Raw tokens fell 39–60%. The bill went up 14–36%.
A re-read of a public 945-run Claude Code benchmark, priced with the costUSD Claude Code writes into its own logs. /compact and subagents both look cheaper in tokens and cost more in dollars.
Everyone deserves LLMs
Recommender systems & agents, from the papers.
Close reads of new research on LLM agents, generative recommendation and reinforcement learning: what each paper measures, what it does not, and how I would test it.
Latest
Recent writing
Harness²: What an Agent Can Learn by Changing Its Harness
Propose, probe, compose: how a frozen model adapts its tools and instructions, and why better-looking execution is not proof of correctness.
Reachability Is Not Ranking: Finite Beams, Continuous Queries, and One TIGER-Style Baseline
A close read of continuous-query generation for generative recommendation, and why "can the target be found?" and "is it ranked well?" need separate measurements.
A Semantic Tokenizer Cannot Be Judged Without Its Decoder
FLASH, SimHash, and the hidden assumptions in generative recommendation
Knowing When to Intervene: Six Papers on RL, Search Agents, and RAG
RL, Agents & Applications · Issue 002 · arXiv submissions from October 5–9, 2026
Search Agents Need Better Decision Boundaries, Not Just Bigger Models
SearchJev, calibrated delegation, and the cost of deciding when to stop
What a Recommender Can Reach, Remember, and Afford
LLMs & Recommender Systems · Issue 002 · arXiv submissions from October 4–9, 2026
When Should a Personal Agent Override Your Recommender?
Personal-agent mediation, rescue–harm accounting, and why outcome rewards are not enough
语义 ID 如何兼顾记忆与泛化，Agent 怎样学会修正？
第 001 期｜AI 研究与应用周报
Semantic IDs, Reasoning Budgets, and Agents That Learn to Revise
Issue 001 | AI Research & Applications Weekly
From the archive: learning LLMs from scratch
Build your own ChatGPT from scratch in Python numpy
Generative Pre-trained Transformers, or GPTs, represent a significant leap forward in natural language processing technology. These models, especially notable ones like OpenAI’s GPT-3, are celebrated for producing text…
From Zero to Hero: A Deep Dive into Andrej Karpathy’s Neural Networks Course
Artificial Intelligence (AI) is rapidly evolving, and for those eager to dive into the world of neural networks and machine learning, there’s no better place to start than Andrej Karpathy’s “Zero to Hero” course. This…