From AI Software Engineers to AI Knowledge Workers
summary
Argues that AI should empower end users to build automations directly, not just augment developers, and proposes a Prolog-like language Universalis plus runtime Automind that treats LLMs as neural computers (weights = ROM, context = RAM) and models tools as relations and horn-clauses. Claims automatic formal verification via proof-carrying code is required for safety, limits language features (no full recursion, bulk ops and monoid comprehensions instead), and shows a working prototype with RAG, AST-based indexing, JSON schemas and guided generation.
AI-written summary · about
chapters (11)
- 0:00AI software engineers to AI knowledge workers
- 1:45Software failure for end users
- 4:58Humans don't scale like hardware
- 7:29Low-code and AI-to-code fail users
- 11:22Testing limits and AI safety
- 16:45Formal verification as path forward
- 17:48Knowledge worker apprentice
- 21:47UniversalES and AutoMind vision
- 29:56LLMs as neurocomputers architecture
- 33:52Proofs over tests for bug-free software
- 36:04Q&A
related talks
33mVerifiably Safe Agentic Compute
26mLanguage Model Agents Enable Semi-Formal Programming
47mHow AI Will Bring Computing to Everyone
43mAI Systems: Foundations for Artificial Minds
20mWill AI Replace Me? A Philosophical Mind Dump.
59mIt AI n't What You Think!
26mAl and Software: The Good, The Bad, The Scary
34mProgramming Languages for AI Programing Agents