I used to run a subscription box called "Candy Japan", and got #1 on Google for "japanese candy" for after posting about it on HN. The position lasted for years AFAIK.
I was going to make an argument that it's inevitable, because at some point compute will get so cheap that someone could just train one at home, and since the knowledge of how to do it is out there, people will do it.
But seeing that a company like Meta is using >100k GPUs to train these models, even at 25% yearly improvement it would still take until the year ~2060 before someone could buy 50 GPUs and have the equivalent power to train one privately. So I suppose if society decided to outlaw LLM training, or a market crash put off companies from continuing to do it, it might be possible to put the genie back in the bottle for a few decades.
I wouldn't be surprised however if there are still 10x algorithmic improvements to be found too...
I think there's a category of these kinds of things where you apply AI to do something humans could do, but could not be bothered to do. Or could not profitably do. At least no human would categorize all these reviews just for lols.
Another recent example from HN would be that site which just lists hotel rooms that have a desk and a chair. It would be an incredibly dull task for a human to look at a million hotel room pictures and just select if they have a desk or not.
What else somewhat useful/fun could we do applying perhaps a little worse than human attention at something, but a lot of it?
We've been hosting both AirBnB and Couchers/CouchSurfing.
AirBnB has the vibe that you as the host are a provider of a service, which will be rated by the "customer". Couchsurfing is just some people hanging out.
Agreed, the TikTok one will be vastly more popular, as it requires no creative input. Also much more near-term. Some version could be shipped within months.
It's already partially that, just with humans still supplying the prompts and doing some cherrypicking before posting their AI-generated videos. I wonder if there will be some "for you (AI)" and "for you (non-AI)", and which one will end up being the default?
Try Roblox (YouTube but for games essentially). You can publish a game just a few clicks, and even quite simple games can get popular enough that you get the satisfaction of seeing others interact with the thing you made, which is very motivating.
This will probably have some cool non-obvious benefits.
For instance if the scenes are a blob of input weights, what would it look like to add some noise to those, could you get some cool output that wouldn't otherwise be possible?
Would it look interesting if you took two different scene representations and interpolated between them? Etc. etc.
