|
|
|
|
|
|
by mmaunder
894 days ago
|
|
|
A platform play like this assumes a captive market that is using the platform and who faces significant pain to leave. iPhone, Android and Windows had this. OpenAI does not because other APIs and open source models are hot on their heels. For many dev needs, an open source model can be used far cheaper, without vendor lockin, with more flexibility, free of censorship, and with lower latency. To make this play and sustain it, OpenAI has to continue to have a model that is significantly better than open source models in particular- I’d say perhaps as much as one order of magnitude. They are vulnerable. If Meta, for example, releases another Llama that is close to the current GPT state of the art, it may be a silver bullet. My world view is tainted. I’m profoundly opposed to AI via API, and I think supporting open source models is so important that it has the potential to affect the future of our species, given how important the AI revolution will be this century. |
|
|
Consider some points:
1. GPT-4 was trained in Summer 2022. OpenAI already has better models.
2. It's not just the model, but the infra around it: ChatGPT has tool use — image generation, web search, and API calls through "actions" — built in.
3. More infra: ChatGPT has a builtin moderation endpoint. This is not sexy, and although many of us hackers want uncensored AI, most applications will need some moderation.
4. ChatGPT has >100M users, and there is some lock-in already. ChatGPT users don't want their chats split over multiple apps.
5. Open source (and proprietary models like Grok) are fine-tuning of synthetic data generated by GPT-4. This fine-tuning process limits them to be sub-GPT-4 level.
6. Even the best open source models (eg Mixtral) are significantly worse than GPT-4. Their low cost makes them attractive, but if you believe, as I do, that sub-GPT-4 level models are just not that compelling, the open source AI ecosystem has a lot of catching up to do.
As long as there exist proprietary models that are an order of magnitude more capable than open source models, I expect the bulk of the value and usage will accrue to the ecosystems of the proprietary models. I do hope that at some point soon this changes. Maybe open source models achieve a flywheel of data, crowdsourced algorithmic optimization, and perhaps some form of efficient distributed training on consumer hardware is possible. This would be awesome, IMO.