Bad but common LLM criticisms

a year ago

AI is both overhyped and underhyped, with widespread discussion but limited meaningful impact observed so far.
LLMs have high variability in output, and user dissatisfaction often stems from mismatched expectations rather than poor performance.
Effective use of LLMs requires detailed prompts, similar to composing an email, to achieve desired results.
The definition of AI constantly evolves, with past benchmarks like chess and Go now considered trivial.
LLMs today can perform tasks that were once multi-year PhD projects, making criticisms based on past standards irrelevant.
LLMs, like humans, make mistakes, but this doesn't negate their value.
The 'fancy autocomplete' criticism is outdated, as LLMs' capabilities far surpass simple autocomplete features.
The rate of improvement for LLMs has slowed, and it's unclear if exponential growth will continue.
LLMs have ingested nearly all human knowledge, raising questions about future learning from synthetic data.

Hasty Briefsbeta