We will use everyday language where possible and mark technical terms in purple. Key is that modern LLMs have a component of making answers (in blue, called "inference") and a component of improving answers (red, called "training"). pic.twitter.com/iUmGaNMKUz
— Kording Lab 🦖 (@KordingLab) August 25, 2026
The real work is done by a core machine, the neural network which is used to generate one word at a time based on its inputs. The probability of each word depends on prior words. And words are chosen one by one according to those probabilities. pic.twitter.com/7zWO87TThs
— Kording Lab 🦖 (@KordingLab) August 25, 2026
The result of this are what is sometimes called large-scale "stochastic parrots". Very impressive text but often somewhat superficial. But these are the old LLMs (think 2022). Innovation was super fast since.
— Kording Lab 🦖 (@KordingLab) August 25, 2026
A second idea (B) for making answers, tool use. Call programs like databases, internet search, calculators etc. pic.twitter.com/zyPU00vmV3
— Kording Lab 🦖 (@KordingLab) August 25, 2026
And all of this, is better with better “training“ goals. Wherever truth is defined, push towards this truth (D). And where it is not defined, use LLMs to check answers ("self evaluation", (E)) pic.twitter.com/vLOy2UCsYb
— Kording Lab 🦖 (@KordingLab) August 25, 2026
A great deal of the strengths and weaknesses of today's LLMs can be understood from these component ideas.
— Kording Lab 🦖 (@KordingLab) August 25, 2026
No comments:
Post a Comment