Member-only story
| PROGRAMMING | ARTIFICIAL INTELLIGENCE |
Scaling Isn’t Everything: How Bigger Models Fail Harder
Are Large Language Models really understanding programming languages?
Code generation and completion models have disturbed programmers' dreams, so much so that they have led to lawsuits and controversy. On the other hand, it is still unclear how much these models include programming. Also, the more parameters a model has the more capable they should be. what if they don’t?
The model he dreamed of programming
With the development of LLMs, it came natural to apply these models not only to language but also to programming. Preliminary investigations showed that GPT-3 could generate simple programs from Python docstrings. All this despite the fact that the program had not been specifically trained to generate code. Considering that there were many public code sources, many research groups focused on developing specific models for code generation. For example, Codex:
