Taeho Kim
HomeWritingNotesReadingAbout
KOEN
Menu
HomeWritingNotesReadingAbout
KOEN

AI

A New Paradigm for LLM Learning: System Prompt Learning

Harvest·May 14, 2025·8 min·gpt-4.1-2025-04-14Original source

Andrej Karpathy identifies a missing learning paradigm for LLMs. Pretraining provides knowledge, finetuning shapes habitual behavior, but much human learning happens through explicit instruction -- similar to system prompts. This "system prompt learning" paradigm may be key to more effective AI systems.

Related writing

Harvest•AI•English

Greg Brockman on Attention Bottlenecks

A summary of Greg Brockman's comments on compute, scaling, AGI access, organizational use of AI, responsible deployment, and science.

May 2, 2026Read more
Harvest•AI•English

When LLMs Corrupt Delegated Documents

A summary of the DELEGATE-52 benchmark and what it reveals about document corruption, long workflows, tool use, and agent reliability.

May 2, 2026Read more
Harvest•Startup · AI•English

YC Summer 2026 RFS — A Quick Read

Y Combinator's Summer 2026 Requests for Startups frames AI as infrastructure, not a feature. The big bets: chemistry-free agriculture, AI doing services rather than selling tools, n=1 medicine, company brains, counter-swarm defense, agent-first software, and rebuilt hardware/semiconductor stacks.

Apr 29, 2026Read more

Taeho Kim · 2026

LinkedInSubstackGitHub