How many people graduate from a US software engineering degree each year? About 100k? If they (the 100k in the US) earn $100k each in the first year, before gaining the skills to earn more, that's $10 billion a year, every year. If you can capture that market for next 20 years, it's worth $200 billion.
Except… can you capture it? A junior dev is… not exactly someone you want connecting to your business-critical database without supervision, and a real human dev will get better with a predictable rate. Will LLMs get better? The makers are betting on that, but we'll only know after the model releases, and even then after we play with them for a bit to differentiate between the record performance on whichever benchmark and the actual work we want them to do.
Then there's the question of can you really keep an edge for 20 years with investments today: Sometime between 2030-2035, there's likely to be models matching 2025-SOTA performance that run on ${year}'s high-end smartphone.
(Well, unless we all die in WW3 because of Russia getting desperate from its failure to remove Ukraine's sovereignty, or because China has a hot war with Taiwan and/or the USA messing with global consumer electronics supplies, but I don't think those get priced into the market…).
But that's just the USA's software developers in just their first year after graduating. Software devs are 1% of the US job market, the first year after graduation is (66-21=45 years, 1/45 ~= 2%) of a working life, the US is just 4% of the world's population/25% GDP.
For the 1% to matter, there have to be other jobs that LLMs can do as well as a fresh graduate. I don't know, are LLMs like someone the first year out of law school or medical school, or are those schools better than software? Certainly the home robotics' AI are nowhere near ready yet, no plumber, no driver (despite the news about new car AIs), would you trust an Optimus to cut your hair? etc.
For the 2% to matter, depends how seriously you take the projections of improvements. Myself, I do not. Looks like exponential improvements come at exponential costs, and you run out of money to spend for further improvements very quickly.
For the 4% to matter, depends on how fast other economies grow. 4% by population, about 25% by GDP. I believe China is still growing quite fast, likely to continue. Them getting +160% growth, and thus getting 2.6x times the money available to burn on AI tokens, over the next 20 years would be unsurprising.
All in all, I don't think the USA is competent enough at large-scale projects to handle the infrastructure that this kind of AI would need, so I think it's a bubble and will burst before 2030 because of that. China seems to be able to pull off this kind of infrastructure, so may pull ahead after the US does whatever it does.
> For the 1% to matter, there have to be other jobs that LLMs can do as well as a fresh graduate. I don't know, are LLMs like someone the first year out of law school or medical school, or are those schools better than software?
Before looking to medical and law schools, I might look to middle-manager school or salesperson school or bookkeeper school.
I don’t know enough to speculate even beyond those crude guesses, but as I thought about this question, I found it interesting to skim the US’ employment-by-detailed-occupation chart:
That analysis only makes sense if companies value AI tools as much as equivalent human productivity. Hypothetically say you have a company with 100 junior developers. An AI service comes around that doubles the productivity of your junior developers, so you can keep only 50 of them. Would the company pay 5M dollars a year forever for that service?
In my experience, the answer is a resounding no. They’ll nickel and dime some kind of per-seat licensing on a monthly basis that costs less than 100 USD or whatever. So for every 100$ in salary you can automate away, you might get 2$ in subscription payments if you are lucky, at current rates.
Except… can you capture it? A junior dev is… not exactly someone you want connecting to your business-critical database without supervision, and a real human dev will get better with a predictable rate. Will LLMs get better? The makers are betting on that, but we'll only know after the model releases, and even then after we play with them for a bit to differentiate between the record performance on whichever benchmark and the actual work we want them to do.
Then there's the question of can you really keep an edge for 20 years with investments today: Sometime between 2030-2035, there's likely to be models matching 2025-SOTA performance that run on ${year}'s high-end smartphone.
(Well, unless we all die in WW3 because of Russia getting desperate from its failure to remove Ukraine's sovereignty, or because China has a hot war with Taiwan and/or the USA messing with global consumer electronics supplies, but I don't think those get priced into the market…).