New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex scripts with a snippet of YAML template

Mathematics is not something distant. It is a structure that quietly shapes reality. I’ve been working on a new video for algebrica.org

Everyone can use the same phone - broke or billionaire. Buffett's "democratization of luxury" Is this ending for AI with Mythos class models? Or is this just a bump before intelligence becomes a utility (while we figure out how to do it safely?)

if you’re freaking out about Mythos, remember: Never make any major life decisions within 30 days of a meditation retreat, psychedelic trip, or first encounter with a frontier AI model.

time for some school! cool work by @zeke - I just had another friend tell me I should spend some time in opencode - now I have no excuse

Chat, my nanochat (left) with its onboard wasm-interpreter is now clearly exceeding @karpathy’s nanochat (right) on a range of computation tasks. The wasm interpreter plus cross attention only adds about 300 million params, a marginal increase in params for a big boost! You could call it tool use but it’s a single transformer that can both predict the next token and is a functioning wasm machine, there is no external tool.

I haven't dove into the details yet .. .but @brendanh0gan is the real deal. He has been pumping out deep work for ages. I first found him when he shared a perfect for learning GRPO/RL repository after the "deepseek" moment. Follow him / check out his work

This work from @voooooogel was pretty ground-breaking: vgel.me/posts/representation…

Great stat in here: Claude Code went from 17% to 92% on our eval set once it had access to LangSmith traces and Skills. A coding agent without trace data is just guessing at fixes

til: code --no-sandbox serve-web --host 0.0.0.0

openai is cooking with this release... codex codex (codex) codex codex v-codex-high

Software dev has already changed a lot since the beginning of the year And seems like both Anthropic and OpenAI will have much better models by the end of the year. Software has always been very inefficient to make. And now it will be not perfectly efficient but orders of magnitude more so. Is there a new kind of Efficient Market Hypothesis for the software industry? Ie if you only have public information that everyone else also has, you probably shouldn’t use it to trade stocks or build a startup on. There’s little alpha there, and roaming apex predators with more GPUs than you. I find enterprise interesting because it’s a non-public slog. Long procurement processes, no public docs, bespoke fractals of internal processes and jargon and messy human context that isn’t public or legible to labs or others yet. Previously, context (eg a well written internal Google doc) was cheap relative to the cost of building software. Now it’s flipped, weirdly If building software becomes more efficient, where are the “private context slogs” worth making? Ie curating non-public context that (when combined with public agents) unlocks new value to businesses Has anyone been thinking about the Efficient Software Hypothesis, and its implications?

If you are not yet following @Prince_Canuma do it now! He is the man behind many of the engines powering local AI on your Apple Silicon, leveraging Apple MLX framework. 🚀

All the demos I'm seeing are text moving all over the screen. But what about an editor with pixel perfect syntax highlighting? You can do it with only little effort on the web now.

devrel trying to get attention by claiming SOTA smashes fiber benchmark at 100% - while actual scores are closer to ~20% ⚕️ fibergate ⚕️ 😂

Nice share by @Teknium. I see these "say don't do" patterns too often in codex-cli / gpt models - that @opencode attempts to prompt away. "when you say you are going to make a tool call, make sure you ACTUALLY make the tool call, instead of ending your turn" "When you say 'Next I will do X'or 'Now I will do Y' or 'l will do X', you MUST actually do X or Y instead just saying that you will do it."

Introducing the factory of the imagination @phyzify fastcompany.com/video/the-fa…

At the studio with @sougwen 🎧 What is drawing in the 21st century when confronted with new technologies? Can mark-making manifest the tensions new these tools pose to traditional forms of image making? How does history affect one’s practice and how can materials manifest these personal and historical relationships?

Seeing @sougwen's performance yesterday at @ArtBasel made me think of something we both had talked about a few weeks back. We discussed ideas about data, labor, and what it means to be an artist today. Sougwen had mentioned how artists have that ability to understand the dignity of work and labor amidst our current conditions, and that struck a chord. This form of mark-making: using data, protocols, and systems, is not commonly associated with labor. Maybe more with automation and efficiency. But here, watching this performance, we are confronted with the body, with labor, with time spent for the work to become an artwork. Though one can't fully attribute artistry with labor (many hours of painting don't automatically make a great work of art), in this context of "AI art," I like the conundrum this kind of mark-making infuses into the cliché and preconception of what doing art with these tools ought to be. Nothing is immediate or efficient in Sougwen's work. Everything is conditioned by the body. Data is empty until it is embodied. "...while there is meaning in the data, the meaning is the meaning we make by working with it as practitioners..." @sougwen