PWV Portfolio Feed

atuin 18.17 is out 🐢✨

- 78x faster to open and search
- ai can see your terminal
- more resilient sync
- switch ai models with /model

mild take: any theoretical benefits of the "public benefit corporation" have been disproven by the major ai labs

Not just yet but in the next week or two!

We’re releasing a self hostable ai server, you can point it to local models or remote ones with your own keys

Shipped: GPT-5.6 Sol, Terra, and Luna

Sol, the flagship for long-horizon work.
Terra balances intelligence and cost.
Luna for fast, cheap, high-volume jobs.

Switch with /model

Format-on-save is now disabled by default, except for languages with official formatters.

For existing users, this means their custom configurations will stay, but any values that were previously "default" will be updated.

We've moved all the key agent-related settings to the settings editor. Configure LLM providers, external agents, and MCP servers directly from the Settings UI.

We now have a setting for rendering the current-line git blame in the status bar instead of inline at the cursor. 

Thanks Koshkaj!

None of this needed a smarter model. A 2018 XGBoost and a 2025 one score about the same.

What never scaled was the engineer running 20 feature loops a week. The agent runs 60 overnight for pocket change, and trying things was most of the job anyway.

it didn't get there by brute force. Nobody told it crystals are periodic.
It worked out that raw atom-to-atom distances misread the chemistry across a cell boundary, then rebuilt them as metal-to-oxygen nearest neighbors. 

677 features by the end. 😳

The headline run was a materials-science task. Predict formation energy and bandgap of a transparent conductor from its crystal structure.

It landed 1st of 685. OOF score 0.04853 against a 0.05589 gold line. The humans who won it originally spent weeks.

Sixty cycles later, across four MLE-bench datasets at once, it had two golds and three of four finishes above the Kaggle median.

No human in the loop the whole run. One NVIDIA L40 on @CoreWeave and a few dollars.

Tabular ML didn't die. It was just waiting for agents. 

We rented one GPU for the price of a coffee and asked an agent to top a Kaggle leaderboard on its own. 

It came back with two gold medals. Here's the run 👇

62 apps shipped at Hot AR Summer and counting. 🚀

Builders are turning out some seriously creative AR experiences. Browse the showcase and see what you could add to it! ✨