I went to Caojiawan Station (yes that "ghost" one) in Chongqing and was not greeted by an empty station with barren fields overgrown with weeds but beginnings of dense mixed-use Chinese suburbia. (Older pic not mine) A 🧵of my visit. (1)

Could have just done computer use and call it a day but no... i just had to do it the hard way

Great advice from Bryan Bischof fka Dr. Donut on "What Evals Framework should you use" 🌶️ He's a guest speaker in our evals course! maven.com/parlance-labs/…

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

the biggest bottleneck in my workflow is communicating my intentions clearly and no amount of model intelligence increase will solve that for me unless we are truly symbiotic which i don't really want anyway. or i can cede my agency to its choices which i don't want either

you don't need a framework to train models pretraining? for loop SFT? for loop RL? two for loops inside a for loop

I think there is a confusion about strength training as if it's something you should like I hate going to the gym, I hate lifting very heavy weights, it's exhausting, obviously HEAVY, and I wanna quit and end it ASAP But that's the entire point, it's choosing to suffer on

it’s incredible how many ppl get promoted for adding complexity instead of removing it. engineers especially in big co’s. what does that tell you about the kind of people who end up climbing the corporate ladder?

> The proof-of-concept (PoC) hinges on users enabling the Gmail integration, but the attack can be extended to any connector that ChatGPT supports, including Box, Dropbox, GitHub, Google Drive, HubSpot, Microsoft Outlook, Notion, or SharePoint, effectively broadening the attack

Forget MCP, what I want is an industry consortium backed standard for a JSON API protocol for talking to LLM providers Everyone building half-baked copies of the OpenAI Chat Completion API worked until it didn't any more!

I don't know what labs are doing to these poor LLMs during RL but they are mortally terrified of exceptions, in any infinitesimally likely case. Exceptions are a normal part of life and healthy dev process. Sign my LLM welfare petition for improved rewards in cases of exceptions.

Jack Ellis Sublime text … all organic hand written code

TIL you can just run `networkquality` in your terminal on MacOS

Until ~2015, GitHub Pages hosted over 2 million websites on 2 servers with a multi-million-line nginx.conf, edited and reloaded per deploy. This worked incredibly well, with github.io ranking as the 140th most visited domain on the web at the time.

Why can’t programmers tell the difference between Halloween & Christmas? Because oct 31 = dec 25.

Fewer frameworks, more libraries please.

If you're non-technical, AI coding tools are for pitching/communicating to engineers, not replacing them. I use Claude to make interactive prototypes that show exactly what I want to build. Then I hand it to a real developer. The prototype isn't the product. Conflating the