[minimaxir]

GPTs are essentially a proprietary moat-building of several open-source workflows, most especially system prompt engineering, tool-usage, and RAG.

Unfortunately there's no magic open-source solution since there's a lot of moving pieces involved that are bespoke to a given use case, and the ones that claim to be magic are libraries like LangChain, which aren't.

[eshack94]

This is the most accurate and clear-eyed take I've seen on GPTs so far. They might be useful, but they're not magic and they're intended to enhance OpenAI's moat-building operations to make it harder for people/companies to walk away from the platform amid future competition.

[visarga]

I believe GPTs are an attempt by OpenAI to generate training data. How can you get data at level N+1 when you have a model at level N? You give it more resources - more tokens (CoT), more dialogue rounds, code execution, web search, local KB, human-in-the-loop. A model with feedback from human and tools can do so much more. And by training on this data they can incorporate these skills in the next generation. It's like RLHF in the sense that the training data contains portions generated by the model itself (specifically model errors) and feedback. It's on policy data, generated with the involvement of the model, not something you can scrape from the web.

Let's do an estimation - if they have 100M users and each of them generates 10K tokens in a month, that's 1T tokens per month. In a year they have generated 12T tokens, which is very close to the GPT-4 training set size of 13T. Looks like they can generate serious data with this method. They don't even need to train directly on it, they could rewrite it as high quality training examples, without copyright and PII risks, because LLMs are great at rewriting and rewording and MS has already shown that synthetic data is better.

Google lost the start and they don't have the human-AI chat logs OpenAI sits on. So they are trying to do the same trick but without the human in the loop. Hence the declarations that Gemini will use some techniques from AlphaZero. They are teaching models by feedback too.

[david_shi]

There really should be open source versions.

It's perplexing why there aren't, especially when the individual components are relatively straightforward (code interpreter, RAG, search, function calling, image generation).