OpenAI announced these API updates 3 days ago:
- new function calling capability in the Chat Completions API
- updated and more steerable versions of
gpt-4andgpt-3.5-turbo - new 16k context version of
gpt-3.5-turbo(vs the standard 4k version) - 75% cost reduction on our state-of-the-art embeddings model
- 25% cost reduction on input tokens for
gpt-3.5-turbo - announcing the deprecation timeline for the
gpt-3.5-turbo-0301andgpt-4-0314models
gpt-3.5-turbowith the 16k context can now fit about 20 printed pages in its context. This is a game changer for summarization and documentation-based question answerint applications. I tried it in the API playground ant it works really well!Function calling also seems very useful for tool-using apps. No more crossing fingers and hoping the LLM will return a syntactically valid call!
The only thing is, haven’t wearied our lesson with reddit? Using these proprietary APIs are not to be trusted. I don’t think I would ever bud anything, even at a hobby or experimental level that relied on this.
Until someone trains a model (and it will happen) that match or outperform GPT4 and i can run i locally, i will use this to experiment and prototype random stuff that i find interesting ^^
Edit; a big difference here too is that Reddit just fetch data from an database, it’s just as expensive as the main site and app while GPT is generative and use quite lot of RAM and VRAM per action.
There are open source LLMs. I am not saying it is wrong to consume LLM as a service, the issue is openAI seems intent on intent on not being very open.
Ah, i think there is a miss understanding of their name, it’s not open as in open source, it’s open as in open research. They publish all their research for others to duplicate. And yes, there is other models out there, but on one as good as GPT4. Unless you have a computer with 640 ram, you can’t run it. So yeah, compared to fetching data from a database that could be done on a Raspberry and generating data that requires a monster computer, i understand that they wanna put a price on the API.
Well yeah, I do get that they conduct open research, but I still think it is disingenuous for the company to not release their source code. Or at least their LLM, if we are going to be somewhat charitable and allow that their specific tooling and API infrastructure should be proprietary so that they can maintain a business. There is no guaranteed that the code running on the other end of their API adheres to any of the research that they have revealed!
I’m not too worried though, because other LLMs and parameter sets have gone open source, so the cat is already out of the bag. I also don’t really believe in the commercial viability of LLMs either, because there is no way to automate the verification that that they are generating correct content so whatever.
