Why is that obvious? Even with effectively complete stagnation and just existing technology + limited RLVR, I can see how this could be trillion-dollars level useful.
That's more of a UI problem than a limitation in Diffusion tech.
That's a customer who'll pay, it might be worth a lot. But a $trillion per year?
The glaring issue with it back then was that unlike an LLM that can be understanding of what you try to explain and bit more consistent the diffusion models ability to read and understand your prompt wasn't really there yet, you're more shotgunning keywords and hope the seed lottery gives you something nice.
But recent image generation models are significantly better in stable output. Something like qwen image will care a lot more about your prompt and not entirely redraw the scene into something else just because you change the seed.
Meaning that the UI experiments already exist but the models are still a bit away from maturity.
On the other hand, when looking at how models are actually evolving I'm not entirely convinced we'll need particularly many classically trained artists in roles where they draw static images with some AI acceleration. I expect people to talk to an LLM interface that can take the dumbest of instructions and carefully adjust a picture, sound, music or an entire two hour movie. Where the artist would benefit more by knowing the terminology and the granular abilities of the system than by being able to hold a pencil.
The entertainment and media industry is worth trillions on an annual basis, if AI can eat a fraction of that in addition to some other work-roles it will easily be worth the current valuations.
ChatGPT's $10b per year is not insignificant tho.
800M active users aka 10% of worlds population.
The promise of this future demand is what is driving the inflation of the stock market, with investors happy to ignore the deep losses accruing to every AI software player...for now. Pulling the plug on the capacity-building deals is effectively an admission that demand was overestimated, and the market will tank accordingly.
It says it all about current market mania that Nvidia (who sells most of the future chip capacity) is valued at $4 trillion, more than every publicly traded pharmaceutical company (who have decades of predictable future cash flows) combined.
That's the lowest of the low and even you accept it doesn't work (yet), how can LLMs be worth 50% of the last years of gdp growth if it's that bad. Do you think customer support represents 50% of newly created value ? I bet it isn't event .5%
Maybe if companies would wire up their "oh a customer is complaining try and talk them out of canceling their account offer them a mild discount in exchange for locking in for a year contract" API to the LLM? Okay, but that's not a trillion-dollar service.
Maybe it's because I find writing easy, but I find the text generation broadly useless except for scamming. The search capabilities are interesting but the falsehoods that come from LLM questions undermine it.
The programming and visual art capabilities are most impressive to me... but where's the companies making killings on those? Where's the animation studio cranking out Pixar-quality movies as weekly episodes?
I work in the industry and I know that ad agencies are already moving onto AI gen for social ads.
For VFX and films the tech is not there yet, since OpenAI believes they can build the next TikTok on AI (a proposition being tested now) and Google is just being Google - building amazing tools but with little understanding (so far) of how to deploy them on the market.
Still Google is likely ahead in building tools that are being used (Nano Banana and Veo 3) while the Chinese open source labs are delivering impressive stuff that you run locally or increasingly on a rented H100 on the cloud.
There are always a few comments that make it seem like LLMs have done nothing valuable despite massive levels of adoption.
> You can easily google "generative AI success stories" and read about them.
notice you suggested asking Google and not chatgpt.
Search engines are better at certain tasks than others.
If I said should FLY to Spain is it a cheap shot against sailing because I didn't mention it?
Check out Neural Viz. Unthinkable for one guy without AI. And we're still in the Geocities stage of this stuff.
A non-Pixar animation studio, with presumably a >10x lower budget than Pixar itself, cranking out weekly Pixar movies would be like >1000x acceleration. And indeed, that's not a thing yet for animated movies. The example I gave shows that it's already quite a big X, though.
It's either world-ending hard to believe conjecture, like the death of scarcity, or it's... ads. Ads. You know, the thing we're already doing?
So, it's not looking great. Maybe we will find monetization strategies, but they're certainly not present now, even by the largest players with the most to lose.
But second of all, companies do not need to 10x their software production output. Rather the goal, if the 10x productivity is achieved, is to _reduce_ human labor while retaining desired levels of output.
If ultimately you're replacing humans with AI agents, you're exchanging one value for another.
What if the amount of slop generated counteracts the amount of productivity gained ? For every line of code it writes it also writes some BS paragraph in a business plan, a report, &c.