The object of this article will be an overview of this past year’s major trends and developments in the subset of the tech industry that works with Artificial Intelligence.
We are primarily interested in AI as an industry rather than as a field of scientific research, so we will be looking at the economics, the politics, and the social reality of AI, as opposed to listing the best academic papers or the most salient technical achievements of 2022.
We have identified four topics that we believe are most significant and representative for AI in the year of 2022. These are heterogeneous, and – as may be expected for such a variable and complex field – they frequently overlap with each other. Taken in combination, however, they offer the most comprehensive picture of the state of the industry at the present time.
Let’s start with the first of these topics, by looking at a technology which you no doubt have already come across.
1. Text-To-Image Takes Centre Stage, Points To Text-To-Video
The domains of Artificial Intelligence continued to expand this year, with some particularly surprising and exciting applications in mathematics and the sciences. But the field that saw the most memorable and globally impactful advances was that of Text-To-Image (TTI) generation, and this for several reasons.
One of these, inevitably, was its ubiquitous cultural fallout. Social media this year has been flooded with AI-generated pictures of all sorts, posted by users amused by the endless flexibility of these new tools and amazed at the results, while the case of a fine arts competition being won by an AI has fuelled vibrant debates on the nature (and future) of art.
Another reason why TTI was so significant was the extent to which it was indicative of the broader trend of decentralisation in AI development (more on this below). While the most prominent TTI software this year belonged to Google (Imagen and Parti) and OpenAI (DALL-E 2), new AI Labs like Midjourney (which was used to win the aforementioned fine arts competition) and Stability.ai have also entered the fray and are doing surprisingly well, both in terms of technical achievements and economic performance. The race is not just to the swift this year, but to whoever wishes to run.
Finally, the TTI trend of 2022 points to the very likely next trend for 2023, which is that of Text-To-Video. Diffusion models this year definitively overtook Generative Adversarial Networks (GANs) for dynamic image generation, and as it turns out, these work quite well for video generation too. Meta’s Make-A-Video and Google’s Imagen Video (neither very imaginatively named) are both based on diffusion models, and they give us a taste of what to expect in the near future.
2. The Industry Decentralizes (Unexpectedly)
Commercial AI research has, for the last few years, been a domain almost exclusive to a few big players like OpenAI, Google’s DeepMind, and Meta. However, 2022 may be remembered as the year when all of this changed, as two separate trends converged to suggest a rapid – and largely unpredicted – decentralisation of the industry.
Firstly, a great deal of the leading engineers working for the aforementioned ‘big players’ have left those companies and moved on to create their own AI startups. This includes, for example, almost all of the authors of the groundbreaking Attention Is All You Need paper (2017), formerly employed at Google and now heading smaller collectives like Adept, Co:here, Character.ai and Inceptive. (Meta did not see a similar exodus, but note that they already proactively decentralized their AI research.)
On the whole, these new companies have been remarkably successful, particularly in the context of a year of global economic difficulties which saw investment in startups slowing down compared to 2021. For example, Anthrop\c raised a staggering $580 million this year, Inflection managed $225 million, and Co:here $125 million.
Secondly, open sourcing has successfully built on and refined on private AI models formerly believed to be beyond the technical skills of the collective. To mention just a few of various possible examples, this year we got Stable Diffusion, Bloom and OpenFold, which all either replicate (or even improve on!) DALL-E, GPT-3 and AlphaFold 1, the models developed by OpenAI and DeepMind.
Stable Diffusion was developed by Stability.ai, the entire business model of which is based on decentralisation and open sourcing. Their motto is ‘AI by the people, for the people’, and they are perhaps best described less as an AI lab than as a central hub for a variety of dispersed labs.
The most notable exception in this trend towards decentralisation is that of hardware. NVIDIA’s chips still enjoy a stranglehold on the market, and there are no signs that this will change in 2023.
3. Sentient Google Chatbots – The Blake Lemoine Controversy
In June this year, Blake Lemoine, then a senior software engineer working for Google, made the sensational claim that the company’s ultra-advanced chatbot LaMDA was self-aware. Lemoine was subsequently fired, and his claims were dismissed by Google and by AI experts at large.
The story made international headlines in a way that very seldom happens in this industry, in no small part thanks to the transcript published by Lemoine of his dialogues with LaMDA. Here the engineer is seen querying the chatbot on the topic of consciousness, and the exchange is undeniably fascinating to read.
To dismiss this episode as a blunder with neither consequence nor scientific basis is to miss the point entirely. The real lesson for the AI industry here is the ease with which the indicators of consciousness can be replicated, or more accurately, falsified – to the point that a leading expert in the field can be so thoroughly bamboozled by a machine.
While we still seem to be relatively far off (emphasis on ‘relatively’) from developing self-aware Artificial Intelligence, the Lemoine controversy reminds us that we also have no reliable model to verify self-awareness – not even for animals, let alone machines. The question is not when will AI become sentient – but how will we know that it did?
4. Regulation: A Tale Of Two Cities
After an extensive period of ratification that began last year, the European Union is expected to finally pass the Artificial Intelligence Act before the end of 2022. The law introduces mandatory risk assessment in AI development and regulates its commercial use.
The law itself is, in its current form, rather light – it is more a collection of general guidelines than anything else. But its significance for the international AI industry could be extremely broad. The EU’s tech regulations have a tendency to be adopted globally by what is known as the Brussels effect, as we’ve already seen with the European Union’s General Data Protection Regulation (GDPR).
This is not the only example of emergent and growing concerns – especially in Europe – related to AI safety. It follows on the National AI Strategy document published last year by the UK government, which is notable among other things for its recognition of the existential risks posed by superhuman artificial general intelligence (AGI), and the founding this year of Conjecture, a British startup dedicated to mitigating those risks.
Unsurprisingly, China appeared to be going in very much the opposite direction this year. This is, of course, equally if not more significant for the global AI industry. While European regulation may have greater soft influence, China leads the world in AI research in just about every metric. According to the State of AI Report of 2022, they have published more papers (counting CNKI publications) on the topic than Europe, North America and India put together, and more than quadruple the number published by the country in second place, the USA.
China not only does not balk at but proactively leads the way in AI research with surveillance applications. Topics like face and speaker recognition or object tracking are much more frequently cited in Chinese papers than in American ones, for example, although China is also the world leader in many of the less controversial fields (autonomous driving, for example).
Tensions between China and the West are not news, of course, nor are they exclusive to tech. We should therefore expect trends in AI and security to continue to diverge in the future, with European markets increasingly regulating the sort of technologies that China will increasingly develop.
No Predictions This Time
This brings to an end our overview of what we believe to be the most significant trends and events in the AI industry for 2022. It is not infrequent for articles of this nature to end with predictions for the coming year, but this is a practice we will refrain from, and instead close our report here.
Some of the topics in this report – the surprising turn towards decentralization, for example – should caution us against trying to foresee the unforeseeable. As the saying goes, the best way to predict the future is to create it – and that’s a task we’ll happily leave to the engineers.