Hello. I made this. I'm a bit shocked at how many people are getting things wrong and taking things out of context, so perhaps I could answer some questions.
To clarify: I have not been with MIT in years. I was paid the minimum hourly rate (roughly $14 an hour) to work on a related project during my undergraduate years, which eventually evolved into this project years down the road. (In fact, I had to pay for my own compute to get my work started - MIT never offered me any credits.) Everything else has been paid out of pocket since then. Yes, it does indeed cost several thousands of dollars a month - that is not an exaggeration. This has been optimized so many times that the technology needs to improve first before I can cut down on costs.
The timer in the "TOS" was put there in hopes that people could understand where I was coming from in regards to misappropriating this kind of research. I did not expect people to get this riled up over a 10 second timer. (Especially not on Hacker News, of all places...)
Edit: I suppose it makes sense to include other information in this post so that others don't have to go hunting for my comments in this thread:
- >This is unrelated but what's with the fascination with HN users and My Little Pony? I've noticed this on a lot of posts in the past few months.
- Twilight Sparkle's voice is indispensable in getting emotional contextualizers to work properly. The logo and profile picture is an homage to that fact.
- >But seriously: how did you get the domain `15.ai`?
- I purchased it. It was definitely not cheap.
- >They named their tts "Deep Throat"? Why would you?
- It was a suggestion from a Twitter user, and I found it clever.
- >I heavily doubt it's "several thousands of dollars"...
- It is indeed several thousands of dollars a month. I can show you AWS invoices, if you're skeptical. Just send me an email and I'd be happy to show proof.
- >The disclaimer is a little ironic considering the site owner doesn’t own the model (MIT does) and doesn’t own the training data (the various shows and games do)
- I'm sorry to tell you that I do, in fact, own my own model. I have not been with MIT in years.
I have not been with MIT in years. I had a successful exit not too long after graduation, and I've been spending most of my earnings on this project.
As an undergrad, I was completely broke. I figured that keeping the project free to use was the best thing I could possibly do with my research as I continued to work on it.
>Can't you continue/do your research without a public website?
Yes, but the website has multiple purposes. It serves as a proof of concept of a platform that allows anyone to create content, even if they can't hire someone to voice their projects.
It also demonstrates the progress of my research in a far more engaging manner - by being able to use the actual model, you can discover things about it that even I wasn't aware of (such as getting characters to make gasping noises or moans by placing commas in between certain phonemes).
It also doesn't let me get away with picking and choosing the best results and showing off only the ones that work (which I believe is a big problem endemic in ML today - it's disingenuous and misleading). Being able to interact with the model with no filter allows the user to judge exactly how good the current work is at face value.
Despite others here, I personally certainly think this is admirable. I've played with your models a long time ago with some colleagues of mine and we were all shocked how good it was, and that it was free.
I'm no stranger to passion projects, I have a lot of respect for people like you. This is great stuff.
Thank you for the kind words. I know that HN is a tough crowd to please (I myself as well), so I hope that my next update will be well worth working for.
The Rise model in particular is amazingly good quality. I pranked a friend with some text from her a few minutes ago and he chastised me for wasting my money on hiring voice actors just to troll him.
Do you cache results from this (especially the random samples provided)? It seems to be regenerating those, which might be expensive if lots of people are using the same prompts
Edit: also wanted to thank you for the Chell voice, it sounds completely true to life to me! (minus some jumping noises)
Yes, I do. For the past three years, I have done nothing but work on this project nonstop. I've been working on massive improvements (that some have pointed out in this thread) that I've been stuck on for the past several months, but I'm getting close to finishing that up.
I don't feel comfortable publishing or releasing anything until I know for a fact that I can make no further improvements. It's not out of corporate greed or anything like that - I'm just really paranoid about getting out the best work possible.
Respectfully, the perfect is the enemy of the good, and it’s entirely reasonable to publish what you have now. If later you make further improvements, you can simply publish again.
You're completely correct, but I'm afraid this is more of a personal problem. I know I'll never be able to forgive myself if I figured out a solution to one of the more obvious problems with the model after I've already published it. I'd just be far more comfortable being happy with my own work before I release it to the wild. I know that this is selfish, and I apologize.
To clarify: I have not been with MIT in years. I was paid the minimum hourly rate (roughly $14 an hour) to work on a related project during my undergraduate years, which eventually evolved into this project years down the road. (In fact, I had to pay for my own compute to get my work started - MIT never offered me any credits.) Everything else has been paid out of pocket since then. Yes, it does indeed cost several thousands of dollars a month - that is not an exaggeration. This has been optimized so many times that the technology needs to improve first before I can cut down on costs.
The timer in the "TOS" was put there in hopes that people could understand where I was coming from in regards to misappropriating this kind of research. I did not expect people to get this riled up over a 10 second timer. (Especially not on Hacker News, of all places...)
Edit: I suppose it makes sense to include other information in this post so that others don't have to go hunting for my comments in this thread:
- >This is unrelated but what's with the fascination with HN users and My Little Pony? I've noticed this on a lot of posts in the past few months.
- Twilight Sparkle's voice is indispensable in getting emotional contextualizers to work properly. The logo and profile picture is an homage to that fact.
- >But seriously: how did you get the domain `15.ai`?
- I purchased it. It was definitely not cheap.
- >They named their tts "Deep Throat"? Why would you?
- It was a suggestion from a Twitter user, and I found it clever.
- >I heavily doubt it's "several thousands of dollars"...
- It is indeed several thousands of dollars a month. I can show you AWS invoices, if you're skeptical. Just send me an email and I'd be happy to show proof.
- >The disclaimer is a little ironic considering the site owner doesn’t own the model (MIT does) and doesn’t own the training data (the various shows and games do)
- I'm sorry to tell you that I do, in fact, own my own model. I have not been with MIT in years.