I was annoyed by the default OpenAI ChatGPT slow-typing animation and logging me out every day, so I built a new UI client on top of its API.
It's a static web app, you can host it yourself, data is stored locally on your browser, API requests are made directly to OpenAI API without any middle server.
It has some more convenient features that make your that experience way better. Like search chat history, prompt library, integrations, etc.
I'm also building a personal ChatGPT client for a similar reason as the OP, which is to address the issues of buggy and slow UX, among others. As you mentioned, many people are currently working on projects using ChatGPT API. In fact, in the past 24 hours, around 12 new apps that have been showcased on OpenAI Discord channel alone. Some of these apps are open source, which makes them a great source of learning and inspiration for me.
I know mate, haha. I imagine someone can do much better than I did in a few minutes, but yeah use the UI that already exists instead of sprinkling electron on something text based.
If anyone fancies a tinker I imagine future editions of a CLI AI having a --context to include former prompts*, --stfu|--concise to make it respond concisely (as I was originally just responding "I know mate, haha" I also need this flag adding), probably a --model= to override that bit too, maybe even a --preamble=./prompt.txt
... it definitely needs a --json flag to prompt and enforce "I only want json and I'll parse it out of the response myself if I need to"
* Not just keeping a buffer and re-sending it, I believe context asks GPT to summarise the conversation so far and then uses the output of that for context memory. It'd need some tinkering.
I'll get round to it myself eventually, for now being able to ask little things like "is it defence or defense" and "what company owns the coco pops brand" is doing me just fine
I did realise I can now potentially pipe things into and out of GPT too, but have yet to come up with a use other than outputting through TTS e.g.
> I was annoyed by the default OpenAI ChatGPT slow-typing animation
I’d assumed ChatGPT’s slow-typing was actually the model’s real-time output, streamed to the browser token-by-token. When I’ve made GPT3 API calls it’s worked like this, with streaming switched off I might need to wait ~20 seconds for a long response back.
I just tried your UI, and it’s so much faster! Was my assumption about real-time output wrong?
Great project, looking forward to using it more :)
Yep, I was surprised I about the response time too. I'm not sure what's going on under the hood, maybe OpenAI did something with the stream output that causes the slowness.
Perhaps it is simply a way to limit the number of queries? If you force people to wait 10s for an answer they aren't writing new questions all the time.
For me it's been consistently the case that after 2-3 messages, the animation would start to visibly lag, and eventually (on 3rd or 4th reply) slow to a halt, sometimes mid-sentence, completely bricking the session. It's been like this since ChatGPT was first opened to public. I initially assumed it's just the usual case of a bloated SPA, but now I think maybe I'm being throttled on purpose?
ChatGPTs paid version is extremely fast, and this is using the paid API. The paid API feels slower than the paid UI, but this may be the difference between streaming results and getting everything in one payload.
I think the slow responses are the real generation rate when you time share the GPU. If you are willing to let GPUs occasionally sit idle though, you can probably get a faster generation rate by dedicating more of the GPU time to that one client.
Note for anybody else wondering, it's not open source, you have to download a binary. More FAQ by clicking the Gift (?) icon at the top and scrolling down.
No. It's real. According to dang, showHN is suppose to be a "safe space". You could make absolute crap, everyone is suppose to be positive about it. Dang will flag people who don't do this.
Nothing dang can do though is hundreds of people hate it though. Things with actual security issues for example he'll likely just be quiet about it.
That is unfortunate. You would think that the intent behind posting ShowHN is to get constructive feedback but based on what you described, it appears that the expectation is to make those creators feel good instead.
I'm not defending anyone here but just strace the program if you're really worried about it. If you don't know what strace is: it exports every syscall the program has made, so you'd see if it's doing anything suspicious.
Well I don't see why not, also I don't see which issues if you're not using it as you couldn't encounter any to begin with. Being closed source is not an issue.
It is open source. Your browser retrives all the JS and html sources. Even if it were somehow closed source, it doesn't really matter. You can view web requests and other, and see for yourself.
All logic is clientside in the downloaded JS and HTML. After the initial page load, it should send nothing to the backend of the website, only API calls to chatGPT API itself. There is no actual backend with logic, it only serves static HTML+JS+assets.
Theoretically, you should be able to block any outgoing connections to the website domain (after it loads all necessary HTML and JS), and it should still work fine. Saying "theoretically" because it might attempt delay-loading some static assets, but you get the idea. No idea how this specific website functions, as I haven't looked deep into it yet, but that's the principle behind "static websites" like this one.
The ability to reverse engineer the source through the browser doesn't make something qualify as open source anymore than reverse engineering a binary with some other tool.
I can view web requests but let's say on the 365th web request it sends my API key to an unknown location. How would I know if this was closed source?
Well done! I wonder if it's worth it to also hide a "select model" option somewhere, since we're putting in our own api keys i'd be interested in trying out other models even though they arent "chat" models.
Also would be great to see which messages are included in the computation, since chat gpt limits total characters per request to 4K? so inevitably some past messages will be rendered useless.
Thank you for the suggestion! I'll add an option to select the model soon.
As per your question, currently all the previous messages of the same chat will be included in the request. If you hit the 4k token limit, there will be an error message. I'll look into improving this too.
This looks very nice, lots of great extra features over the vanilla UI.
One possible improvement: I had accidentally mistyped my API key and it popped up an HTTP 500 error instead of going "That doesn't seem to be a valid API key, could you make sure you pasted it in correctly" or something...
Cool project! Thank you!
I would request a new feature, a page/document upload (in parts if it is big) to update the right\correct context. For example the current documentation for the framework I use version 27, but chatgpt only knows version 17 of it and I have to upload pieces of documentation every time to get correct updated code snippets.
That would be great, something like a button\action with a promt that would load and split if needed large amounts of text and promt it in conversation. (maybe add format some parsing, like markdown or plain text)
Great work on the project and thank you for sharing it! I do have a feature request though - would it be possible to add a page/document upload function that allows for updating the correct context?
Looks amazing - congrats! I haven't checked the API in detail but how do you keep the context of the conversation (awareness of previous questions and answers)? Do you send back the whole conversation on every request?
Shameless plug: I just don't like closed-source software; here is my attempt at re-implementing ChatGPT's UI as a desktop app with the ChatGPT API, and it is open source:
https://github.com/chatgptui/desktop
I've implemented most of features of ChatGPT, text-to-speech via Azure, and keyboard shortcuts such as tab/shift+tab.
Edit: However, as typingmind is earning more than $1000, it should have more features and better maintenance.
I mean, - ChatGPT's UI is closed-source, so I implemented a UI that looks almost identical to ChatGPT, - While ChatGPT's AI model is closed source, it can be accessed through the OpenAI API.
Not to mention, there are already OpenAI API clones which handle running against any of the open-source models. When they get more powerful and act as an actual competitor, it'll be very nice to have open-source components throughout the stack.
It has the one most important feature ChatGPT is missing: add to homescreen/install. Not sure why they haven't added it yet. Thank you for making this!
One feature I would love is passing a prompt (or even a conversation) using queryparams. It would help for the integration with other tools.
Another feature I'd love is to exploit the 'system' agent. From the documentation, there are three possible agents : User, Assistant and System. My understanding is that system can be exploited to drive the expected general behavior of the assistant and seems very useful to that regard.
Last thing, if it was open-source I could clone it and that would dismiss any fears from having my API key stolen. But that's up to you of course.
Can anyone recommend a UI clone as close to the regular ChatGPT as possible, but built on the API? The API traffic is sent directly into Azure while the regular UI is sent via CloudFlare. Vastly prefer the former reading my conversations than the latter
ChatGPT has this annoying habit of crashing or having extended downtime.
If this UI can capture my query in the case of such a crash and keep retrying till I get the answer, and then notify me (browser, email, etc) - that would be great.
94 comments
[ 4.0 ms ] story [ 159 ms ] threadIt's a static web app, you can host it yourself, data is stored locally on your browser, API requests are made directly to OpenAI API without any middle server.
It has some more convenient features that make your that experience way better. Like search chat history, prompt library, integrations, etc.
Would love to hear all the feedback!
Can already see the whisper model makings its way through the product :)
It's basically doing the same thing. I found that using API keys is also significantly cheaper than using the subscription with moderate usage.
Congrats on the release.
Do you then reference those open source apps?
https://gist.github.com/cohan/7d2abfa56b2dfdaae98bf70e84c622...
All of these UIs when there's a perfectly good command line right there! The CLI is always the best way for text.
> ai what is that song with moms spaghetti
"Lose Yourself" by Eminem
Sorted.
different flavours: 'fix' acts as `thefuck` replacement in 30 lines of python
Something like
> ai give me an index match formula to link between sheets New and Current
Then I'll tweak the output as needed to match reality. Context would probably get in the way more than it helps in this use case.
Don't expect perfection from an AI assistant, just expect it to be a little faster than tabbing to a browser or figuring it out yourself.
In Stellaris terms: https://img.imgy.org/D76l.png
If anyone fancies a tinker I imagine future editions of a CLI AI having a --context to include former prompts*, --stfu|--concise to make it respond concisely (as I was originally just responding "I know mate, haha" I also need this flag adding), probably a --model= to override that bit too, maybe even a --preamble=./prompt.txt
... it definitely needs a --json flag to prompt and enforce "I only want json and I'll parse it out of the response myself if I need to"
* Not just keeping a buffer and re-sending it, I believe context asks GPT to summarise the conversation so far and then uses the output of that for context memory. It'd need some tinkering.
I'll get round to it myself eventually, for now being able to ask little things like "is it defence or defense" and "what company owns the coco pops brand" is doing me just fine
I did realise I can now potentially pipe things into and out of GPT too, but have yet to come up with a use other than outputting through TTS e.g.
> ./ai is it defence or defense | festival --tts
I’d assumed ChatGPT’s slow-typing was actually the model’s real-time output, streamed to the browser token-by-token. When I’ve made GPT3 API calls it’s worked like this, with streaming switched off I might need to wait ~20 seconds for a long response back.
I just tried your UI, and it’s so much faster! Was my assumption about real-time output wrong?
Great project, looking forward to using it more :)
Note for anybody else wondering, it's not open source, you have to download a binary. More FAQ by clicking the Gift (?) icon at the top and scrolling down.
Sarcasm?
Nothing dang can do though is hundreds of people hate it though. Things with actual security issues for example he'll likely just be quiet about it.
Theoretically, you should be able to block any outgoing connections to the website domain (after it loads all necessary HTML and JS), and it should still work fine. Saying "theoretically" because it might attempt delay-loading some static assets, but you get the idea. No idea how this specific website functions, as I haven't looked deep into it yet, but that's the principle behind "static websites" like this one.
I can view web requests but let's say on the 365th web request it sends my API key to an unknown location. How would I know if this was closed source?
Your example is valid, but it's like asking "how can I know that my iPhone doesn't record me while I'm sleeping?"
https://github.com/cogentapps/chat-with-gpt
https://chatwithgpt.netlify.app/
Also would be great to see which messages are included in the computation, since chat gpt limits total characters per request to 4K? so inevitably some past messages will be rendered useless.
Overall well done!
As per your question, currently all the previous messages of the same chat will be included in the request. If you hit the 4k token limit, there will be an error message. I'll look into improving this too.
One possible improvement: I had accidentally mistyped my API key and it popped up an HTTP 500 error instead of going "That doesn't seem to be a valid API key, could you make sure you pasted it in correctly" or something...
That would be great, something like a button\action with a promt that would load and split if needed large amounts of text and promt it in conversation. (maybe add format some parsing, like markdown or plain text)
Edit: However, as typingmind is earning more than $1000, it should have more features and better maintenance.
One feature I would love is passing a prompt (or even a conversation) using queryparams. It would help for the integration with other tools.
Another feature I'd love is to exploit the 'system' agent. From the documentation, there are three possible agents : User, Assistant and System. My understanding is that system can be exploited to drive the expected general behavior of the assistant and seems very useful to that regard.
Last thing, if it was open-source I could clone it and that would dismiss any fears from having my API key stolen. But that's up to you of course.
https://github.com/kierangilliam/chatwithme.chat
Available at: https://www.chatwithme.chat/
If this UI can capture my query in the case of such a crash and keep retrying till I get the answer, and then notify me (browser, email, etc) - that would be great.
https://www.reddit.com/r/OpenAI/comments/11ef8ea/superpower_...
[1] https://twitter.com/adamdangelo/status/1631714221171023873
Add serpapi on top of that, and you might just become my phone’s most used app.