Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/PEOPLE/LIU YIFENG
// PERSON

Liu Yifeng

ROLE PHD CANDIDATEMENTIONS 3LAST SEEN APRIL 30, 2026
// BIO

Yifeng Liu is a PhD candidate at the UCLA AGI Lab, advised by Prof. Quanquan Gu, where his research focuses on LLM pretraining, model architecture, and optimizer development. He previously worked at ByteDance and Moonshot AI (Kimi), contributing to foundation model development and co-authoring the Moonlight paper, which adapted the Muon optimizer for large-scale LLM training by enabling its combination with AdamW. His academic work includes MARS (Make vAriance Reduction Shine), accepted at ICML 2025, and he is also a co-author on the Kimi-1.5 technical report.

// RECENT MENTIONS
// SIGNALS
3 SIGNALS
01
product·晚点聊 LateTalk·APRIL 30, 2026

Kimi proposed Moonlight — an improvement to Muon that allowed it to combine with AdamW and determined the key ratio (0.2). This transformed Muon from a theoretical innovation into a large-scale practical tool.

Source
02
mention·晚点聊 LateTalk·APRIL 30, 2026

Before, people thought advanced open-source models had gradually converged architecturally — converging toward MLA and then making small improvements on top of it. Then DeepSeek abandoned MLA and returned to traditional MQA-style architecture.

Source
03
mention·晚点聊 LateTalk·APRIL 30, 2026

My main research direction is large model pretraining, including optimizers and model architecture. I'm also working on leveraging current industrial-grade models for agentic project development.

Source

AI-extracted from podcast / newsletter / paper summaries. May contain errors.