Hacker News Reader

Comment by armchairhacker | Hacker News Reader

This is really cool. It's a text-to-speech and the gist seems to be that they synthesize it from only a little audio.

The results are clearly synthetic and need work. However what's cool is that there are a ton of characters (from popular shows and video games) and there are useful statistics like inferred emotion (which is also in the output).

Honestly it's a big problem how a lot of AIs are like "black boxes" where you really can't customize or see anything. Yeah we have DALL-E and GPT which can generate text images but the lack of customization or fine-tuning the image afterwards severely hinders what's possible with them. Ultimately what you want is something interactive, where you can control how much or little the AI generates, and give it really specific criterion.

But seriously: how did you get the domain `15.ai`?

View on HN

I agree this is an amazing demonstration of what AI can do, but I think that the current method of "learn and repeat" that depends on having tons of computing resources available is still too inefficient in many ways. Personally I'm more interested in what parameterisable formant-based synths can do, since they are extremely efficient and can produce a theoretically infinite variation of voices, although the output quality is still not great. Example: https://news.ycombinator.com/item?id=31604299

>But seriously: how did you get the domain `15.ai`?

I purchased it. It was definitely not cheap.

> But seriously: how did you get the domain `15.ai`?

it's an MIT project so I'm sure that was a factor

.ai domains cost a couple hundred bucks a year so domains are very available / not widely used by domain squatters (Its the country domain for the island of Anguilla, pop 15,000)

$65/year on Porkbun, not terrible.

In the case of text generation, we call this "Constrained Text Generation" and it is an active field of research. Without going into too many details (I have a paper out for review about this), it's pretty trivial to get "interactive control over how much or how little the AI generates" by a combination of filters on the LMs vocabulary, and effective selection of the various hyperparamaters in the decoder (top_p, top_k, temperature)...

Dall-E 2 has customization.

You can remove or add things etc.

And for GPT you can also specify more details.

Only a question of time until you can work with the ai on your art/thing.

There are ai models which keep track of context and others which generate a plan of actions.

AI is not a blackbox

OpenAI itself is a black box. Until I can reproduce their models or download them myself, and have unfettered access to them, it's just gatekept magic behind an API. So much for democratizing machine learning.

> So much for democratizing machine learning.

Unless this is a recent change, their mission isn't that:

> OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.

https://openai.com/about/

If that's your definition of black box, you need to specify this.

Blackbox would mean in a normal setting that no one knows how it works

You can fine tune GPT-3

Only if you're OpenAI

Fine tuning of GPT3 models is available via their public API. Costs credits, and you need to get their permission to use it in an actual application, but it’s not locked in a lab.

So “Only if you’re OpenAI” :)

If the weights were public, the community would figure out a way to fine tune it.

It’s not a matter of ‘figuring out’. The model supports fine tuning. It’s a core feature of the openai API. Running ‘fine tuned’ versions of GPT-3 that are created by customers is literally their SaaS model. They have examples in the documentation. Here: https://help.openai.com/en/articles/5528730-fine-tuning-a-cl...

I get that. But you don’t get the weights. Which means they limit what you can and can’t generate.

A friend of mine was going to build a writing assistant on top of GPT-3. She got a lot of encouragement from them. Then one day the social media storm hit OpenAI, and suddenly safety became a non-negotiable feature of their api. And along with that came the restriction “absolutely no products that can generate unlimited amounts of text, even if you can pay for the credits.”

Poof, no more business.

Imagine buying a keyboard that restricted what you could type.

All of these problems go away when you have access to the weights.

GPT-3 can be quite adaptive given prompt engineering and the uploading of sample files.

Have you used GPT-3 with any of the methods mentioned in the docs?

I’ve seen that GPT-3 can produce quite starkly different results when prompted differently and when samples have been uploaded.

I just used it to make spongebob squarepants say bad things.

This thing synthesizes dolphin squeaks? Wow!

Reply on news.ycombinator.com