For my own personal use, I tend to stick with the best available model, so Opus or GPT-4 But now that I'm building user-facing features on them both cost and speed are more important, so I'm being more thoughtful about which model to use
@simonw
-
OpenAI’s missing middle tier between GPT-3.5 and GPT-4
By
–
Surprisingly the OpenAI lineup seems to be missing that middle piece – something that sits between the cheap and fast gpt-3.5-turbo and their GPT-4 class models The difference between GPT-4 and gpt-4-turbo doesn't quite feel like Opus to Sonnet
-
Three Sizes of AI Models: Haiku, Sonnet, and Opus Comparison
By
–
Something I have learned from Claude 3 is that I really like my models in three sizes: the fast one (Haiku), the slow but "best" one (Opus) and the hard to define but spectacularly useful one in the middle (Sonnet) Mistral 7B / Mixtral 8x7B / Mixtral 8x22B feels similar to that
-
OpenAI Beta Feature Review and Analysis
By
–
OpenAI have a beta feature for this at the moment – I've actually not been very impressed with it so far, I have some notes on how it works here
-
OpenAI AI Functionality Speculation Without Official Confirmation
By
–
That's pretty much my guess too, but I can't confidently tell other people that's what it's doing because I don't have confirmation from OpenAI
-
OpenAI’s vague data usage policies raise transparency concerns
By
–
OpenAI have been infuriatingly vague about what "using ChatGPT conversations to help improve our models" actually means They have various opt-out options even for free usage these days but it's hard to keep track of where they are and how they work
-
Understanding How Large Language Models Actually Work
By
–
It's not an unreasonable mental model to form to be honest – enormous weird blobs of vector floating point matrices and token embeddings are hardly an obvious way that this technology might work
-
Fine-tuning vs RAG: Teaching AI Models Effectively
By
–
This could also explain why so many people instantly assume that "fine-tuning" a model is the obvious right way to teach it new information, as opposed to using more effective but less obvious techniques like RAG (not helped by that being a pretty terrible acronym)
-
Lack of transparency in AI model training data usage
By
–
Right, the most frustrating thing about this is that the complete lack of of transparency about how training works (and how the data is used) means it's impossible to confidently state how it all really works
-
Common Misconceptions About How AI Models Learn
By
–
I wonder how common it is for people to confidently hold an inaccurate idea of how AI models work where they believe that anything they show the model is instantly memorized and added to its "knowledge" of the world