New LLM research, explained by people who build it.
Each issue takes one idea that matters (a KV-cache breakthrough, a new reasoning method, the architecture behind a product launch) and explains what changed, why it works and what it means for people building with LLMs. Read by about 37,000 engineers, researchers and builders.
Plans
Free
$0
No card needed.
- Public issues by email
- 30 visual flashcards as a free PDF in the welcome email
- Upgrade or unsubscribe any time
Paid monthly
$15 /month
US$15 billed monthly. Renews until cancelled.
- Every paywalled issue and the full archive
- The exclusive architecture page
- The Premium Asset Library: the 376 LLM flashcards and the illustrated book
Paid annual
$150 /year
US$150 billed yearly, about two months free compared with monthly. Renews until cancelled.
- Everything in the monthly plan
- Includes the flashcards and the book, US$35.98 if bought separately
US dollars. You subscribe, pay and cancel on Substack, which shows the final price and any local tax before you pay. Choose the monthly or annual option on the Substack page.
Recent issues
- The KV cache is now 890 bytes per token How one new model stores far less memory per token, and what that changes for long contexts.
- Latent Reasoning vs. Chain-of-Thought: How Models Think and Learn New Tasks A map of reasoning approaches, from written-out chains of thought to reasoning inside the model.
- What is an AI agent harness? Why the same model behaves differently across agent systems, and why more scaffolding is not always better.
- We Built TurboAngle: Near-Lossless KV Cache Compression Without Calibration The lab's own KV-cache compression method, explained with results.
Questions
What does the newsletter cover?
New LLM research explained in plain language: inference and KV-cache efficiency, reasoning, agents and agent harnesses, and the architecture behind new AI products, plus the lab’s own research.
What do paid subscribers get?
Every paywalled issue and the full archive, a dedicated architecture page, and the Premium Asset Library, which currently includes the LLM flashcards and the illustrated book How LLMs Happened (US$35.98 if bought separately).
How does billing work?
Paid plans are US$15 billed monthly or US$150 billed yearly, and renew until cancelled. Checkout and cancellation happen on Substack.
How often is it published?
One or two in-depth issues a month. We write when there is research worth explaining, not on a fixed weekly schedule.
Only want the flashcards or the book?
Both are sold on their own: the flashcards for US$19.99 and the book for US$15.99, one-time purchases.