Most probably the difference is - you made you assistant for your specific task and OpenAI made assistants for general usage. At the same time, Assistants are in beta, it is clearly stated, so I really don’t understand why people are trying to assess it as if it was released as a stable version.
Ales is actually not wrong about the cost of using Assistants. In beta they are really not optimized for cost and can eat A LOT of tokens in some cases without giving the same type of value. And you can do nothing with it for now. Probably once they release in prod, it will be much better.
every criticism is constructive when represents somebody’s experience. what about this: if you don’t vectorize an ordinary string as it is (which is an ANSWER) but you push it through a model to generate a series of relevant QUESTIONS for that ANSWER and you vectorize them instead of the original string and then you use that QUESTION vectors to find a row with ANSWER (because user query is usually a QUESTION) you get more than 30% more accurate cosine similarity (even euclidean distance) results? It almost eliminates the need to add side strings to ensure context match.
using this simple trick which I invented and watching the token cost of your assistant makes somebody else comical… I didn’t find anything about this technique anywhere so maybe it is considerable to use it in your Assistant API framework because I’m sure it does make sense and it works for me pretty well.
sorry from a dumb graphic designer
1 Like
Guys, I appreciate this sweet conversation, but let’s not walk this slippery path as we are on a brink of going away from the topic into the personal judgements zone 
oh, im still warming up 
loong night coding
You’re describing a somewhat more complicated and convoluted version of HyDE.
And I, for one, am not mad at OpenAI. In fact, I say, keep doing what you’re doing. So long as they focus on general usage, they create a huge market for developers in specialized niches. And, if OpenAI decides to pull the plug on open-ended development as some have suggested? Well, I’ve got a nice spot just waiting for Gemini Ultra when it arrives.
So, all good on this end.
2 Likes
Getting back to the OP question, I love this question, and think about it a lot. Even if we ignore the physical world, what do I want from a computer literate assistant? Which parts of that are hardest?
How about one I’ve been working on: Generate the most promising set of miRNA, mRNA, methylated-DNA, and DNA expression assays to perform for early-stage lung cancer detection.
Finding relevant information
Accounting for newness of information
Credibility of source
Accounting for inconsistency
digesting information into usable form
Generating hypotheses and validating them
Managing obscure terminology
and mostly
Generating a real answer that goes beyond the typical LLM near-contentless gloss
or, to put it another way, as @StoicJester said, LLMs are near useless if you are at the bleeding edge.
1 Like
I believe the process you explained is already realistic using agent frameworks. But that’s on high level. If we go deeper into details, we see all those yet unsolved problems like retrieval, controllability, etc.
The biggest problem with these frameworks is that they are not production ready and are more for exploration and learning.
2 Likes