DeepSeek is cool and proofs size do not always matter, but it is not a real game changer.
It comes down to few very simple things. These are already under development to get better.
- The amount and versality of data we are pushing to LLM's today.
- The scale of questions and response pair with the context.
- Bad prompting, this is human factor.
Amount and versality of data:
- Insane size. One model has data about anything. From knitting to coding, from ancient Egypt to modern day politics and keep on listing. Result is LLM that can do pretty well everything, but is expert in nothing.
- Well tuned smaller models is the solution. RAG/Agents are the response and those solutions are coming more common.
The scale of questions and response pair with the context. & Bad prompting, this is human factor.
- AI can create simple pong or tetris just after simple prompt. But thats it.
- Human cant ask LLM to create full size FPS. LLM runs out of context. And will do so for distant future.
- Bad prompting; humans tends to believe that everyone knows what just he/she is thinking or wanting.
But LLM's are handy for many tasks. I use them to create all the boring stuff.
"Write index.html for me please.
title = foo bar
Header (w 100%, h 20%), sidebar (left, w 20%, h 100%), main there nicely under header next to sidebar and footer at bottom.
Create index.html and styles.css. Ill adjust them myself more properly."
