It’s interesting to see something that’s so complex reduced to its simplest form.
New art project.
Train and inference GPT in 243 lines of pure, dependency-free Python. This is the *full* algorithmic content of what is needed. Everything else is just for efficiency. I cannot simplify this any further.
gist.github.com/karpathy/862…
I really enjoyed this talk from @3blue1brown where he gives an accessible way to visualize transformers and attention.
Shane Lawrence retweeted
Want to make MCP tools more interactive? @janakiramm explains how to integrate elicitation so your apps can accept real user input instead of fixed parameters.
thenewstack.io/how-to-implem…
TIL I can immediately escape a hung SSH session! I don’t need to repeatedly mash Ctrl+C, Ctrl+Z, and Ctrl+Shift+6 before giving up and closing the terminal. I can press Enter ~ . and it just works.
TIL
After a lot of dead-ends and the usual turning it off and back on again, I found a bug. No matter what channel width I set, the 6 GHz band was stuck using 20 MHz.
At the conference:
AI’s different because is non-deterministic and we need more interpretability research to understand what it’s doing.
Debugging my own code:
It worked on my machine. I’ll add print statements everywhere and try to figure out what it’s doing.
“In a particularly thought-provoking session, Shane Lawrence, senior staff engineer at Shopify, discussed security threats to AI models.”
Thanks @camillecrowelee!
It was a lot of fun to present this demo and talk about AI (and a bit of Supply Chain) today at #KubeCon!
Mind-blowing talk by @shaneplawrence at Cloud Native AI Day at #KubeCon!
He shared a live demo showing how AI can accidentally SPILL SECRETS (!!) and how to deal with these new challenges!
This is what makes KubeCon fun!! Just cool ideas everywhere 💫
Shane Lawrence retweeted
Types of controls:
1. Controls as code.
2. Controls as bureaucracy.
3. Other?
Has anyone generalized a deobfuscator that can be built into pipelines yet?










