I'm not buying it....
There's no way these major AI companies don't know what they're selling is mostly ridiculous...
I just got done watching an AIDevCon "presentation" from some expert developer at Anthropic trying explain how their claude.md are working amazingly well for instructing agents...except now they call them context "memories"...
They then go on to talk about the inevitable memory layer that is now needed to keep track of and allow the agents to properly access all these highly effective "memories"... and how they are allowing the agents to even update these memories themselves now...
Then came the crunch... shocker! This whole setup doesn't work so well in production...
So the solution to all these agents working against a codebase and assumptions being stale etc is versioning and hashing before and after...they're actually calling these guardrails.
Then the presentation smoothly glides into portability for the whole "system"... and of course it's all a nice clean API setup.
The entire upshot is that it's supposed to make these super duper agents EVEN better because they are doing a self learning loop... but they don't say against what baseline or how they measure "better"...
There's also a disturbingly loose use of the word "learning" in connection with the agent which seems to suggest they are getting this entire "loop" process confused with fine tuning?!
I just don't buy it... this has got to be some kind of effort to just get millions of people blindly using these models and setups and burning thru tokens.
Anything else makes no sense... there's no way this works in real production grade complex multi-stack systems.
Top comments (0)