Revolutionary technology + massive adoption ≠ good investment
Investors have poured money into a bottomless pit, attracted by the growth and glamour of the industry. The airline industry since its birth has had a collective net loss, in aggregate, despite moving hundreds of millions of people.
Commodity Product, no switching costs. Infinite competition
True. But it has added enormous benefits to many other parts of the economy.
Airlines do not capture that value.
That is where the AI companies are. Adding value they cannot capture
The industrialisation essentially socializes the cost across a lot more people though, so even though it doesn't make a profit it does mean people can have air travel without it costing millions per flight for the few people who can afford it. Essentially the economies of scale from having lots of flights isn't enough to make it profitable but they are enough to make it affordable.
There's no spare money to extract from the airline industry but it's still very useful. The same could be true for AI in the long term.
Sometimes the goal of an industry is to exist rather than to make a profit, because the benefit to society is more important than profit. People don't like that though so they do a bit of creative accounting or head-in-the-sand denial around it.
Of course it could, let’s start with making the models open weight and entirely open source. Fully publicly owned and not shaped to maximise profits for the shareholders.
Oh wait, Scam Altman entered the chat and turned a non-profit lab into the next biggest IPO vehicle the world has ever seen.
OpenAI launched as a nonprofit research institution. Its announcement explicitly said it wanted to pursue AI “unconstrained by a need to generate financial return,” produce value for everyone rather than shareholders, publish research and share patents broadly.
The number of competitors for SOTA drops by a few every year. The winners make more money, get more revenue, buy more compute, train better model with compute, buy best talent, and the cycle goes.
I think it's easier to fall behind and never catch back up than people think. One disastrous training run can leave a lab months to a year behind. For example, Meta's disastrous LLAMA 4 models. Meta is lucky to have their ads business as a funding source. However, Anthropic's revenue is growing so fast, that ability to use ads as a funding source to stay in the race may not last much longer for Meta.
To me, SOTA LLM training is very much like new chip fab nodes. One disastrous node can put you behind for many years or forever. The cost to build the next chip node doubles every every 4 years (Rock's law). The cost to train the next SOTA model likely has some similar power law which means over time, it's too costly for losers to keep up. The only reason TSMC isn't a defacto monopoly for advanced chip nodes is strictly due to geopolitics.
I can't prove it. It's just my opinion.
If we couldn’t isolate a variable we would never be able to argue.
Using a better model is an advantage even if only for the coders. There are a million ways to turn that into profit, both proper and not so proper but that’s the beauty of ceteris paribus: the other factors do not matter now.
Mind pointing where that profit for companies consuming AI is? I don’t mean hypotheticals. Where are the proof that current AI contributes positively to ROI?
This might actually still hold true now, or or at least many actors in the market behave that way. But I'm not so sure there isn't a cliff to that effect. At some point, if SOTA models remain expensive, it'll turn into a market advantage to figure out how to get things done without depending on the most expensive tooling available.
Similar scenario, different phrasing: if your company relies on overqualified workers to deliver 100% quality, the market may still decide that it's fine to go with 90% quality for 50% the price.
I never said there will never be a diminishing return. I'm challenging the statement that we've already hit.
Note: We're still scaling chip nodes. It's still worth it for TSMC and chip design companies to invest hundreds of billions into every new chip node every 2-3 years. This is after decades of scaling already.
Opus 4.5 vs Fable 5? i think the output difference will be negligible, and the opportunity cost of paying an order of magnitude more will actually benefit my business.
Heck, just take Kimi K3 and start your training run from that.