If you use "AI agents" (LLMs that call tools) you need to be aware of the Lethal Trifecta Any time you combine access to private data with exposure to untrusted content and the ability to externally communicate an attacker can trick the system into stealing your data!

In era of pretraining, what mattered was internet text. You'd primarily want a large, diverse, high quality collection of internet documents to learn from. In era of supervised finetuning, it was conversations. Contract workers are hired to create answers for questions, a bit

X is active income, while Youtube is passive income

I am once again begging you to put your database servers and application servers in the same region.

You will learn things in a startup, of course. But the way to learn the fastest is to work on whatever you're most curious about, and you don't have that luxury in a startup. In a startup you have to work on whatever users want most.

This is the conclusion slide of a talk I gave more than a year ago on RL/Alignment! It still holds true today.

Airbnb should have been called localhost

As agentic AI ushers in a new era marked by tool expansion, systems are converging, and complexity is rising. Microsoft Research explores the Model Context Protocol (MCP) as a new standard for agent collaboration across fragmented tool ecosystems. msft.it/6014scgGq

legendary hacker news comments that aged like milk

As a software engineer, it's very important to learn about Gall’s Law, which states that complex systems cannot be created successfully from scratch. In reality, even large systems, such as Netflix, Google, or Facebook, have started small and built incrementally over the

I was surprised by how many didnt know that (1) per token MLE is whole seq MLE, and (2) PG at token level same as PG at seq level (optimizkng one big combinatorial action). story is different if you introduce fitted critic/Q-values or intermediate resets.

Qwen3-Omni finally, damn it is more than half a year since the release of Qwen2.5-Omni! Last time we thought that we had some successful attempt on unifying audio understanding and generation, yet we were still building small 7B model and we were far lagging behind on data

fyi 200m tokens are more text than you read your entire life

The biggest application of MCP servers: building them

Why this man is still working

I filmed this video on October 4th 2015 when I had 8,000 subscribers and schedule uploaded it to go live tomorrow 👀