This has been my experience too. A junior dev at my last company kept trying to use ChatGPT to generate docket compose files and wondered why they generally didn’t work.
My research has been on time series forecasting which is tangentially related to NLP. People are shocked when I point out to them that all these models do it predict the next token. Using weather forecasting has been a good analogy for why long AI generated texts are extra bad: weather forecasts get worse as the horizon increases.
Despite all my gripes about LLMs, I must say that copilot has saved me writing TONS of boilerplate code and unit tests.
This has been my experience too. A junior dev at my last company kept trying to use ChatGPT to generate docket compose files and wondered why they generally didn’t work.
My research has been on time series forecasting which is tangentially related to NLP. People are shocked when I point out to them that all these models do it predict the next token. Using weather forecasting has been a good analogy for why long AI generated texts are extra bad: weather forecasts get worse as the horizon increases.
Despite all my gripes about LLMs, I must say that copilot has saved me writing TONS of boilerplate code and unit tests.