The head of OpenAI, Sam Altman, recently made this prediction:
Fundamentally our business, and I think the business of every other model provider, is going to look like selling tokens. You know, they may come from bigger or smaller models, which makes them more or less expensive. They may use more or less reasoning, which also makes them more or less expensive. They may be running all the time in the background trying to help you out, they may run only when you need them if you want to pay less. They may work super hard, you know, spend tens of millions, hundreds of millions, of someday billions of dollars on a single problem, that’s really valuable.
But we see a future where intelligence is a utility like electricity or water, and people buy it from us on a meter and use it for whatever they want to use it for.
I agree with his prediction and want to share my thoughts on what it means for individuals, society and the economy, and to unpack some key assumptions behind Altman’s statement.
“Fundamentally our business, and I think the business of every other model provider, is going to look like selling tokens.”
The phrase “is going to be” means he’s speaking about the future. This means Altman foresees most people getting AI services remotely from large tech companies like OpenAI, and not in-house from personal computing devices they own and keep in their houses or pockets. I agree this will be true, though improvements to the price-performance of LLMs and an eventual drop in the prices of GPUs and RAM as supply expands will someday make it feasible for average people to host AIs on their own hardware.
However, it will remain more convenient and probably cheaper to access LLMs over the internet from companies like OpenAI. Additionally, thanks to having better hardware and more of it, and more advanced models, the companies will always have a monopoly over the fastest and smartest AIs. For those reasons, access to AI will stay under the central control of a handful of big companies and government agencies with the resources to build data centers and to do cutting-edge AI research.
“You know, they may come from bigger or smaller models, which makes them more or less expensive. They may use more or less reasoning, which also makes them more or less expensive. They may be running all the time in the background trying to help you out, they may run only when you need them if you want to pay less.”
Access to AI will be tiered and principally based on how much money the user is willing to spend. As is the case today, different AI models will have different levels of capability, and it will cost more money to talk to better ones. As a result, though everyone in the future will have access to AIs, not all AIs will be equal, and the rich will have the best.
“They may work super hard, you know, spend tens of millions, hundreds of millions, of someday billions of dollars on a single problem, that’s really valuable.”
This is an important point. Not every question or task that humans put before an AI is equally hard for it to handle. Asking one to summarize a news article is easier than asking it to make a new discovery in particle physics. A “dumber” AI with only a slower and cheaper processor could handle the former perfectly well, whereas only a cutting-edge AI with a data center at its command could do the latter.
If true AI were available, the vast majority of people would not use it to attempt solving high-level problems that cost millions of dollars to solve–they would use it to meet their everyday needs. This includes things like soliciting advice on routine things like cooking new meal recipes, remembering chores, and making conversation. If endowed with a robot body, most people would have their AIs do mundane work like cooking or cleaning.
Today, the world’s best supercomputer can’t beat a smart nine-year-old kid at tic-tac-toe; the two will tie each other forever. This is because the game is so simple that beyond a certain threshold, adding more brainpower doesn’t help you come up with a better strategy for it. Likewise, we’ll reach a point where nearly all needs that average people have can be sated by AIs and robots that are not cutting-edge.
There is a ceiling to the level of intelligence an AI must have to perform cognitive tasks as well as humans, and a ceiling to how well a robot must be engineered to perform the physical tasks as well as humans. Once those respective ceilings are reached, the vast majority of human demands for cognitive and physical services will be met, yet the fields of AI and robotics will continue advancing. As the cost-performance of technology improves, the prices of AIs and robots designed to satisfy routine human needs will drop until everyone can easily afford them.
We will all have machine servants that can answer every question, provide us with true companionship, and do physical work for us around the house. Today, only the rich enjoy such services, but this century it will become common for everyone, and stand as another example of technology raising the standard of living. As wonderful and this sounds, it will not be a utopia, nor will it necessarily make the world more equal for humans. Rich individuals, corporations and governments will be able to afford access to more advanced AIs and robots and more of them than ordinary people. The disparity will grow when mass unemployment robs most people of their income and leaves them with only enough money (from UBI?) for the machines calibrated to not surpass what is needed to satisfy everyday needs (cognitive, physical, emotional).
For those reasons, AI technology will not be truly democratizing. Ordinary people will have access to tools of enormous power, allowing them to do things previously impossible, but that access will be controlled by large companies and entities. Additionally, the best AIs will also be too expensive for ordinary people to afford, meaning richer people and organizations will enjoy major advantages as they do now.
Consider the realm of economics. Let’s say it’s 2046 and I have a robot butler that can do any physical task as well as a human. I am enterprising, so I decide to start renting my robot out to mow peoples’ lawns in my neighborhood. After all, it’s a capital good, and I’m a capitalist. My business plan isn’t as solid as it seems.
Problem #1: Since robots have gotten cheap enough for anyone to afford, most of my neighbors already have their own, so their lawns are already being mowed.
Problem #2: The RoboChop Lawn Care company offers grass-cutting services in my neighborhood, does it faster and charges less money than I do. That’s because they have robots that are specialized for cutting grass and nothing else. One of their wedge-shaped, coffee-table sized robots with four wheels and an inbuilt rotating blade can zip along four times faster than my robot butler can walk while pushing a traditional lawn mower. And since RoboChop ONLY mows laws and its robot crew does multiple jobs per day, the company can offer volume discounts for its services that I can’t.
Even though I have a robot and access to AI, I’m still so outclassed in the market by bigger competitors that there’s no niche where I can succeed. At first glance, this clashes with my essay “Will future technologies end capitalism? No.”, but the key ideas are in fact in harmony: Yes, the future economy will still be capitalist, and arguably more so than now since most humans will have easy-to-use capital goods (AI and robots), but the resource disparity between average people and businesses will be so large that the former won’t be able to participate in that capitalist economy. A rising tide will lift all boats, but some much more than others.