The research on AI in this report is rather old fashioned, contrary to the newness of the topic and the young age of the researchers. The approach makes this research feel like a university assignment. I mean, it is methodologically correct in the sense that statements are accompanied by data charts, but the mind frame is a decade old classic corporate. It may be a new addition to an AI business overview course material, but it is old by today’s standard of what can be considered as new frontier in AI business. I say this because the focus is on the IT enterprise cloud software and hardware in a classic productivity analysis and procurement scenario, not on the disruptive AI that you would be looking for as a source of inspiration that you can use to identify the next AI business leaders of tomorrow (the future FAANGs).
One glaring omission of this report in line with my above commentary is the metaverse. This term (or omniverse) has not been mention once in the entire document! This was a surprise to me because the publication of the report came three months after Facebook change its name to Meta Platforms, a watershed moment for the high tech industry, and more than a year after Nvidia announced Omniverse Open Beta. AI is at the heart of this new technology where the future dominant platforms will be born, if they are not born already (a highly debated subject).
On the positive side, the report touches on an important aspect of AI: its huge capacity to learning, which over time is bound to have a large impact in most businesses. William Summerlin and Frank Downing (the authors of AI analysis section of this report) use the Wright’s Law as a predictor of consistent cost decline (accompanied by an increase of performance) of AI software and hardware reasoning that this in turn leads to the decline rate of AI training costs twice as fast as Moore’s Law: “Informed by Wright’s Law, we estimate that AI-relative compute unit (RCU) production costs could decline at a 39% annual rate and that software improvements could contribute an additional 37% in cost declines during the next eight years. In other words, the convergence of hardware and software could drive down AI training costs by 60% at an annual rate by 2030”.
The users of AI benefit from this trend through increased productivity:
The drop in the cost of AI training could have significant consequences. This observation doesn’t get the prominent attention I think it deserves. The estimated decline of training cost of a GPT-3 sized model (a language model that generates text) is remarkable, but when put into the context of the human brain capacity to learn based on the number of synapses, the low cost of AI training will make possible the development of ultra-smart applications that will have a high impact in both industrial and social settings. “Given 240 trillion synapses, the cost to train a neural network equivalent in size to the human brain in 2021 would have been $2.5 billion and is likely to drop 60% at an annual rate to $600,000 by 2030”.
This model predicts a significant trend. Imagine what kind of AI applications will be developed when companies can build or afford to buy (or share) neural networks made of 240 trillion artificial synapses and train them for the cost of $600k.
The question is what kind of companies will be able to take advantage of such advanced AI? The report treats AI like a utility service similar to internet or electricity, a potent computing power that can be harnessed straightforwardly to make smarter products. However, AI is much more than that.
There is invisible AI that gradually becomes enmeshed into our world by stealth in form of smart software (online or functional hardware), and then there is entity AI that lives along with us as co-workers or social helpers. There is also the big AI in form of large systems that process information collected from a vast array of sources ranging from computer systems, media, audio-visual devices to micro-sensors.
For all of these AI powered systems to work they need an ecosystem governed by agreed technological and social protocols that determine how all these AI components interact between themselves and us, humans.
Perhaps that is the AI innovation platform that should be at the centre of this analysis. To put it in the context of the previous significant technological transformation, this AI platform is similar to the internet network that links computers and computer systems from around the globe, where computers are the equivalent of AI units. I am inclined to believe this is where the future Microsofts and Googles of AI will rise. Nvidia is a good candidate for becoming one of such AI leaders over the next decade.
AI Could Automate The Tasks Of Knowledge Workers And Boost Productivity
The report considers convergence as being one of the two key characteristics of innovation over the next decade, yet it misses the opportunity to apply this concept. AI is seen here as a distinct, separate industry that produces hardware and software tools used for tasks automation. It is puzzling why AI is just seen as enterprise software running in the cloud powered by capable servers. The idea that AI software leads to an increase in demand for AI hardware is similar to the idea that demand for cloud applications leads to demand for servers. This is the world seen through the glasses of a conservative engineer who believes the future world will be the same, just run with more powerful computers.
My impression is that this section of the report was written in a hurry, because I found two doubles (ex. “57% at a compound annual annual rate during the next nine years”). The editors missed the chance here to use an AI spellchecker…
In relation to AI and its socio-economic impact, the report only mentions how knowledge workers are estimated to achieve a productivity improvement of 140% by 2030. This is what you would expect to see in an an ad for a cloud software product.
ARK’s Best Ideas of 2022 exploration of opportunities in AI stops at the office. It doesn’t venture to take a walk on the factory floor or out to the agricultural fields, industrial parks or the small enterprise workshops. This is where the transformation promises to be the most radical, especially as a result of the convergence between multiple technologies (AI, Robotics, IoT, 3D Printing, etc). The pandemic demonstrated how acute is the problem of limited supply of labour in all these areas of the economy and that is precisely where AI can have the most disruptive and consequential impact.