Summary: The history of computing shows that increased availability of cheap compute, especially on the client side, creates opportunities for improved usability. AI will continue this trend, with vastly better usability to come. Sadly, slow growth in electricity capacity will likely delay the ability of currently-rich countries to realize the full benefits of AI for as much as 20 years.
The history of computing is characterized by Moore’s Law:
Moore's Law was proposed by Intel co-founder Gordon Moore in 1965. It states that the number of transistors on a microchip doubles about every two years. This doubling effect scales down costs while exponentially increasing performance and creating a compounding increase in computing power and efficiency since the 1960s. The impact of Moore's Law can be seen in the rapid development of smaller, faster, and more affordable electronic devices, from smartphones to laptops, which have transformed the way we live, work, and communicate.
The main effect of this cheaper and more powerful compute has been the ability of computers to support better user interface styles:
Batch processing: very efficient in CPU, but terrible for users.
Line-mode UI: still very efficient CPU use, still terrible usability, but a little better than batch processing.
Full-screen UI: this interface style required more computation, mainly on the back end, which had to keep track of everything on the screen instead of just the last few characters typed by the user. Full-screen terminals also needed more electronics to render the full-screen UI and allow for form-filling across this expanse, compared to earlier terminals that simply scrolled the previous lines of text up the screen.
Graphical user interfaces (GUI): very computationally intensive, both in terms of calculating what should be shown where, and in terms of rendering these pixels on the screen.
Intent-based outcome specification: the new AI-driven interaction style where users specify what they want to happen rather than the detailed commands needed to achieve this outcome. Super-demanding of compute power for AI inference. (Also, as a one-time expense, even more demanding of the AI learning compute to build the foundation model driving the inference.)
Even though AI is the first new interaction paradigm in 60 years, it thus follows along a long tradition of ever-more computer power being needed to support the user interface.
I should point out that increasing compute only creates the possibility for improved usability. It’s also possible to ship designs that employ new UI technology but deliver worse usability. Every new UI generation has seen this sad effect: We get new opportunities, but they’re abused, as the initial product launches are driven by engineers rather than UX professionals. AI is no exception, as proven by the substandard usability in the early releases of several leading AI products, which violated elementary usability guidelines.
More compute doesn’t automatically equal better usability. We still need to follow the UX design process to actually get good results. But even though I’m a usability guy, not a hardware guy, I must admit that most of the halfway-good user interfaces we enjoy today have only happened because computing has become cheaper, faster, and more widespread. In particular, the personal computer has driven a level of usability that would never have been possible on shared computers.
Future of AI Compute: Limited by Electricity?
Many AI experts believe in the scaling law: the more compute you throw at AI, the more intelligent it becomes.
This scaling law even seems to apply to the growth of biological intelligence: various pre-human species had gradually increasing brain sizes, corresponding to their gradual increase in ability to use tools. Among current humans, there’s a positive correlation between brain size and IQ. Since this correlation is fairly low (estimated at 0.24, rather than the value of 1, which would indicate a perfect match between the two measures), there are clearly other things at stake than pure size, but size does matter.
We shouldn’t make too much of the biological analogy since computer hardware and software are obviously different than meatware. But it does seem that AI is driven by both sheer size (amount of training and inference compute, as reflected in the number of model parameters) and the quality of the “thinking” (as reflected by the software).
Current GPT-4 level AI is good enough to use, with the ability to improve knowledge worker productivity by about 40%. But AI is still distinctly subhuman intelligence, with many well-documented weaknesses. We need super-human intelligence in our AI to realize the many hopes for a quantum leap in the lot of humanity.
Upscaling AI power by 1000x will allow a hugely improved user experience by using the resulting super-human intelligence. (Midjourney)
What’s holding us back? Some people say the lack of pure compute because high-end GPUs are not shipping in sufficient numbers. Others claim that the true limit will be electricity: even as more GPUs are manufactured every year, data centers are not able to get sufficient electricity supply to power these chips.
Both are clearly serious problems. But both are solvable in the long term.
For the last decade, there has been a phenomenon similar to Moore’s Law in effect: the amount of compute delivered by a given amount of electricity has doubled every two years due to hardware advances. The recent NVIDIA Blackwell chip jumps ahead of this curve, by only needing a quarter of the power of their previous Hopper chip. (It is unclear whether this better-than-trend improvement is a one-time benefit of a particularly strong NVIDIA invention or whether we will now see faster improvements in the power-efficiency of AI chips based on the vastly increased investment in R&D in this previously esoteric field.)
Hardware improvements are thus strong. But software improvements are even more impressive: the amount of compute needed to produce a given level of AI results has been cut in half every 8 months or so. This means that over a two-year period, compute needs are likely to drop to one-eighth of current levels due to better software.
Combine hardware and software improvements, and the electricity needs for a given amount of AI results will drop to 1/16 of current levels in two years.
Midjourney (a generative AI tool that is the second-most used AI tool by UX professionals): 2 grams of CO2 per image
Human designer: 5,500 grams of CO2 per image (if living in the United States — the emissions to sustain a human designer for enough time to create an image are only 700 grams of CO2 if the designer lives in India)
I typically generate between 10 and 50 images on Midjourney for every illustration I use in my articles. Thus, I emit between 20 and 100 grams of CO2 per illustration — let’s say 60 g for an average. For the human designer, let’s say that I only needed to request one redesign per image: that would put the emissions at around 11,000 grams of CO2 per published illustration by a human designer, or 183 times as much pollution as would result from using generative AI. (Which is what I actually use, so I guess I’m doing my part to save the planet.)
Carbon Emissions | Design an Image | One Published Image |
Midjourney Generative AI | 2 g | 60 g |
Human Designer (USA) | 5,500 g | 11,000 g |
Human Designer (India) | 700 g | 1,400 g |
The point that my actual emissions for AI use are much higher than the per-image estimate points to a third opportunity for improving the performance of AI systems: usability advances! Yes, AI is an intent-based interaction paradigm, but current AI is poor at divining user intent and requires substantial iteration to deliver acceptable results.
Making AI better at interacting with users should substantially reduce the need for iteration. Probably not eliminate it, if we take the example of working with a human designer as a precedent. It’s very rare to get perfect results from a designer on the first attempt. My first ambition for AI is to become as good as humans. (In my example, this would mean an improvement factor of between 5x and 25x for creating the images I want.)
However, there’s no reason for the limited ambition of wanting AI to become as good as humans. AI should be better than humans, and will be better once it has somewhere between 100x and 1000x more power.
Let’s say this takes 10 years, and let’s go with the high estimate of needing 1000x the AI power to achieve superhuman performance from “GPT-10” (or whatever will be the leading model in 2034). Well, 10 years will see 5 of those two-year periods where AI’s electricity consumption for a given level of AI output drops by a factor of 16. 16 to the power of 5 is one million. Furthermore, usability advances should add another factor of 10x to the efficiency of using AI. Thus, if we need 1000x the AI power for a given outcome, the electricity consumption to deliver that will be 0.01% of current use.
This results from multiplying together these expected advances:
Hardware: 32x
Software: 32,000x
Usability: 10x
(I think it’s likely that hardware improvements will be more than this estimate and that software improvements will be less. But the overall improvement in the end-to-end system is still likely to require 10 million times less electricity per unit of AI power and 10,000 times less electricity per deliverable.)
Better AI = More Expensive AI
Does this mean that electricity will not be a limiting factor for AI? Sadly not. If AI actually delivers superhuman performance in 10 years, everybody will want to use it, all the time. Let’s say a million times more AI use worldwide in 10 years than the rather sporadic AI use we see now. In that case, 0.01% power consumption per use still equates to 100 times more overall electricity consumed.
Currently, data centers account for about 1% of the energy consumption in the United States, and AI-specific data centers are only about one-tenth of that, or 0.1%. Well, 100 times 0.1% equals 10%. Meaning that in 10 years AI will need 10% of the total current energy consumption of the United States. Adding 10% energy use for more than doubling our standard of living doesn’t sound that bad. (In fact, this will be in line with historical data showing a drastic drop in carbon emissions per unit of GDP delivered by the economy in recent years.) The problem is that all the energy for AI has to be electricity. Steampunk is science fiction and can’t drive AI inference.
Building enough power plants in 10 years to supply the need is not likely to happen in the United States or Europe. China and South Korea build electricity infrastructure at the required speed, but very few other countries have the political will to match them.
Something has to give. My guess is that AI prices will increase dramatically in jurisdictions like the United States and the European Union that can’t build infrastructure.
This might mean a subscription fee of $200 per month for frontier foundation models (stated in fixed prices — the actual number will have to be adjusted for inflation). Much more than the laughably cheap $25/month current fee for ChatGPT Plus.
Companies will still pay up. Remember that I’m predicting super-human AI performance. Instead of the current 40% productivity improvement, this will likely result in maybe 200% productivity improvement for most knowledge work. (In other words, one AI-augmented human will do the work of 3 non-aided humans.) As long as $200 for AI is less than the monthly cost of two additional employees, it’ll be worth paying.
For example, the monthly salary of a mid-level UX designer in India is about US $1,000. Thus, saving two such salaries corresponds to 10x the cost of super-human AI. This doesn’t even account for also saving the overhead cost of more staff. The true ROI will be about 15x. Actually, the most likely is an ROI of 30x (corresponding to 3,000%) when also accounting for the expected doubling of Indian salaries over the next decade. (In the United States, the ROI from using Super-Human AI in UX will be closer to 12,000%.)
One of the world’s poorest countries, Uganda currently has a GDP per capita of US$964 per year, or $80/month. But that’s the average for everybody in the country, including babies and subsistence farmers. Knowledge workers almost certainly make more than $200 per month, meaning that a Ugandan company would still have ROI >100% by replacing two knowledge workers with one AI subscription.
So far, so good. We can afford super-human AI for corporate work, even in the poorest countries. But something has to give. I am afraid that many of the non-corporate advances I expect from AI, such as better education and better healthcare, might well be the victims. (For example, private schools in Uganda often cost around US $2 per day, which is the amount better-off parents can afford to improve their child’s education. This is far below the $200/month that might be needed to add AI to improve that child’s education even more.)
I seem to have written myself into a sad conclusion, but only for the next decade. Let’s look two decades out, and either sufficient electricity infrastructure will be built (remember that more electricity is also needed for other desirabilities, such as emissions-free cars), or AI will benefit from subsequent advances, driving costs back down again.
I’m not predicting unlimited happiness with more powerful AI. But I do think the improved user experience it’ll bring and the doubling of GDP resulting from the associated productivity advances will benefit humans more than most other technologies. (Midjourney)
Comments