The AI files: Making sense of artificial intelligence🍋

We don't know how AI will turn out. But that doesn't mean we know nothing

The news these days is dominated by artificial intelligence (AI), and predictions of its immediate and longer-term effects on people, occupations, countries, society and the human race.

Like many others, I have been trying to make sense of this torrent. The issues are complex, and many positions are overstated and often entrenched.

Some things about AI are likely truly unknowable. However, that does not mean that everything is unknowable. Humanity has a lot of experience with new technologies; and of the consequential economic and social disruption that is frequently predicted but rarely real.

This is the first of a series of posts exploring what we can reasonably know about AI, and what we cannot reasonably know. I’ll be writing from the perspective of an economist, albeit one originally trained in computer science and with a few decades of IT industry experience to draw on. Treat this post as a taster — raising issues to be explored in detail in later posts.

Ridges above Rangiwahia Hut, Ruahine Forest Park. Dave Heatley

Is artificial intelligence a useful term?

The term AI itself goes back to 1955, and for most of the intervening years has referred to techniques and technologies that, while they enjoyed varying degrees of success, generally fell well short of the ambitions attached to them. The AI that concerns us now dates back to around 2012. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton showed that a combination of software, hardware and large datasets that imitated some features of neuronal networks could significantly improve on previous attempts to solve some previously intractable classification problems.¹ Their work led to large language models (LLMs), the core technology in modern AI.

Imitation is a recurring feature in AI. LLMs, trained on enormous quantities of human-written text, have got better and better at answering a deceptively simple question: given the text so far, what comes next? The technology might be better described as “imitation intelligence” or, reflecting the era, “fake intelligence”.

More generally, if LLMs are very good at imitating humans, then how do we judge it? Is it intelligent, or just a clever fake? And if it is to improve beyond imitating humans, then where is the training data going to come from?

Most new technologies fizzle or offer incremental change …

As a first approximation, LLMs are just a new technology, built largely from well-understood components that were themselves new technologies years or decades ago.

Economists have spent a long time thinking about new technologies. Most are announced to the world with a lot of fanfare. Their potential, and their dangers, are over-hyped.

Many technologies flop, whether quickly (e.g. Segway, 3D television, Google Glass) or slowly (e.g. supersonic airliners). Some survive, often taking much longer to become widespread than initially expected (e.g. self-driving taxis) or serving a much narrower market (e.g. 3D printing, virtual reality).

Others change life as we know it

A few — just a tiny fraction — technologies end up having deep, long-lasting effects across a wide range of applications. Economists call these general-purpose technologies (GPTs; an acronym that unfortunately clashes with “generative pre-trained transformer” as in ChatGPT).² GPTs, such as steam power, electricity, digital computing, and internal-combustion engines, are obvious in retrospect but less clear at the time.

I think there is a better-than-even chance that AI will turn out to be a GPT.³ Time will tell. Regardless, I think there is much to be learned from experience with past GPTs that can be applied to the current situation.

The claims around AI are extreme — for good and for bad

Some of the claims surrounding AI are unusually extreme, and a poor fit with human history. For example:

• AI will cause mass unemployment

• AI will take control of all resources, making humans extinct or (at least) irrelevant

• AI will solve all human problems, offering (e.g.) boundless energy and vastly extended lives

Accepting any or all such claims requires us to make strong assumptions about states of the world and causality. These may (or may not) be reasonable, but it is crucial that they be clearly identified and, where possible, tested.

Policy to deal with extraordinary consequences

Governments are under intense pressure to “do something” about AI. But what? Policy analysis tools seem inadequate to the task, and the set of viable policy options may be small, and ineffective. Moreover, problems will be global, and international coordination mechanisms are at their weakest since the late 1940s.

On the policy analysis side, the cost–benefit calculus of extreme outcomes is straightforward:

• it’s worth accepting near-infinite costs to avoid certain extinction from AI; and

• it’s worth accepting near-infinite costs for the certainty of gaining utopia (i.e. unlimited benefits) from AI.

Things get much harder once we abandon certainty. We then need probabilities: how likely is extinction? How likely is utopia? And how much would any particular policy change those probabilities?

Enough serious people attach non-zero probabilities to one or both outcomes that it would be imprudent simply to rule them out. Which raises a difficult question for policy analysis: how should we deal with outcomes that are extraordinarily consequential but very difficult to assign probabilities to?

A snappy title

I was never a fan of the 1990s TV series The X files. But I did like the series’ tagline: “The truth is out there”. It manages to suggest simultaneously that the truth exists, that it may be strange, and that it may be beyond our reach.

Science fiction appears to inform opinions and policies about AI as much as anything else, so The AI files seems a suitable title for this series of Asymmetric Information (A↬I) posts.

Stay tuned.

By Dave Heatley

Te Hekenga, Ruahine Forest Park. Dave Heatley

1

Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet Classification with Deep Convolutional Neural Networks, Advances in Neural Information Processing Systems 25 (NIPS 2012).

2. Bresnahan, T. & Trajtenberg, M. (1995). General Purpose Technologies ‘Engines of Growth’?, Journal of Econometrics.

3. I’ll explain my rationale in a future post.