Squish Meets Structure — Designing with Shoggoth
summary
Claims that language models become useful when broken into tiny specialized reasoning engines, with composition and prompt chaining tying models to trusted data sources. A working product, Elicit, extracts dosages and outcomes from thousands of papers into a table and cuts a 6–12 month literature review to days–weeks, showing source quotes and confidence warnings below 90%. Uses: semantic embeddings for search, prompt engineering, fine‑tuning and RL to constrain hallucinations.
AI-written summary · about
chapters (10)
- 0:00Designing with language models
- 1:06Practical guidance focus
- 2:09What language models are
- 8:52Beyond text generation
- 11:58Hard to design with
- 15:28Language models as squishy
- 19:00Squish versus structure
- 23:11Balancing squish and structure
- 25:17Models as reasoning engines
- 32:41Design principles for tools
related talks
28mTalking to Objects in Natural Language: Toward Semantic Tools for Exploratory(…)
49mBen Hall & Barbara Fusinska
27mUnlocking the Potential of Large Language Models
01h 00mMalignant Intelligence: Prompt engineering and software archeology
48mStop Fighting Your AI Engineering Prompts That Actually Work
01h 00mNext-Gen Software Architecture with Large Language Models (LLMs) like ChatGPT
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools
42mText-to-Struct: Fine-tuning SLMs for Query Intent