Against vibes: When is a generative model useful

3 days ago

The author critiques the indiscriminate use of generative models without scientific evaluation of their utility for specific tasks.
A model for evaluating generative model utility is proposed, focusing on three factors: encoding cost, verification cost, and task dependency on artifact vs. process.
Generative models are more useful when encoding and verification costs are low and when the task is artifact-focused rather than process-dependent.
The author provides examples where generative models fail (e.g., generating complex code) and succeed (e.g., installing a package with a simple prompt).
Verification of generative model outputs is challenging due to their probabilistic nature, often producing plausible but subtly incorrect results.
Process-dependent tasks, such as education and certain engineering tasks, are poorly suited for generative models as they require human involvement for skill development and knowledge creation.
The author emphasizes the need for domain expertise when using generative models to ensure useful and correct outputs.
Generative models are seen as tools that, while capable of producing vast amounts of output, often fail to deliver genuinely useful results without careful consideration of trade-offs.

Hasty Briefsbeta