5 ms·

So these "GPTs" are the combination of predefined prompts and custom API access? Not customly trained LLMs? If so, I guess you can make such a "GPT" on your ow

by mg 3y ago

So these "GPTs" are the combination of predefined prompts and custom API access? Not customly trained LLMs? If so, I guess you can make such a "GPT" on your own server and independent from a specific LLM service by using a prompt like ...you have available an API "WEATHER_API". If you need the weather for a given city for a given day, please say WEATHER_API('berlin', '2022-11-24') and I will give you a new prompt including the data you asked for... Or is there some magic done by OpenAI which goes beyond this?

jondwillis 3y ago
“GPTs” (terrible name) === Agents
- arcanemachiner 3y ago
  Agreed, but one is a recognizable trademark (I assume) and one is a generic catch-all term that means a lot of things and would probably be worse from a branding/marketing perspective.
- golol 3y ago
  Not even. An agent should primarily operate in some kind of self-prompting loop. Afaik xou can not specify complex and branching looping behaviors for GPTs.
  - zer0c00ler 3y ago
    ChatGPT and plugins were already agents. An agent just does stuff on somebody else’s behalf. Browsing, browses the Internet for the user, code interpreter creates and runs Python code for the user… What you describe are what’s called autonomous agents, and i agree that I also expected some interesting announcement there but probably too many security issues.
    - golol 3y ago
      Well, I would say that when people at the moment say agent they mean what you call autonomous agent. Browsing or code interpreter is just LLM + tool use. I think there is a really large difference in quality between a system that just interacts 1-2 times with some tool/API before giving an answer and one that runs in a loop with undefined length (until it decides to terminate). It's like the difference between programming without loops or recursion vs with them. Night and day.
      - jondwillis 3y ago
        Currently the loop results in compounding errors and attention failure. Even with reviews, grounding information, and veracity checks in place by other LLMs it still happens.
        golol 3y ago
        Yes, agents don't really work yet, I agree.
  - jondwillis 3y ago
    You can have a GPT/Assistant/Agent (OpenAI is using all three terms) call out to an API that has a looping agent behind it.
- ddmma 3y ago
  Exactly my point, then you can register .agency domain based on your area of ‘agent’ expertise. Sorry, but I won’t bite the marketplace as normally wipe the top agents with official ones, like apple did with native apps. It’s crowdsourced AGI
- dgreen899 3y ago
  they are trying to trademark "GPT" and "GPTs"..in order to do this...they need to use it widely and specifically...i don't believe that it has gone through yet..so in order for it be approved they will use this name...then prevent others from using it....thats the game...
BoorishBears 3y ago
If you want to be independent for academic/personal reasons, sure you can. If you want reasoning capabilities that Open Source hasn't matched before today, and I'm guessing just got blown out of the water on again today... there's no reason to bother.
- mg 3y ago
  You don't need to use an open source LLM for the approach I described. You can still send the prompts to OpenAI's GPT-4 or any other LLM which is available as a service.
  - BoorishBears 3y ago
    What other LLM will compete with GPT-4 Turbo (+ V)? At most you're hedging that Anthropic releases a "Claude 2 Turbo (+ V)": is complicating your setup to such a ridiculous degree vs "zero effort" worth it for that? If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways, absolutely no reason to roll your own on this.
    - dragonwriter 3y ago
      > If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways If things change down the road such that your tool (or a major potential downstream market for your tool) is outside of OpenAI's usage policies, the fact that you invested even a few developer-weeks into combining the existing open source tooling to support running your workload either against OpenAI's models or a toolchain consisting of one or more open source models (including a multimodal toolchain tied into image generation models, if that's your thing) with RAG, etc., is going to be a win. If it doesn't, maybe its a wasted-effort loss, but there's lots of other ways it could be a win, too.
      - BoorishBears 3y ago
        Let's go back to my very first comment. > If you want to be independent for academic/personal reasons, sure you can. If your goal is to be in business, or to get some sort of reach, or anything other than "have fun tinkering"... wasting developer weeks on "could be a win" is how to fail.
burningion 3y ago
That's the thing, we don't know. The lack of transparency for how the product works behind the scenes will most likely make it difficult to build something effectively.
- burningion 3y ago
  Actually, the documentation here is pretty great. Too late to edit my comment, but the transparency level here isn't too bad: https://platform.openai.com/docs/assistants/tools https://platform.openai.com/docs/assistants/tools
jatins 3y ago
Yeah I don't _think_ there is a lot "magic" in custom GPT. However creating something like this previously required a Jupyter notebook but now just...asking for it. Makes it accessible to 10x more people

by mg 3y ago

jondwillis 3y ago

“GPTs” (terrible name) === Agents

arcanemachiner 3y ago
Agreed, but one is a recognizable trademark (I assume) and one is a generic catch-all term that means a lot of things and would probably be worse from a branding/marketing perspective.
golol 3y ago
Not even. An agent should primarily operate in some kind of self-prompting loop. Afaik xou can not specify complex and branching looping behaviors for GPTs.
- zer0c00ler 3y ago
  ChatGPT and plugins were already agents. An agent just does stuff on somebody else’s behalf. Browsing, browses the Internet for the user, code interpreter creates and runs Python code for the user… What you describe are what’s called autonomous agents, and i agree that I also expected some interesting announcement there but probably too many security issues.
  - golol 3y ago
    Well, I would say that when people at the moment say agent they mean what you call autonomous agent. Browsing or code interpreter is just LLM + tool use. I think there is a really large difference in quality between a system that just interacts 1-2 times with some tool/API before giving an answer and one that runs in a loop with undefined length (until it decides to terminate). It's like the difference between programming without loops or recursion vs with them. Night and day.
    - jondwillis 3y ago
      Currently the loop results in compounding errors and attention failure. Even with reviews, grounding information, and veracity checks in place by other LLMs it still happens.
      - golol 3y ago
        Yes, agents don't really work yet, I agree.
- jondwillis 3y ago
  You can have a GPT/Assistant/Agent (OpenAI is using all three terms) call out to an API that has a looping agent behind it.
ddmma 3y ago
Exactly my point, then you can register .agency domain based on your area of ‘agent’ expertise. Sorry, but I won’t bite the marketplace as normally wipe the top agents with official ones, like apple did with native apps. It’s crowdsourced AGI
dgreen899 3y ago
they are trying to trademark "GPT" and "GPTs"..in order to do this...they need to use it widely and specifically...i don't believe that it has gone through yet..so in order for it be approved they will use this name...then prevent others from using it....thats the game...

BoorishBears 3y ago

If you want to be independent for academic/personal reasons, sure you can. If you want reasoning capabilities that Open Source hasn't matched before today, and I'm guessing just got blown out of the water on again today... there's no reason to bother.

mg 3y ago
You don't need to use an open source LLM for the approach I described. You can still send the prompts to OpenAI's GPT-4 or any other LLM which is available as a service.
- BoorishBears 3y ago
  What other LLM will compete with GPT-4 Turbo (+ V)? At most you're hedging that Anthropic releases a "Claude 2 Turbo (+ V)": is complicating your setup to such a ridiculous degree vs "zero effort" worth it for that? If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways, absolutely no reason to roll your own on this.
  - dragonwriter 3y ago
    > If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways If things change down the road such that your tool (or a major potential downstream market for your tool) is outside of OpenAI's usage policies, the fact that you invested even a few developer-weeks into combining the existing open source tooling to support running your workload either against OpenAI's models or a toolchain consisting of one or more open source models (including a multimodal toolchain tied into image generation models, if that's your thing) with RAG, etc., is going to be a win. If it doesn't, maybe its a wasted-effort loss, but there's lots of other ways it could be a win, too.
    - BoorishBears 3y ago
      Let's go back to my very first comment. > If you want to be independent for academic/personal reasons, sure you can. If your goal is to be in business, or to get some sort of reach, or anything other than "have fun tinkering"... wasting developer weeks on "could be a win" is how to fail.

burningion 3y ago

That's the thing, we don't know. The lack of transparency for how the product works behind the scenes will most likely make it difficult to build something effectively.

burningion 3y ago
Actually, the documentation here is pretty great. Too late to edit my comment, but the transparency level here isn't too bad: https://platform.openai.com/docs/assistants/tools https://platform.openai.com/docs/assistants/tools

jatins 3y ago

Yeah I don't _think_ there is a lot "magic" in custom GPT. However creating something like this previously required a Jupyter notebook but now just...asking for it. Makes it accessible to 10x more people