Focus and Context and LLMs

a year ago

LLMs can generate usable SQL statements, reducing time from 4-8 hours to 15 minutes.
Agentic coding tools claim to complete multiple implementations of software, but results are exaggerated.
Only one known end-to-end AI-written project exists: a standards-compliant HTTP/2 server.
Building the HTTP/2 server required intense algorithmic supervision and context management.
Tool calling, a foundation of agentic hype, is unreliable for complex tasks.
LLMs struggle with producing JSON and perform better with ad-hoc formats.
LLMs require exceptional supervision and context curation to produce impressive results.
Current agentic programming resembles the inefficient brute-force approach of 90s genetic algorithms.
Without better context curation, LLMs will only excel with exceptional engineers.
Mediocre inputs lead to mediocre outputs from LLMs.

Hasty Briefsbeta