Talking to/with Machines
summary
Engineers must treat LLMs as a separate audience and build explicit, shared assumptions as part of prompts and documentation rather than relying on vague natural language; a concrete example uses Claude to make a peanut butter sandwich, discovers mismatched defaults (cream vs crunchy, square cuts) and only gets the wanted result after asking the model for its assumptions and aligning them. The talk names limits like context window and token exhaustion, highlights emerging fixes like context compression and memory or “dreaming,” and argues that LLMs perform reliably once make explicit assumptions are supplied.
AI-written summary · about
chapters (8)
- 0:00AI and existential angst
- 1:44Human-machine models
- 4:13Talking to machines like humans
- 7:21Natural language fails machines
- 8:49Cooperation with machines
- 10:56LLM assumption testing
- 12:44Explicit assumptions strategies
- 13:47Being human in the loop
related talks
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
32mWhat can we expect of LLMs as software engineers?
01h 04mBuilding Expertise Into Generated Conversations
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
38mNobody Pages the LLM: Engineering Rigour for Vibe Coding
45mRewriting the Role: Developers in the Age of LLMs
26mContext Engineering: Practical Techniques for Improving Agent Quality Today
10mRethinking collaboration: LLMs, teams, and cognitive load