Meanwhile we're struggling with autonomous delivery at 35km/h 😢 Ontario is the only province where self driving pilots are legal apparently.

Tomorrow I'm running a session on training models with RL. We'll look at training loras in a free Google colab all the way to techniques to scale training across multiple gpu nodes. All practical examples you can take away to get started. Epic week at New

POV it's an hour before you present a workshop and Karpathy calls the topic "super sus" 😭😂 I'm adding this to my slides.

Successfully radicalized at least a few people to start training their own models. New great hosts as usual. Thanks to Tarun Sachdeva for organizing!

Impressive result and this has been my takeaway with GRPO more generally. 1. SFT to teach a basic skill 2. GRPO improve accuracy In my experience so far skipping SFT has never worked well, even when starting with a model which is already instruction tuned. Awesome that this

This is a great example of training an agent. Many interesting infrastructure problems to solve.

Awesome!! Can't wait to get chipped.

GPT-5 (reasoning high) is really the first time that I've felt a model can do actual work. It's smart enough that it can recover from fairly significant mistakes and is accurate enough over long context that it can be left alone. It's insane to imagine this but multiple hours.

"My first instinct was exactly what such people would delight in happening. To watch the rest of us recoil, then retract, and perhaps even eject. To leave the internet for a while or forever. But I can't do that. We shouldn't do that." world.hey.com/dhh/words-are-…

Awesome to have programs like this running here supported by incredible companies like Modal 😍

In Canada self driving pilots are illegal in every province except for Ontario where it is still extremely restricted. Imagine the damage this one regulation has done. There are many more like it.

From OpenAIs GDPval paper. Also a nice list of problem spaces worth building for.

Evan Solomon Yes, Canada pioneered AI and played an outsized role on the academic side, but has underperformed on commercialization. Many such stories in Canadian history. My take: The best way to ensure Canadian AI success is a guarantee that government understands that the existing body

A half dome lap on the way to SF. Feeling like I got hit by a truck today.

I'm surprised that we didn't see any fine-tuning updates at OpenAI's dev day. Personally, I really don't have a lot of interest in visual chatbot builders or fancier APIs. I only want better, task specific models. Seems like GPT-5 fine-tuning isn't supported at all still.

“SF is basically a Canada wrapper” is an excellent meme 😆

The hero Canada needs