MemAgent uses fixed number of tokens as long memory and applies chunk-wise reinforcement learning across the entire context window to select useful content from each chunk and update long memory, performing this iteratively until whole context window is covered. cool idea :)

funny how science brings you closer to reality by separating you further from reality

War rooms be 🫶 Let's go Syed Amaan

you've been asking for on how to catch up with new models, ok revealing a life hack I use Hugging Face MCP a lot, the prompt: "what are the trending models datasets and spaces for past week, created past week on Hugging Face? bring 10 for each, their model size, details"

is irreducibility worth it if it kills your margin of safety

Google has shared the system prompt that got Gemini 2.5 Pro IMO 2025 Gold Medal 🏅 paper now #1 trending on alphaXiv 📈

You're not a true engineer, if you don't have a strong opinion that contrasts with another engineer 😂 Jayanth Krishnaprakash Syed Amaan 🫶

New video on the details of diffusion models: youtu.be/iv-5mZ_9CPY Produced by Welch Labs, this is the first in a small series of 3b1b this summer. I enjoyed providing editorial feedback throughout the last several months, and couldn't be happier with the result.

claude code really spends a lot of tokens, not sure how many are actually useful. need to be cognizant of only feeding relevant context instead of full files and docs.

fun sunday, reading this: blog.wilsonl.in/search-engine/

Saying that deep learning is "just a bunch of matrix multiplications" is about as informative as saying that computers are "just a bunch of transistors" or that a library is "just a lot of paper and ink." It's true, but the encoding substrate is the least important part here.

Anecdotally, I’ve found the people most vocal and showy about grinding hard (9-9-6) tend to have less throughput than a garden variety workaholic. I suspect this is because they’ve internalized endurance pace all the time. And loose the ability to sprint when needed.

the 9-9-6 local maxima trap you can optimize for looking busy, hitting metrics, being “productive” – but you might be climbing the wrong hill entirely. real breakthroughs happen in the spaces between. when you’re walking and your mind wanders. when you sit with a problem long

AI for Security has never been more exciting. Let me present MAPTA, our multi-agent framework that found multiple (now confirmed!) Remote Code Executions (RCE's) in flagship web products of Tier-1 companies. Why the secrecy? We're good boys, letting them cook patched through

I also don’t know any highly skilled people who think in terms of hours relative to their output. They mostly think in terms of results. Most of your best ideas come from doing things that don’t even look like working. So if you ask people to pretend, you’d hurt yourself in two

In our investigation, we uncovered three separate bugs. They were partly overlapping, making diagnosis even trickier. We've now resolved all three bugs and written a technical report on what happened, which you can find here: anthropic.com/engineering/a-…

Given its status as one of the great American business biographies, I decided to read Titan, on Rockefeller, a couple of months ago. (I also paired it with a little of Tarbell’s history.) Titan is an impressive work, clearly the product of prodigious research. Assorted points