KevinSimback

It should be obvious by now that just selling inference into a highly competitive and increasingly commoditized market is not going to justify the trillion+ $ valuations
And we should also remember that if the product is subsidized you are, or eventually will be, the product

Can a small open model with $3 of fine-tuning beat a production RAG setup?
This is what I tested, and the results were impressive
Based on a 100-question eval set, a fine-tuned version of Qwen3.5-9B outperformed Gemini Flash with RAG and Opus4.8/Sonnet 5 without RAG
Working on a full write-up of the end-to-end flow

Over the weekend there was a lot of chatter (and slop) about Graph Engineering vs Loops, thanks @steipete
Just a reminder that Looper is already very graph-aligned
> it thinks in nodes and decision points (goal, plan, gates, deliver, stop)
> gates act as conditional branches (pass / revise / fail)
> it maintains explicit state and produces a flowchart preview
It basically addresses the main weaknesses of "naive loops"
So no need to totally discard loops, just build better loops
That said, graphs can be better when you need a combo of LLM-based and deterministic-based actions together in a proc

As a parent, managing screen time is an ongoing battle
So every morning my Hermes agent sends a quiz to my older son, if he gets any questions wrong his daily screen time gets reduced
Evil dad!

I once has a boss who would often give the same assignment to 2-3 different people
It caused resentment across the team and was terribly inefficient
Agents have no such problem, and it costs next to nothing