements.

Major Battles:

1. The Siege of Dairyhold: At Dairyhold, a major Sapient Cow city, the pizzas deployed their elite Pepperoni Phalanx, using their scalding hot cheese as a weapon. However, the cows, with their organized defenses and milk cannons, managed to fend off the attackers after a long battle.

2. The Ambush at Wheat Fields: A herd of Bovinae sapiens set a trap in the vast wheat fields, luring in the pizza forces. When the pizzas attacked, the cows used their sun-harnessed weapons, focusing intense beams of light to cook and immobilize the pizzas.

3. The Crustacea Counterstrike: Using their technology, the pizzas developed the "Dough Driller," a massive contraption that could tunnel through Lactopia's surface. They launched surprise underground attacks, causing chaos and devastation in several cow settlements.

Turning Point:

Realizing the extent of the pizzas' threat, the Sapient Cows formed the "Moo Alliance," uniting all the cow cities. They developed "The Great Cheese Neutralizer," a device that could nullify the pizzas' molten cheese weaponry. With this, the tide began to turn in favor of the cows.

End of War:

After years of intense battles and heavy losses on both sides, a faction of vegetarian pizzas, tired of the violence, approached the Sapient Cows with a peace offering. They proposed an alliance, sharing their knowledge of temperature control to help the cows counteract the aggressive pizzas. With this newfound advantage and the internal division among the pizzas, the Sapient Cows successfully pushed the invaders back to Crustacea.

Aftermath:

With the war over, Lactopia underwent a time of healing and rebuilding. The vegetarian pizzas that had aided the cows were granted sanctuary on Lactopia and coexisted peacefully. Both species established an interplanetary council, ensuring that such a conflict would never arise again.

Legacy:

The Battle of Moos and Crusts became a legendary tale in the annals of Lactopia, serving as a cautionary tale about the perils of unchecked aggression and the importance of unity and understanding.

---

All without being able to walk, learn new things beyond a 6000 token limit, or, until recently, see anything. We thought basic interaction with the world and continuous learning would come long before coherent, and even, in a limited sense, creative, writing. It is, of course, not especially good fiction, but if that’s what you’re looking for, if you think that’s the splitting line that would make all the difference, you're losing the plot for the paragraphs.

The form AI has come to us in is from outside the range of our experiences, our theories, and our understandings- even as it intimately reflects features of the world as it is. AI has slammed into us. In light of that, what we need is a little bit of curiosity about these new technologies. Instead, what we have is a surplus of people who seem:

1. To be wholly uninterested in the technology, except insofar as it relates to their existing hobby horses.

2. To be working on an assumption that the technology will be static, that it will not get drastically better from here. But it’s only just been rolled out a consumer-grade, of course, it’s going to get much better!

Not just curiosity, but a little humility for humanity’s first encounter with an alien (simulated) intelligence capable of flexible language use would be prudent. But humility seems to be very rare indeed, especially in an age where sneering dismissive confidence is algorithmically prioritized, along with equally misplaced rapturous applause.

There are of course numerous people who are taking this technology very seriously. Boosters on Twitter who are writing threads like “Ten GAME CHANGING uses of AI (number 7 will shock you!), but though they are enraptured by the technology, they are not intellectually serious people. Most of the serious people have given over taking AI seriously and on its own terms to the cranks.

Auto-complete on steroids, stochastic parrots- these are thought terminating clichés, ironically quite hypocritical epithets in that regard. The super-autocomplete metaphor has some value, but it has limits as well. It tells us what language models do, in a sense (although this is complicated a little by RLHF training), but it doesn’t tell us how they do it. Increasing evidence, some of it pretty high quality, suggests language models are implicitly models of the world- a cute little toy example is here. By contrast, the stochastic parrots model is useless. It has no predictive utility and people who rely on it have been consistently blindsided by developments. It’s an everlasting thought-stopper.

An example of the kind of overconfidence I’m talking about:

Sebastian Budgen, editor of Verso books writes in praise of Eye of the Master, a book about AI:

Every day, it seems, we are assaulted by articles about AI, each more idiotic than the previous, whether in the form of the chittering of economic journalists or the pseudo-philosophical sub-Netflix chuntering regarding the “singularity” and other impending scenarios of doom for the human race. It is remarkable to see the abstract concepts of the Western Marxist tradition such as “alienation”, “reification”, “commodity fetishism”, “determinism” and so on come alive and skitter across the page or screen when this subject is discussed.

Drawing from the well of Italian operaismo – particularly figures such as Romano Alquati, Raniero Panzieri (as well as thinkers such as Alfred Sohn-Rethel) – which sought to enact a proletarian Copernican revolution in the understanding of the relationship between technology and the labour process, capital and the working class, Matteo Pasquinelli, in his new book The Eye of the Master: A Social History of Artificial Intelligence, offers us welcome relief. Professor of media philosophy at the University of Arts and Design Karlsruhe (HFG), but currently based at Ca’ Foscari University of Venice (and introduced to me via McKenzie Wark, who has a sharp eye for promising scholarship), Pasquinelli offers us in this work a fascinating new, revisionist genealogy of AI, which ranges from Adam Smith, Charles Babbage, the English Ricardian socialists Thomas Hodgskin and William Thompson, and Karl Marx (the famous, much chewed-over lines in the Grundrisse, in particular) to Friedrich von Hayek, via the what to me are the unfamiliar debates regarding the self-organisation of organisms and machine, artificial neural networks, the Gestalt controversy and the “perceptron”. But, more than this, he informs his account with a “labour theory of automation” that reconnects this history of AI to struggles over the measurement, discipline, and intensification of work, from Taylor onwards.

Now I’m not going to say anything critical about Eye of the Master, which I have not read, and which sounds quite promising in some ways. Rather I want to say something critical about Budgen’s comments here. If one reads between the lines, the angle is obvious “Don’t worry, Marxist theory can subdue the new AI monster. You’re not going to have to fundamentally revise your web of beliefs. The more exotic claims of a singularity certainly can be laughed out of the room (if with a little nervousness). In fact, AI just makes everything like I was already saying it was but more so.”

Freddie de Boer, usually a sharp critic, is trying to run a line roughly to the effect that it’s basically like Eliza and won’t amount to something because something something Chomsky, something something Stochastic Parrots, something something, every age believes itself to be living in the end times. Chapo are refusing to think about it on the grounds that there’s some crossover in it’s fan base with Bitcoin. The linguists in particular really seem in a tiz, which reminds me of that famous quote about natural language processing “Every time I fire a linguist, the performance goes up”.

A bunch of scholars are talking about it as if the only issues or main issues it raises were algorithmic bias. It certainly does raise issues of algorithmic bias, and I truly think these are deeply important, but I strongly suspect it’s potential capacity to replace 95% of tasks which can be done by exchanging words, and eventually a lot more tasks than that, might matter more. The right is banging the table about something something wokeness because they can’t get ChatGPT to say the N word. Even the AI focused- rationalists and the like- aren’t really grappling with the technology. No one in the “AI safety” space predicted the AI would be like this 10 years ago, and the work of revising beliefs and plans in light of that is an unfinished task.

You can’t more-cynical-than-thou-cultural-criticism your way through this one. No one knows what’s going to happen next, an event has barged through the door. Things could get better, worse, or remain about the same but go far sideways. They will not remain unchanged.

A friend of mine, broadly in the historical materialist school of thinking, once said to me that he really resented AI because it’s another thing he has to think about and fundamentally reframe his social thinking in terms of. In an age where civilization is under threat of barbaric violence induced by climate change disasters and scarcity, it’s just too much to sift through. Well, I’m sorry, but this is it now. You’re going to have to spend the rest of your life engaging with it. I don’t know what’s going to happen- there are a lot of things that could- but the variance of the next 10-20 years has increased greatly, and it won’t go away after that either.

II

Imagine you just wrote out whatever popped into your head one word at a time, no going back, no correction, any ‘planning ahead’ purely implicit and pattern-based, not explicit. You wouldn’t do so well! This is especially true of working with complex subjects.

The HumanEval task is a programming task for language models. I don’t have the tools to evaluate how good a benchmark it is for that purpose, for our goals though that won’t matter much. All that matters is that it’s a difficult reasoning test.

GPT-4, raw, the version you can play with online if you pay OpenAI for it, gets 67% on the task, GPT-3.5, the free LLM you can play with if you sign up gets 48% in it.

Now, add LATS. LATS is a complex system, one I don’t fully understand, but I think a fair metaphor is that it involves the model thinking through what it’s going to do, rather than just coming up with one plausible word at a time. A bunch of work has been going on to try to make language models do this, starting with chain of thoughts and only getting more complex from there. Think of it as a bit like thinking through possible chains of moves in chess, plus a process of self-guided experimentation. I have included the LATS abstract here:

While large language models (LLMs) have demonstrated impressive performance on a range of decision-making tasks, they rely on simple acting processes and fall short of broad deployment as autonomous agents. We introduce LATS (Language Agent Tree Search), a general framework that synergizes the capabilities of LLMs in planning, acting, and reasoning. Drawing inspiration from Monte Carlo tree search in model-based reinforcement learning, LATS employs LLMs as agents, value functions, and optimizers, repurposing their latent strengths for enhanced decision-making. What is crucial in this method is the use of an environment for external feedback, which offers a more deliberate and adaptive problem-solving mechanism that moves beyond the limitations of existing techniques. Our experimental evaluation across diverse domains, such as programming, HotPotQA, and WebShop, illustrates the applicability of LATS for both reasoning and acting. In particular, LATS achieves 94.4% for programming on HumanEval with GPT-4 and an average score of 75.9 for web browsing on WebShop with GPT-3.5, demonstrating the effectiveness and generality of our method.

LATS takes GPT-4 to 94%, and takes GPT-3.5 to 87%. Notice that LATS makes GPT-3.5 much better than GPT-4 vanilla.

And this kind of plan-ahead stuff is only one area of expansion. Even if, from this point onwards, there were no further improvements to language models whatsoever, the low-hanging fruit that remains in improvements external to the models themselves means that current language models are nowhere near their full power but there will continue to be improvements to the models themselves.

Personally, I would be fascinated to see this kind of iterative mechanism applied to problems like making creative models. A lot of people think that language models can’t be genuinely creative because they simply recombine what already exists- but I don’t think this is the major problem- with language, we all, in a sense, simply recombine what exists. I believe the big barrier to creativity at the moment is a lack of iteration. Creative thinking necessarily involves an evolutionary process of somewhat random generation, testing and selection, I think. I’d be terrified too.

It would be very foolish to bet against language models when there are literally dozens of easy-ish avenues of improvement that are already great promise and also, just making the damn things bigger isn’t over yet.