LLM Field Notes
The Next LLM Breakthrough Will Not Be a Bigger Model
Every LLM hits physical limits. NVIDIA's Nemotron 3 Puzzle shows what serious compression…
LLM
9 posts tagged
LLM Field Notes
The Right Model for Each Task, Not the Biggest Model for Every Task
Kimi K3 is a frontier push, but 2.8T total parameters is not the same as 2.8T active…
Hermes Agent
1 Billion Input Tokens Later: My Personal Agent Field Report
90 days, one billion input tokens, roughly thirty dollars a month. The number is…
Hermes Agent
Running a Team of LLMs: Mixture of Agents in Practice
One model is a soloist. A team of them, coordinated by an aggregator, is an orchestra.…
LLM Field Notes
Prompt Engineering Didn't Die. It Built a Skeleton.
Prompt engineering didn't die -- it evolved. The clever prompt is gone, replaced by…
LLM'ler kas belleği oluşturmuyor — sadece kiralıyor.
Bir araç senin yerine düşündüğünde, refleksi geliştirmeyi bırakırsın. Hız gerçek. Erozyon…
LLMs don't build muscle memory — they rent it.
When a tool thinks for you, you stop growing the reflex. Speed is real. The atrophy is…
LLM'leri kendi kodlarımızla eğittik. Şimdi onlara daha iyi kod yazmalarını tembihliyor, bir sonraki modellerin çıktılarının sadece kendi kodlarıyla eğitilmiş modeller olacağını umursamadan kullanıyoruz. Döngü iyiye gitmiyor.
Artık kendine ihtiyacı duymayan bir döngüde insan sezgisinin kalitesine ne olur?
We trained LLMs on our own code. Now we instruct them to write better code, without caring that the next models will be trained only on their own output. The loop is not improving.
What happens to the quality of human intuition inside a loop that no longer needs it?