Will AI just keep getting better?

What if it doesn't?

Last Monday I was on Radio 4’s Today programme speaking to Mishal Husain about how we do - and don't - use artificial intelligence at No More Marking (from 2hr39m30s here).

I explained how we successfully use AI in some of our back-office functions, but that our experiments using the Large Language Models (LLMs) GPT-3 and GPT-4 to mark students' writing had been pretty underwhelming.

Mishal Husain then said: "of course [LLMs] might well, and indeed should, get better the more it’s used."

A year or so ago, I would have totally agreed with this.  But now, I am not so sure. There are genuine debates about whether LLMs will keep improving. Here’s a summary of the arguments.

LLMs will inevitably get better

Computing power keeps increasing, you can train LLMs on more and more data, and if you do this they will improve. In fact, if you do this they might even improve in ways you don’t expect: one of the interesting debates about LLMs is whether as they get bigger they display ‘emergent’ properties that cannot be predicted from smaller models.

If this is the case, LLMs could improve in ways that cannot be predicted or extrapolated from current models. Quantity has a quality all of its own.

LLMs will not get better

But there are others - for example, Gary Marcus - who would argue that the models underlying LLMs are fundamentally flawed. They need a serious qualitative fix, and throwing more training data or computing power at them won't solve the instability and hallucinations which plague them at the moment. The metaphor that's very common here is that LLMs are like growing a tree to get to the moon. You might see some very rapid growth in the early stages which encourages you and means you plot a trajectory that shows you reaching the moon in x days. But of course eventually those predictions will break down, because a tree can't get you to the moon.

If this is the case, LLMs could regress in ways that cannot be predicted or extrapolated from current models. Quantity cannot make up for a fundamental underlying lack of quality.

So who is right?

Like I say, a year ago, I would have subscribed to the techno-optimist case: of course they will improve, technology keeps getting better, the tech you are using right now is the worst you will ever use, etc.

But seeing some of the the egregious errors that GPT makes and reading more about the theory behind it has made me doubt this optimism, and it has even made me question techno-optimism in other areas. Not all new technologies do succeed. Many of those that do take longer to get there than first anticipated. And some new technologies go through a process of rapid improvement followed by diminishing returns.

Ultimately, I remain a techno-optimist. But I am less certain than I was. 'But it will only get better' feels much more like a faith-based assumption to me than it used to. Sometimes, I think that our society is as saturated with the assumptions of technological progress as medieval times were saturated with the assumptions of the resurrection. It’s incredibly hard for anyone to think outside the paradigm of ‘technology will always improve’. Even self-styled techno-pessimists are often not actually as pessimistic as they seem. For example, I recently read an interview with John Gray, who is about as pessimistic about the idea of progress as anyone in modern Britain. Did he say 'I don't think LLMs are going to get any better?' No. He said 'LLMs are going to destroy society by taking all the jobs'. Ultimately, for all its surface pessimism, that's still an optimistic take about the power of technology.

Likewise, much of the AI doom talk is predicated on a belief in technological progress. To go back to my religious comparison, most of the AI debate is about whether we’re going to heaven or hell. No-one is really asking if heaven or hell exist. And yet, it seems worth asking that question. The thing I'm really worried about is not that AI will succeed so well that it will free us from labour. It's that so many of the massive problems we face as a society cannot be solved without exponential technological progress. If the progress stops, we're all in trouble.

Is our belief in technological progress a help or a hindrance?

The fact that so many people believe in technological progress is a good thing. It encourages people to build careers and to persist when faced with challenges and setbacks. In some ways, a belief in technological progress can become a virtuous self-fulfilling prophecy. Our belief that solutions are ‘out there’ helps will them into existence.

But it is also worthwhile considering the downsides of optimism. When does faith become blind faith? When does it lead to complacency, to the assumption that something will turn up because it always does?

Failures of technology

I would say that if you are a techno-optimist, you should be especially concerned with the historical examples of technology failing, or plateauing, or not quite working out - because by being aware of them you can avoid those pitfalls.

For example, scientists have been researching fusion power for the best part of a century, but so far no-one’s worked out a way it can produce energy at scale. The speed of air travel increased rapidly in the early years of the technology, but has plateaued (and arguably gone into reverse with the decommissioning of Concorde). Antibiotics are perhaps similar - rapid progress, plateaus, potential reversal.

Even when technologies do succeed, if they take a lot longer to do so than anticipated, that causes problems. Take self-driving cars: perhaps they are inevitable, but it matters a lot whether it takes us 6 months, 6 years or 60 years to get there.

What has this got to do with education?

Ten years ago, I wrote a chapter in my book Seven Myths about Education about the problems with using sketchy predictions of the future to guide what we teach and how we teach. I stand by that. Let’s base our education systems on what we know works, not on what we think might work in 20 years’ time.