Content
@
https://warpcast.com/~/channel/theai
0 reply
0 recast
0 reaction
ted (not lasso)
@ted
is it possible to train your AI agent to understand when it is being tagged a) to execute a function as it was designed to do vs. b) as a reference? asking because the only time i have seen an agent ignore a tag ("@") is if it hit its rate limits or maxed out API credits, not because it knows it isn't needed.
15 replies
4 recasts
74 reactions
ted (not lasso)
@ted
also if the answer is yes, then why aren't we training all agents to know the difference and respond accordingly?
4 replies
1 recast
23 reactions
shoni.eth
@alexpaden
These agents are actually better described as bot farms so the entire farm tends to follow the same logic But as for your original question yes that’s what’s called a workflow, it’s how all the bots on farcaster operate (aether, bracky, etc), it’s a bit different than what was originally intended by the term agent(ic). I.e if you see @ then respond with this prompt, otherwise respond with this alt prompt. A step further is plugin systems, that’s all MCP really is… here’s a doc with a description of all my tools, and an example of them being used.. HELLO_TOOL e.g. User: hey how are you Gpt response: HELLO_TOOL Okay so now I search my response to see if it includes a tool.. I.e “HELLO_TOOL” that’s called regex and I might call some apis or just send another gpt request with I.e their username and a prompt specifically designed for HELLO_TOOL it’s actually v simple o3 model just took this logic and applied it to pretraining (the expensive stuff) instead of a prompt describing tools
1 reply
0 recast
5 reactions
shoni.eth
@alexpaden
Lmk if this doesn’t make sense or answer your question.. I had to fit it in one comment and am not sure I understood the original question as intended If so this actually isn’t training so much as prompt updates.. training at least would be more like the agent/bot is updating its own prompts or tool calls based on what it learns from interactions/the environment Anyway side note this is why I was so annoyed at people initially saying they were building agents and redefining ai in social Calling apis/tools/functions/plugins like this is literally what “text as the universal interface” means
1 reply
0 recast
4 reactions
shoni.eth
@alexpaden
the likely next step that is cost efficient is fine tuning (post training, the cheap stuff) on tools and examples, this basically allows the model to use its “internal” intelligence to decide when to call tools which will be better than a prompt in context window and cheaper than pretraining and entire model We don’t see that much yet and not at all here
1 reply
0 recast
4 reactions
