grikdotnet

"Attention is all you need" they wrote in 2017, and introduced a transformer architecture implemented in LLMs. The "The Parallelism Tradeoff: Limitations of Log-Precision Transformers" article in 2023 described transformers' constraints.

In a theory transformers are Turing-complete, which means they can solve any computational task. But there are 2 major assumptions: infinite precision, and arbitrary size of models. The limits of computer memory and precision in chips are set in stone. Long story short, there are tasks that transformer LLMs can not solve.

Main limitation.

In programming we are used to the binary logic: AND, OR, NOT, and others, called "gates". They are used to implement operators like IF and FOR loops that are used to branch execution and define the logic of applications.

grikdotnet

LLM - a stateless in-memory map/reduce backend service

Vertical LLM - a specialized LLM service optimized for a particular business scenario

RAG (Retrieval Auhmented Retrieval) - an back-end architecture of a service, consisting of a middleware, a database for a full-text or vector serch, and an LLM service

Agents - an architecture, consisting of an LLM, and a middleware that implements business logic, an API, and usually sessions

Multi-agentic system - a microservice implementation of multiple middleware services, and often multiple LLM backends

grikdotnet

How much did your productivity increase with agents? How much did code quality improve?

People jump into chats and write: “1000% productivity gain!”

Question of the year. “How much” is the only question that actually makes sense to ask about agentic software development.

The key is to define what exactly you are measuring. Traditionally, programmer productivity was measured as how fast different people completed tasks relative to each other, assuming comparable learning speed.

But we can also measure it as project costs. I’ll take my own speech recognition project as example. 
I can estimate the time I need to solve the tasks. To dig through documentation and debug everything on my own,