Smaller models, smarter prompts, and the lessons learned from "strawberry-gate"

Published on Convert on 06/08/2025.

When AI failed to count the R’s in “strawberry,” it highlighted two ways of thinking: technologist (how the model works, how it’s improving) and practitioner (how to get the job done with the right tools). The article explains how LLMs actually work (pattern matching, no real “counting” or “facts”) and why “progress” can be misleading—cost, benchmarks, data contamination.

Five practical tips:

1. Use smaller models (and local tools like LM Studio) when they’re enough.
2. Improve prompting: Chain-of-thought and few-shot, with clear steps and human review.
3. Self-reflection: Get the LLM to critique its own (or another model’s) output using clear criteria.
4. Break tasks into small steps and use calculators or APIs where LLMs are weak.
5. Put tools in workflows (e.g. n8n) so results are stable and under your control.

Bottom line: treat LLMs as tools, learn how to use them well, and favour smaller models and smarter prompts to support your thinking—not replace it.

Read the full article on Convert →