I think it when it starts having a broad impact on the way we live, like the internet and smartphones did.

Last night the AI and I teamed up to write bedtime stories on demand and on the fly for whatever themes my daughter mused. I think about a billion families would enjoy an app that did that.

And yet people aren’t doing this so I guess something is still missing. Awareness? UX? Something else? All? The company that figures it out might earn a lot.

Price. I have a dozen ideas that could be done using GPT-3. Most of them aren't financially viable, I'm not burning half my income to build an MVP.

See what is happening with StableDiffusion, a model was released opensource, performance in the same league as closed source, usable on consumer hardware and (non AI) techies start to modify it. The biggest steps/modifications are by specialists no doubt, yet still opensource, but others are happily glueing parts together to make something else. The key really seems to be access, an hosted API is rather hostile to innovation, using and especially experimenting is expensive, modification can only happen within whatever the API allows. For the tech to get bigger and more noticed faster more people need to be able to tinker with it.

The real reason LLMs are hard to sell is alignment issues. The models reflect our biases. As an app developer, are you willing to risk it? You app—using GPT-3—might output racist remarks and then you're in trouble.

Maybe so, but unless there are a billion families out there who are already privileged enough to have internet access and are willing to pay for it, I don't see it happening.

I do think there's a good chance this stuff will have that level of impact, I just don't think it has yet.

(Though I'm not convinced it will. If you wanted to be a skeptic, you could argue that we're already in a huge content glut; there's basically infinite content available for almost free. So does radically lower the cost matter that much?

(Maybe in low-level ways, like it'll increase the abilities of small indie studio to produce high-level content. But to your average consumer, maybe it's not noticable.)

> So does radically lower the cost matter that much?

It's not about lowering the cost, it's about generating content that fits your specific preferences in real time. Forget on demand streaming of off-the-shelf content, people will want on demand content generation.

And then, as is tradition, the next step after that is giving you content you did not even know you wanted. That can be perverse marketing, feeding you sensationalized stuff that keeps you hooked 24/7¹ but the better alternative is something that is tailored to your general tastes, genres, writing styles, etc. so that you receive the content that is best received by you at that particular moment (matching your mood, your goals, your style)

––––––––––

1. This image comes to mind... https://www.wallpaperflare.com/dystopian-cyberpunk-sad-virtu...

It's an interesting scenario. Sort of like a mega-tiktok.

Your image seems apropos; I'm reminded of David Foster Wallace, as well, with his whole obsession w/ entertaining ourselves to death...

Really?

Because my experience with reading bedtime stories to kids is that they want to hear the. same. story. every. night.

For us, it's more like an interest in the continuing adventures of existing beloved characters.

It's all smoke and mirrors.

Ironically, so is most of real life.

care to elaborate?

I'll admit it when I can actually run any of it locally without needing 37946 GB of VRAM and 37 Nvidia Teslas to even load it.

You can run stable diffusion on a MBP and produce images in under a minute. It's training these models that takes the crazy GPU power - running them is quite reasonable.

Still can't run GPT-3 or even GPT-J locally though, which is what the article is about. Learning takes a whole datacenter with actual terabytes loaded into VRAM, sure, but even running it requires you to have enough space on the card to actually load the model. Which is usually still in the 20G+ range.

Stable diffusion is about the only one of these useful groundbreaking models that can run on normal hardware to some extent, and even that's extremely limited with only like what, 256x256 being possible with a 6G card and 512x512 on a 10G card? So thanks for pointing out the one partial exception.

With stable diffusion Im making right now an image every 26 seconds at 512x512 with 50 sampling steps with https://github.com/JoePenna/Dreambooth-Stable-Diffusion

The training with a beefy GPU from vast.ai (RTX 3090 with 24vram) and Im generating the images with a GTX 1080 with 4vram, so no need for 6 or even 10 GVram from my testing