Squish Meets Structure – Designing with Shoggoth
summary
Calls language models a squishy Shoggoth vs structure problem and argues for breaking big tasks into small, verifiable steps. Describes building Elicit — a tool that uses embeddings, RAG, prompt chaining and model scoring to extract tables from thousands of papers in days instead of months, including a three‑tier confidence score that took months to tune. Shows temperature and fine‑tuning tradeoffs, warns about hallucinations, and recommends designing focused, tiny reasoning engines rather than one open‑ended text box.
AI-written summary · about
chapters (12)
- 0:00Introduction and speaker intro
- 1:39Language model basics and capabilities
- 10:45Language models in real-world products
- 17:13The squishy nature of language models
- 23:27Balancing squish and structure
- 30:20Designing with tiny reasoning engines
- 33:53Closing remarks and resources
- 34:18Q&A on model limitations and RAG
- 37:42AI in interface design and prompt engineering
- 42:27Integrating AI into workflows
- 44:00Costs and sustainability of AI
- 46:36Final Q&A and wrap-up
related talks
47mSquish Meets Structure — Designing With Shoggoth
33mCognitive Shortcuts: Models, Visualizations, Metaphors, and Other Lies
43mMoldable Development & Glamorous Toolkit in Practice
30mCognitive Shortcuts: Models, Visualizations, Metaphors, and Other Lies
43mLanguage Games
27mThe Shape of AI : How designers integrate AI into their workflow
30mUsing AI to Create Better Requirements and Designs
16mWorking Against Predictability