ChiChieh HuangFOUNDER · AI ENGINEER
Writing LLM Plateau

Notes on Karpathy's 2025 Review

Date2026.01.08
Length47 words
Reading~1 min
ChiChieh HuangFounder · AI engineer

Translated from the Chinese original · Read the original

LocatorLLM Plateau
2 wks
Overview

Karpathy’s year-end review is well worth reading. A few notes:

  1. Reinforcement Learning from Verifiable Rewards (RLVR)
  2. Ghosts vs. Animals / Jagged Intelligence
  3. Cursor / new layer of LLM apps
  4. Claude Code / AI that lives on your computer
  5. Vibe coding
  6. Nano banana / LLM GUI
End of the trail

47 words, and you made it to the end.

RSS feed Back to the map
Newsletter

Get the next essay by email.

One email when a new essay goes up, nothing else. Unsubscribe in one click.

← Previous · 2026.01.08One Week with OpenClaw: Getting Started Next · 2026.01.11 →AI Agents Are Becoming Infrastructure
Nearby cairns · the closest essays on the map
LLM Plateau3 days apart 2025 in Review: LLMs Started Doing Real Work Looking back at 2025, AI finally stopped just talking and started doing things. Notes on Simon Willison's year-end review and why the real shift is AI fitting into how people work. LLM Plateau24 days apart Pretraining from Scratch to Production: What I Learned Notes from Liang-Hsun Huang's DevFest Taipei 2025 talk on pretraining Gemma-3-270m for Traditional Chinese, from building the FineWeb-Edu-zhtw dataset to the learning rates that make a small model actually learn. LLM Plateau38 days apart The Multi-TW Dataset Multi-TW is the first multimodal benchmark built for Traditional Chinese, testing image, listening and text comprehension with real exam questions, and measuring latency too. A guided read.
ChiChieh Huang
Surveyor

ChiChieh Huang

I build generative AI products and write about them, first in Chinese. Lately I’ve been researching agent memory and testing the ideas in Cairn.

About me Invite me to speak