We gave some of our partners early access to n1.5 — the most capable computer use model for the web. It is in production at FAANG scale as we speak, replacing a computer use model from a frontier lab. If your product can benefit from web automation — extracting structured data from dynamic webpages, filling forms, completing workflows on the web, testing vibe coded web apps — you should try out @yutori_ai's Navigator n1.5! Save your GPT / Claude / Gemini capacity for something else :)

My @yutori_ai scout notified me of @ohmaryplay tickets for the price point and evening of my choice! Bought!! Still learning, though, because I haven’t given it abilities to buy on my behalf yet.

playing w/ claude fable

Fine fine. I mean MANGO(S) 🙄

By popular demand, Claude Code-written JavaScript Warli generator.

"honest" is the new emdash.

Claude Code wrote a JavaScript Madhubani generator.

goin iso

pixels-to-action

Would you rock a @yutori_ai cap?

Here's the video of my talk @southpkcommons Demo Day! Featuring all new visualizations for why grokking works, how you can make grokked models forget, and what this says about memorization in LLMs

Polishing tools 🧰

I used to do these little algorithmic art projects some number of years ago. I was curious what it is like to do those now with coding agents. I pointed Claude Code to my past projects and asked it to come up with a new project that it thinks I'd make. I asked it to make it a "create your own" interactive tool like others I've made before. It came up with something reasonable, I can see myself making it. deviparikh.com/create_your_o… Wasn't any more of an interesting experience than any hand coding --> coding agents project would be. But I think that's because I offloaded both the creativity and the coding to it. I'd be curious to try it when I do have an idea of my own, and use coding agents to implement it.

Others are probably way ahead of me on this, but: I've refused to turn on the auto archive feature in Superhuman this whole time. Felt too dangerous / I wasn't okay letting go of the control. Spam / poorly done cold outreaches have been getting out of hand though. So I decided to try it out. A couple of weeks in, I like it! I go through the auto archive folder every other day or so. And feels like a good balance.

Go Jasper!

Browser-use agents are inference-intensive in a subtle and interesting way. Every step: task, screenshot, trajectory history → action. 30-100 steps per task. Long inputs, short output, context grows the whole time, but has high cache potential. Scouts run in the background. Delegate runs in real time. Two very different latency requirements, one platform. What running on @togethercompute gets us: ↳ 2x faster per-step inference ↳ 4-5x lower cost ↳ Scales on short notice, 99.9% uptime together.ai/customers/yutori

Browser-use models have a fairly interesting workload — long inputs, short output, long trajectories with growing input, but high cache potential. Been fun solving these problems for Yutori Navigator models with Together.

For browser-use AI agents, every task is dozens of model calls in a tight loop. The inference layer isn’t background infrastructure. It’s what the product runs on. @yutori_ai runs Scouts, Delegate, and Navigator on Together AI’s inference platform: • 2x faster per-step vs. frontier models • 4-5x lower inference cost • 99.9% uptime, elastic scaling on demand Hear from the Yutori team on how they built it on Together AI, the AI Native Cloud

Future that Yutori is showing ! Ft. @abhshkdz

Confluence #810 & #915 // @deviparikh Confluence matters because Parikh trains a GAN on a sketchbook that was never composed as a public style. The machine is fed draft matter, private notation, scraps of visual thought. The authorship sits in the edit. She takes the output back through rerendering and hand-built palettes, then lets AI pareidolia misread it one more time. That loop gives the work its edge.

😮💙