AI will upend your world

For the better. Ignore the pearl clutchers

Chris Johns

Nobody knows anything.... (William Golding)

I used to think I was tech-savvy. Silly me. The title of this piece is carefully chosen: I am the idiot. That’s meant in all seriousness. While I consider myself to be a proud early adopter of all things computing, I am a total neophyte when it comes to awareness and understanding of AI. But, I do have some tech skills. At least I thought I did.

I was a teenager when pocket calculators were cool and in the sixth form when a terminal appeared linking my school to the local council’s mainframe computer. We mucked around with coding, something called, for a reason, BASIC. I spent a few months between school and college at a computer centre that, 24/7, 365 days a year, produced only telephone bills. It was all punch tape, 80-column cards and a few mag tape drives all interacting with a LEO 326 computer with hardly any internal memory capacity. Not a disc drive to be seen. LEO stood for ‘Lyons Electronic Office’ a machine built for a now extinct chain of tea-houses.

At college a mainframe enabled linear programming and econometric research. As a graduate student I learned a now extinct language, FORTRAN, on an ICL terminal linked to a DAP (Distributed Array Processor). We solved and simulated models using algorithms (Gauss-Seidel in case you are interested). At a UK research institute, we ran economic models via an expensive telephone link to a large ‘computer bureau’ in Kansas. We were thrilled in ways only fellow geeks can imagine when a standalone ‘mini-computer’, now also extinct, appeared in the office. It was called a Wang.

Next, at the UK Treasury, I worked both with a mainframe and the new Personal Computer (IBM PC). We used Lotus-123 and word processing programs that would look like clay tablets to today’s users of Word or Excel. On to the City in the 1980s where computer skills were fast becoming a valued commodity. Often, the uber-nerds would know more about tech than the failed-at-everything-else occupants of the IT ‘Help Desk’. When everyone came to us looking for answers we would answer, as only arrogant youth can: ‘RTFM’ (read the fucking manual). And so on to the present day where I am typing this on a more-or-less state of the art Apple iMAc.

All this credentialism is not to prove how much I know about (old) technology. It’s to emphasise the sense of astonishment a self-described semi-expert feels when exploring all the new toys being developed daily by the Artificial Intelligence industry. The Economist calls it ‘intellectual vertigo’. That’s well put.

How can we organise our thinking, form views, about something that looks like science fiction and improves at light speed daily?

Here are some announcements and stories - just a selection out of many - from just the past few days:

Another:

Another:

Another:

Another:

I could go on.

How do they do it? What, precisely, is ‘it’? What does AI portend for the economy? For politics? For jobs? For the military? And a thousand other questions. Things are moving so quickly. Everyone seems to have a different answer to those questions.

I’ve read learned articles that describe AI as ‘just programming’. Others argue that AI represents an existential risk to our species. ‘It’s gonna be huge, as significant as the discovery of electricity, the transistor and beer.’ Alternatively, ‘it’s a nothingburger’.

Conspiracy theorists and other nut jobs have long posited something called the Great Replacement - its a racist thing. Maybe these wackadoodles were right for the wrong reasons: we are all about to be replaced. By machines, not other humans. It’s a jobs apocalypse!

Or just maybe we are on the verge of super-abundance thanks to massive increases in productivity, the cure for cancer and the explanation of how Liz Truss ever became Prime Minister of the UK.

There are many guides and explainers. This ‘A-Z’ guide to AI is a good primer - read it and you will acquire more knowledge about AI than everyone else at your next drinks party.

The Opportunity

Here’s The Economist:

ChatGPT embodies more knowledge than any human has ever known. It can converse cogently about mineral extraction in Papua New Guinea, or about TSMC, a Taiwanese semiconductor firm that finds itself in the geopolitical crosshairs. GPT-4, the artificial neural network which powers ChatGPT, has aced exams that serve as gateways for people to enter careers in law and medicine in America. It can generate songs, poems and essays. Other “generative ai” models can churn out digital photos, drawings and animations.

That’s about as succinct as it gets. If that doesn’t fire up the imagination about what these tools can do, nothing will. Moreover, the creators of these tools are muttering already that they are not sure about the ‘how’: some of the answers generated by these bots have baffled their creators. They don’t know how the machine got there. And, therefore, what the machine might be able to do next. Which leads us on to:

The Threats

Much has been made of the threats from AI. Indeed, ‘Alignment’ is the issue of the day. From that A-Z guide:

The “alignment problem” is one of the most profound long-term safety challenges in AI. Today’s AI is not capable of overpowering its designers. But one day, many researchers expect, it might be. In that world, current ways of training AIs might result in them harming humanity, whether in pursuit of arbitrary goals, or as part of an explicit strategy to seek power at our expense. To reduce the risk, some researchers are working on “aligning” AI to human values. But this problem is difficult, unsolved, and not even fully-understood. Many critics say the work to solve it is taking a back seat as business incentives lure the leading AI labs toward pouring focus and computing power into making their AIs more capable.

A gift to scammers, a problem for privacy, election manipulation will explode, we are all going to die.

Everything that was said about the printing press is being said about AI. Consider this bit of pearl-clutching from a learned Professor writing in the Irish Times:

[Chat GPT3] was able to find somewhere to make chemical weapons, write hate speech and find unlicensed guns online. Based on this evaluation, OpenAI amended GPT-4 before release. But this is just one limited set of problems. Clearly a huge range of as yet-undiscovered negative consequences remains.

The author has clearly never used Twitter, Facebook or Google if he thinks that we can’t do any of these things already. Maybe he means that wannabe criminals actors may now be able to break bad at warp speed. With that, at least, I have some sympathy. Running scared from undiscovered threats is not something anyone should take seriously. But being able to do awful things, quicker and better than before is something to be worried about.

Noah Smith explores here the existential (and other) threats from Chatbots and concludes, convincingly in my view, there is nothing much there.

That’s not to say there are no threats at all. For example, as with computer gaming, ‘Jailbreaks’ of ChatBots have become a thing: self-taught users of AI are figuring out ways to get round internal ethical and safety controls. Here is someone fooling a bot into giving instructions on how to make Napalm:

I haven’t tried it, but a Google search could, I would guess, come up with similar results, albeit more slowly and without having to pretend to be enquiring on behalf of your deceased granny.

Here is a recent Bloomberg story about, supposedly, Google’s ethical lapses in its race to catch up with Microsoft:

Google’s Rush to Win in AI Led to Ethical Lapses, Employees Say

The search giant is making compromises on misinformation and other harms in order to catch up with ChatGPT, workers say

That race with Microsoft is one example of how businesses, not just individuals, are being turned upside down. From being utterly dominant, AI risks doing to Google what Google did to thousands of other businesses.

It’s not all bad news. And here’s the thing: there are as many surprising good news stories as there are concerns about ethics, safety, privacy and an extinction level event for our species:

AI Weed-Killing Drones Are Coming for the Mega Farms

This Canadian startup has developed an AI-enabled crop-spraying drone that reduces chemical use by up to 90%. 

Another story, also from Bloomberg, that contains a mixture of good and not-so-good stuff:

I’m going to write a lot about AI, today and in the future. Just summarising what everyone else has written about Bard, Sydney, ChatGPT3 and 4 would fill a large (old-fashioned) library. So some of this is inevitably idiosyncratic (what interests me, what seems to be most significant to me). I will not (yet) explore the technical aspects of the ChatBots - machine leaning, training, Proximal Policy Optimisation and all the rest of it. But one technical point should be made.

GPT-4 is an example of how quickly things are moving. It is a massive upgrade of GPT-3.5, itself only released a few months ago. GPT is a based on statistical modelling that aims to predict, given a word or a letter in a user ‘prompt’ (usually a question), what comes next. The basic unit is called a token, roughly analagous to a word. GPT-3.5 could handle 2048 tokens in a single prompt, GPT-4 can be prompted with 32,000 tokens. That exponential increase in power comes at a cost, in terms of computing power, electricity and money. From The Economist:

But there are good reasons… to think that this growth cannot continue indefinitely. The inputs of LLMs—data, computing power, electricity, skilled labour—cost money. Training GPT-3, for example, used 1.3 gigawatt-hours of electricity (enough to power 121 homes in America for a year), and cost Openai an estimated $4.6m. GPT-4, which is a much larger model, will have cost disproportionately more (in the realm of $100m) to train. Since computing-power requirements scale up dramatically faster than the input data, training LLMs gets expensive faster than it gets better. Indeed, Sam Altman, the boss of Openai, seems to think an inflection point has already arrived. On April 13th he told an audience at the Massachusetts Institute of Technology: “I think we’re at the end of the era where it’s going to be these, like, giant, giant models. We’ll make them better in other ways.”

Predictions about the future of tech have often been notoriously wrong. But if this line of thinking is correct, there are reasons to think the rapid improvements in AI may face constraints going forward.

Finally, the ‘we don’t know’ threat (again). Everyone seems to worry about AI because we don’t understand it. I’m tempted to say I don’t care that I don’t know how my IPhone or car works, but the obvious riposte is that ‘somebody does’, probably the person that built it. Today, the AI fear is that nobody understands it. When GPT-5 comes out, I plan to ask it to explain itself to me. Meantime, all of this represents my attempts to understand.

First question: What is intelligence?

Lots of people grapple with the question of intelligence: what is it exactly? Lots of theories, but, paradoxically when you think about it, we don’t seem to know. It’s therefore somewhat moot to ponder the nature of artificial intelligence if we can’t agree on what intelligence actually is.

Some scholars think we need to ask: ‘what makes us human?’ And, relatedly, ‘what is consciousness?’

Does AI deserve the I? Whether or not we can define human intelligence (see question 1, above) plenty of deep thinkers reckon that ChatBots don’t - and will not - come anywhere close to replicating whatever it is that we do. Here’s the ‘father of modern linguists’ Noam Chomsky (alongside another linguist and a philosopher/AI specialist:

Note, for all the seemingly sophisticated thought and language, the moral indifference born of unintelligence. Here, ChatGPT exhibits something like the banality of evil: plagiarism and apathy and obviation. It summarizes the standard arguments in the literature by a kind of super-autocomplete, refuses to take a stand on anything, pleads not merely ignorance but lack of intelligence and ultimately offers a “just following orders” defense, shifting responsibility to its creators.

In short, ChatGPT and its brethren are constitutionally unable to balance creativity with constraint. They either overgenerate (producing both truths and falsehoods, endorsing ethical and unethical decisions alike) or undergenerate (exhibiting noncommitment to any decisions and indifference to consequences). Given the amorality, faux science and linguistic incompetence of these systems, we can only laugh or cry at their popularity.

Chomsky’s point about getting things wrong is well made. I’ve a long list of things I’ve asked AI that it has got wrong. I’ve even asked “do you make stuff up’ and been told, unambiguously, ‘yes’.

Others have a very different view. One of America’s most prominent public intellectuals, Professor Tyler Cowen, sees it this way:

I take great care to spell out just how special recent times have been, for most Americans at least.  For my entire life, and a bit more, there have been two essential features of the basic landscape:

1. American hegemony over much of the world, and relative physical safety for Americans.

2. An absence of truly radical technological change.

Unless you are very old, old enough to have taken in some of WWII, or were drafted into Korea or Vietnam, probably those features describe your entire life as well.

In other words, virtually all of us have been living in a bubble “outside of history.”

Now, circa 2023, at least one of those assumptions is going to unravel, namely #2.  AI represents a truly major, transformational technological advance. 

Cowen argues that most of us haven’t experienced massive technological change. As Peter Thiel famously said, tech promised us flying cars and all we got was Twitter. We have great toys - this iMac, my iPhone etc - but they are, in the way that most of us use them, just toys.

Cowen also argues, disagreeing with Chomsky et al, that all of this is about to change in fundamental ways. and in ways that none of us have experienced before. (He’s also worried about his first point, a topic for another discussion.) I’m with Cowen, AI is going to be massive.

Plenty of others disagree. The physicist Roger Penrose argues that the structure of the human mind is beyond any computer program to emulate. Chat GPT3 summarises Penrose argument in this way:

[Penrose says] consciousness is non-algorithmic, and humans have access to non-computable mathematical truths that machines cannot access. He proposes that these non-computable mathematical truths are essential for human understanding and creativity and that they cannot be replicated by AI.

Yes, I’ve asked the various AI ChatBots my list of questions: the answers are less than illuminating. At least AI has given philosophers and linguists something to do. Maybe that’s where the (only?) jobs of the future lie

Second question: how will the gains from AI be distributed?

On the economic consequences of AI, read this. David Autor looks at research that reckons 80% of jobs are going to at least affected, if not abolished, by AI. He also considers the possibility that we will all be ‘upskilled’ if we learn to use AI properly. Another of those two-sided stories about AI.

Noah Smith is a relatively lonely voice arguing that the threat to jobs is exaggerated. As with most things, I’m with Noah on this one. First, conventional wisdom:

Lots of research reports from entities like investment banks and universities say things like ‘57% of jobs in OECD countries are at risk from AI’. That number varies - a lot. 20% to 80% seems, to my eye, to be the range. These warnings emerged a long time before we heard of ChatGPT. Previously, the buzzword was not AI but ‘automation’. As Smith correctly notes, in the years since these jobs apocalypse warnings began to emerge, the number of people at work in many economies has gone up to the point where labour markets are beyond tight.

The subtle point is the one described by Autor: it seems uncontroversial to assert that parts of people’s jobs are going to be automated or simply made easier by AI. Tasks are automated, jobs not so much. That’s upskilling, that’s a potential productivity boom. History teaches us another important point: new jobs, previously unimaginable, will be created.

Some jobs will disappear, we simply do not know how many. A jobs boom is at least as likely as a jobs apocalypse. More likely in my view. Skilled ‘prompters’,a brand new job, can command six figure salaries right now.

If AI elicits an economic boom, the issue will be how the gains from abundance are distributed. Piketty and others have demonstrated that economic gains in recent decades have accrued to a small slice of the population: a gently rising tide has only lifted a few boats. Will the owners of AI, all three or four of them, take all of the productivity gains? Will AI usher in the final denouement of the winner-takes-all economy? And thereby lead to a revolution of a very non-technological kind - will we finally find our pitchforks?

Some of the economic questions that have already been asked are as profound as they are unexpected. Here is one discussion of the taxation of AI, making a very simple point:

The Taxman Will Eventually Come for AI

The taxman, as Al Capone would attest, eventually sorts everything out.

I’ve tried to argue that there is, indeed, much that we don’t know. I’ve taken an essentially optimistic stance, based on how our economies have evolved in the past when facing rapid technological change. Fear of the unknown is, however, pointless and verging on irrational. So is paying much attention to precise forecasts of the future. My imprecise forecast is this: our lives will be changed radically by AI, but mostly in ways for the better.