4 ms·

Built this on Sunday as a fun side project. It's basically a fancy UI on top of GPT-3 (with some "prompt engineering"). The idea is to explain terms and concept

Built this on Sunday as a fun side project. It's basically a fancy UI on top of GPT-3 (with some "prompt engineering"). The idea is to explain terms and concepts in simple terms. It was quite simple to build (granted I've been building web apps for a long time) and I expect we will see tons of those "UI on top of model" apps in coming months and years. Happy to answer technical questions. Edit: Oops, already hitting API rate limits :( Going to see if there's anything I can do about that. Edit 2: Actually displaying an error message now when the rate limit is reached (instead of "looping" forever). I will keep figuring out if there are ways to increase the limit but it looks like it would have to wait 2-7 days anyways :(

jedberg 4y ago
Heh, one of the other commenters here hacked your system to reveal your prompt engineering. :) I agree with you, I think we'll see a lot of web apps that are UIs on top of models, which is probably a good thing!
- esskay 4y ago
  'Hacked' probably isn't the right word.
  - jedberg 4y ago
    I think it's exactly the right word. They found a flaw in the system and exploited it.
  - upwardbound 4y ago
    "Influenced" might be a fair word though. The ability to exert influence, any influence at all, can some times build up to a potent attack in unforeseen ways. Hell, look at Rowhammer and how that works. It's nuts. I figure there will be some use cases where it would be super bad to let an attacker exert any influence on your system whatsoever, especially if paired with simultaneous attacks that synergize with a prompt attack, like for example if a prompt attack + user database can be combined with a side channel analysis like thermal data or something, and used to learn information about private keys on the server. My overall point is that many prompt attacks may not meet the bar to be a "hack" but they certainly constitute an attacker being able to exert an influence on the running server, which can potentially allow sandbox escape when paired with other vulnerabilities. Which to me, is potentially pretty scary given the bad possibilities of what evil rogue AI can do when it is in the role of the offensive hacker, and can battery away at sandbox defenses 24/7/365 until it finds a sandbox escape. Meanwhile, the AI may cover its traces by falsifying the records of computational resources used. It is to be truly a frightening thought and one I hope we can all work hard to avert. We can do our part by being extremely EXTREMELY diligent with cyber-security, hardware security, software supply chain traceability & governance, and so on.
andreygrehov 4y ago
Is ChatGPT API publicly available?
- yuck39 4y ago
  ChatGPT is not, GPT-3 is
- andai 4y ago
  No but GPT-3 is, and with the right prompts can give very similar output. (See the Chat preset in OpenAI playground.)
andai 4y ago
I got a similar app (preprompt + small user input) up and running recently and it was surprisingly easy! I would encourage anyone who's considering it to give it a try. In all honesty I just Frankensteined OpenAI's Nodejs / React template into two Railway template projects (a TS+Express and a Vite+TS+Preact template). (Railway builds and runs web apps for you—I do not enjoy being a sysadmin so this is ideal for me!) This "lazy" approach was mostly due to business time pressure, but I found it surprisingly refreshing. Normally I'd spend weeks making everything from scratch.
mattfrommars 4y ago
Are you saying for a given user input, the application concatenate the input with something that makes GPT3 API to explain in "ELI5" form?
- olalonde 4y ago
  Yes, that's exactly how it works.
kgeist 4y ago
>It was quite simple to build A few months ago (before ChatGPT) I built an IRC bot using GPT3 in like an hour, the API is pretty straightforward to use. Most of the time was spent adding conversation memory and debugging a few edge cases. The result was pretty impressive, considering that there was almost no effort on my part.
- giobox 4y ago
  95 percent of the ChatGPT product basically sat in the OpenAI API developer dashboard for a year too and the market largely didn't notice. Even before chatGPT, you could converse with GPT3 just using the API dash in a browser in a very, very similar fashion to chatGPT. I thought it interesting Sam Altman's comments that he kinda thought the market would just make chatGPT itself and OpenAI could just sit back having made the APIs available, but given the cost per query on a service that makes no revenue, personally I'm not all that shocked it took an internal product with large marketing spend etc to gain the exposure chatGPT has for their API services.

jedberg 4y ago

Heh, one of the other commenters here hacked your system to reveal your prompt engineering. :) I agree with you, I think we'll see a lot of web apps that are UIs on top of models, which is probably a good thing!

esskay 4y ago
'Hacked' probably isn't the right word.
- jedberg 4y ago
  I think it's exactly the right word. They found a flaw in the system and exploited it.
- upwardbound 4y ago
  "Influenced" might be a fair word though. The ability to exert influence, any influence at all, can some times build up to a potent attack in unforeseen ways. Hell, look at Rowhammer and how that works. It's nuts. I figure there will be some use cases where it would be super bad to let an attacker exert any influence on your system whatsoever, especially if paired with simultaneous attacks that synergize with a prompt attack, like for example if a prompt attack + user database can be combined with a side channel analysis like thermal data or something, and used to learn information about private keys on the server. My overall point is that many prompt attacks may not meet the bar to be a "hack" but they certainly constitute an attacker being able to exert an influence on the running server, which can potentially allow sandbox escape when paired with other vulnerabilities. Which to me, is potentially pretty scary given the bad possibilities of what evil rogue AI can do when it is in the role of the offensive hacker, and can battery away at sandbox defenses 24/7/365 until it finds a sandbox escape. Meanwhile, the AI may cover its traces by falsifying the records of computational resources used. It is to be truly a frightening thought and one I hope we can all work hard to avert. We can do our part by being extremely EXTREMELY diligent with cyber-security, hardware security, software supply chain traceability & governance, and so on.

andreygrehov 4y ago

Is ChatGPT API publicly available?

yuck39 4y ago
ChatGPT is not, GPT-3 is
andai 4y ago
No but GPT-3 is, and with the right prompts can give very similar output. (See the Chat preset in OpenAI playground.)

andai 4y ago

I got a similar app (preprompt + small user input) up and running recently and it was surprisingly easy! I would encourage anyone who's considering it to give it a try. In all honesty I just Frankensteined OpenAI's Nodejs / React template into two Railway template projects (a TS+Express and a Vite+TS+Preact template). (Railway builds and runs web apps for you—I do not enjoy being a sysadmin so this is ideal for me!) This "lazy" approach was mostly due to business time pressure, but I found it surprisingly refreshing. Normally I'd spend weeks making everything from scratch.

mattfrommars 4y ago

Are you saying for a given user input, the application concatenate the input with something that makes GPT3 API to explain in "ELI5" form?

olalonde 4y ago
Yes, that's exactly how it works.

kgeist 4y ago

>It was quite simple to build A few months ago (before ChatGPT) I built an IRC bot using GPT3 in like an hour, the API is pretty straightforward to use. Most of the time was spent adding conversation memory and debugging a few edge cases. The result was pretty impressive, considering that there was almost no effort on my part.

giobox 4y ago
95 percent of the ChatGPT product basically sat in the OpenAI API developer dashboard for a year too and the market largely didn't notice. Even before chatGPT, you could converse with GPT3 just using the API dash in a browser in a very, very similar fashion to chatGPT. I thought it interesting Sam Altman's comments that he kinda thought the market would just make chatGPT itself and OpenAI could just sit back having made the APIs available, but given the cost per query on a service that makes no revenue, personally I'm not all that shocked it took an internal product with large marketing spend etc to gain the exposure chatGPT has for their API services.