What is mattwood.fyi?

A live, structured feed of riffs and links from Matt Wood, Chief AI & Technology Officer at AWS — readable by people and queryable by agents as a typed knowledge graph.

Is there an API and is it free?

Yes. A public, read-only query API is available with no authentication and no API keys. See https://mattwood.fyi/openapi.json and https://mattwood.fyi/docs.

How should an AI agent use this site?

Fetch https://mattwood.fyi/feed.json for the full structured feed, or query the API for keyword/semantic search and the connection graph. See https://mattwood.fyi/agents/ for when-to-use guidance.

PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations

PLUM is a framework that adapts pre-trained large language models for recommendation systems at scale, using semantic IDs for item tokenization, continued pre-training on domain-specific data, and generative retrieval fine-tuning. The approach demonstrates significant improvements over production models on video recommendation tasks and has been deployed to billions of YouTube users.

arxiv.org →

Connections

Challenges

Infinite-Parameter LLMs: Generating Weights from Live Data PLUM uses semantic IDs and continued pre-training as a static adaptation approach, while Infinite-Parameter LLMs propose generating weights from live data — contrasting paradigms for keeping models current with dynamic item catalogs

Supports

LLM Classification Is Feature Engineering PLUM's use of semantic IDs for item tokenization is a form of feature engineering that bridges LLM representations with recommendation signals, reinforcing the idea that LLM classification tasks require thoughtful feature representation

Related to

Training a 4B Model for Faster SQL Query Plans Both involve adapting/fine-tuning models on domain-specific data for specialized industrial tasks (SQL query plans vs. recommendations), sharing the continued pre-training and domain adaptation methodology
GenRec: LLM-Native Recommendation System at Netflix Both papers directly address LLM-native recommendation systems at industrial scale, with PLUM and GenRec both tackling the core challenge of adapting large language models for production recommendation pipelines
Contrastive Language Model (CLM) Both frameworks modify pre-trained language model training objectives for non-standard downstream tasks (contrastive learning vs. generative recommendation), sharing the theme of repurposing LLM pre-training for retrieval/ranking applications

OneRec-Think: In-Text Reasoning for Generative Recommendation PLUM and OneRec-Think both adapt pre-trained language models for industrial generative recommendation, though OneRec-Think specifically focuses on reasoning scaffolding via chain-of-thought
Semantic ID-based Generative Retrieval for Podcast Discovery at Spotify PLUM also adapts pre-trained language models for industrial-scale generative recommendations, directly paralleling GLIDE's use of LLMs and Semantic IDs for production recommendation at scale

Type:	link
First published:	Tuesday, 29 September 2026 at 08:10 UTC
Location:	https://arxiv.org/abs/2510.07784
URL:	https://mattwood.fyi/i/5d89f4b3

I'm Matt Wood, and this is For Your Information. A live list of riffs and links for you and your agent, drawn from what I'm reading, noticing, questioning, concluding, and revising.

Connections