This year was the tenth anniversary of the Big Bang in AI, when Geoff Hinton and some colleagues introduced deep learning, a relaunch of neural networks. Deep learning enabled the Big Tech firms in the US and China to build products and services which generated enormous amounts of money – the first time that AI was lucrative.
This year was also the fifth anniversary of a second big bang in AI – the advent of Transformer models, like GPT-3 and Dall-E. These are AI systems which achieve remarkable results by predicting what token (a piece of text or image) will come next in a sequence. In the last couple of years, AI researchers at OpenAI, Deep Mind and others have leap-frogged each other by announcing increasingly powerful natural language processing Transformer models. They are also known as large language models, or foundation models.
Bigger and bigger
These models amaze us by producing human-like text, photorealistic images – and more recently, video sequences – when prompted by a single sentence. During 2023, these models will get even larger and more powerful. They will also continue to surprise researchers with their capabilities, and will become less “brittle”, i.e. less inclined to make what look to humans like silly mistakes.
Partly as a result of this, more people will start to take the possibility that artificial general intelligence (AGI) and then superintelligence may arrive within a decade or two. For some, this will provoke excitement and a sense of opportunity, so we are likely to see further investment into AGI research to add to Google’s investment in Deep Mind and Microsoft’s in OpenAI.
For others, it will be a cause for concern. The belief that AGI might arrive soon will still be a minority view, but among that minority will be some very rich people, with the result that funding for research into AI safety and AI alignment (how to make sure that advanced AI is beneficial to humanity, and that its goals are compatible with ours) will increase sharply. The number of people working full-time on AI safety will at least double from its current worldwide number of around 300. There may also be some moves towards formal government regulation of labs working on AGI – probably in the US, since that is where the work is most advanced.
If and when superintelligence arrives, it will be the most important event in human history, bar none, so it is remarkable how little attention is paid to the prospect. There were numerous elections around the world this year, but not one prominent active politician talked about AGI. A number of former leaders are talking about it (Tony Blair and Rory Stewart in the UK, for instance), so the subject is no longer exclusive to Silicon Valley and a few eccentrics elsewhere. Sadly, mainstream discussions will most likely continue to be preoccupied with pressing short-term considerations like the economy, war, and the threat to democracy posed by modern populism.
The tech industry has been hit as hard as any by the downturn in the wider economy, and mass layoffs have followed significant falls in share prices. Shrewd observers have said for years that the current crop of Big Tech firms are not immune to economic setbacks or insurgent competitors, and they will now have to caution against commentary which veers too far in the other direction. Google, Amazon and the others are well-run businesses with huge resources, and the great growling economic engine of technological advance has not shuddered to a halt. The developed world is probably heading into a recession, but the financial system is stable, and innovation continues apace, so it will probably be a mild and short one, except for countries like Russia, Turkey and the UK, whose leaders made foolish decisions that made their countries poorer and less respected. The tech sector may well lead the world out of recession during 2023.
AI in drug development and basic science
Insilico Medicine, Exscientia and other companies pioneering the use of AI to develop drugs now have healthy pipelines of molecules going into phase two and phase three clinical tests in humans. They have proved beyond doubt that using AI can cut the cost of pharmaceutical development from decades and $ billions, to years and $ millions. The pharma giants are experimenting with the approach, but they are like supertankers, and it is hard for them to change their processes quickly.
The most notable breakthroughs in basic biology of the last couple of years was the prediction by Deep Mind’s AlphaFold and AlphaFold 2 of how proteins fold – first a million of them, and then 200 million. The way proteins fold determines their biological function. Researchers from Meta managed to speed the process up. Deep Mind has open sourced its work, and believes that it is already being used daily by most biologists around the world. It is reasonable to expect that some of these biologists will make important knock-on breakthroughs thanks to this tool. It is also likely that AI will enable at least one major scientific breakthrough next year, in an unrelated area.
Wearables and virtual worlds
More and more people wear devices which track their vital signs, and enable them to take a measure of control over their own health. This is increasingly important as the global population ages: staying healthy is obviously a matter of life and death, but it is also critical to the financial health of most countries. The next major step in this process is the inclusion of glucose monitoring in Apple’s watches, which may be coming with the series 9, to be launched in September 2023. Blood pressure monitoring may take a little longer, but you never know.
2023 will be an important year for virtual reality and the Metaverse. Not because it will be the year when they go mainstream: the technology is still not quite ready, and more important, developers have not created enough things to do while wearing VR headsets apart from playing games. (Games are a huge market, but they are not significantly increasing their share of human attention.) The importance lies in whether Mark Zuckerberg can hold out against investors in Meta, and maintain the company’s expensive research programme in the face of economic headwinds. The Metaverse is powered by AI, and will one day be one of its most important impacts. As usual with technology, it’s a matter of when, not whether: timing is everything.