Makes me wonder if they stumbled onto some emergent behavior with the new Assistants API. You can have an Assistant thread spawn other Assistant threads, each with their own special instructions, plus the ability to execute custom code, reach out to the internet for other data and processing as needed, etc. Basically kicking off a hive mind that overcomes the limitations of a single LLM.
Except this was entirely possible with API, and the dead stupid obvious thing to do, even as far back as OG ChatGPT (pre-GPT-4). Assistants don't seem to introduce anything new here, at least not anything one could trivially make with API access, a Python script, and a credit card.
So I don't thing it's this - otherwise someone would've done this long time ago and killed us all.
Also not like all the "value adds" for ChatGPT are in any way original or innovative - "plugins" / "agents" were something you could use months ago via alternative frontend like TypingMind, if you were willing to write some basic JavaScript and/or implement your own server-side actions for the LLM to invoke. So it can't be this.
I'd agree that what is available publicly isn't anything that hasn't been in wide discussion for an agent framework since maybe ~march/april of this year, and many people had just hacked together their own version with an agent/RAG pipeline and API to hide their requests behind.
I'm very sure anything revolutionary would have been more of a leap than deeply integrating a agent/RAG pipeline into the OpenAI API. They have the compute...
This does work to a certain extent, but doesn't really converge for significantly more complex tasks. (Source: tried to make all sorts of agents work on complex problems in a divide and conquer fashion)
Did you make a framework for the agents so they could delegate problems to an appropriate model, query a dataset, etc, or was it just turtles all the way down on GPT4?
My hunch is that one big LLM isn't the answer, and we need specialization much like the brain has specialized regions for vision, language, spatial awareness, and so on.
To take the analogy of a company, the problem here is that management is really bad.
What you described is rather akin to hiring better workers, but we need better managers.
Whether it’s a single or multiple models is more of an implementation detail, as long as there’s at least one model capable of satisfactory goal planning _and_ following.
okay, i could be convinced... but what is the compute for this? you can't just "spawn threads" with reckless abandon without considering the resource requirements
As long as your checks clear and the HVAC in the data center holds up I think you're good to go.
The beauty of the Assistants is you're not limited to OpenAI models. You can wire them up to any model anywhere (out they can wire themselves up), so you can have specialist threads going for specific functions.
