Hacker News Reader

Comment by danpalmer | Hacker News Reader

> SlickGPT allows you to run your own local ChatGPT instance

I find it interesting that so many projects claim to be "run your own ChatGPT", when they're in fact "run your own web UI". Are people really so ignorant of how this works that they're equating these concepts, or believing that the web UI is anything more than a razor thin veneer?

I guess maybe people use ChatGPT to refer to the interface, and GPT-N to refer to the models, but that's not very accurate given the amount of tuning, it's much more accurate to say that ChatGPT is a productised GPT instance, with a web UI.

View on HN

It's just a hot marketing strategy, it has a totally different reception outside of HN. There are already people making 5k+ MRR with apps that are just a chatGPT wrapper that passes user input with some additional context and prompts.

I could easily see a wave of small startups that are just a few inputs and the context needed to generate documents people currently write by hand. Like subpoenas, doctors letters to insurance companies, etc. You can stand one up with a bit of industry knowledge and a couple 100 LOC.

I expect this, too. Scammers are gonna scam.

> I find it interesting that so many projects claim to be "run your own ChatGPT", when they're in fact "run your own web UI". Are people really so ignorant of how this works that they're equating these concepts, or believing that the web UI is anything more than a razor thin veneer?

Perhaps, but also there is a huge incentive to blur that line, obviously from the developers of these UIs, but also from the people at media outlets covering them. If they told the truth, it wouldn't be as big of a story. And I'm just waiting for the "7 year old makes their own GPT" story. For a long time now, the overwhelming majority of tech journalism has been just uncritically rephrasing press releases.

For it to be good, it should also have a specific system prompt and chat history, which is fed to the model. With langchain this isn’t too difficult to build, but it’s not just a web ui.

This one doesn’t appear to use a custom prompt, so I’m guessing it doesn’t perform as well as ChatGPT, but it does look like it passes message history.

SlickGPT passes message history including a "context" or "system" message which it clearly explains to the user at the start of every chat. This is probably the most important thing the official chat.openai.com client fails to explain or expose to users directly.

But you can do all that with the OpenAI Playground for the chat API, and it supports GPT-4. I don't really see what your service adds to that.

You can use the same model that ChatGPT uses through the API, and it isn't called 'ChatGPT' it's called 'GPT-4'. So at least officially ChatGPT refers to the webUI not the model.

I'm not sure that's true though. The first line of the ChatGPT blog post is:

> We’ve trained a model called ChatGPT which interacts in a conversational way

It really seems that GPT-3/3.5/4/etc are the generic models, and that ChatGPT is the fine tuning, a safety layer, and so on. This is also why ChatGPT and Bing don't provide the same or even remotely equivalent answers despite both running on GPT-4.

As far as I know the current chatGPT running on the chat.openai website is specifically gpt3.5-turbo .

Its like when everything became an 'app' all of a sudden.

ChatGPT Plus is like $10/mo but bring your own interface and its fractions of a penny per API call. So for most, this will be cheaper.

In this case it is.

Lot of others posted are other models that people could use.

Reply on news.ycombinator.com