Experiments in health, endurance, and tech. Building @FeanorLabs.
San Francisco, CA
Joined May 2018
- Tweets2.3K
- Following297
- Followers1.1K
- Likes16K
Pinned Tweet
I’ve made Auri 4.0 Keyboard even better since this benchmark study.
It’s now #1 in voice dictation and only 0.8 points behind Apple in touch.
You can finally try it yourself:
apps.apple.com/us/app/auri-a…
This is fun
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
Asked Codex to collaborate with Claude for some out-of-the-box thinking, and omg, what an impatient micromanager Astra is
I keep seeing these "No on G," "Yes on E," "Maybe on X" signs all over SF, and can't help but wonder, perhaps naively:
1. What's the point of electing government *representatives*, employing ~35,000 government workers, and spending ~$20,000 per capita if general population is then asked to make these decisions?
2. Who has the time and willingness to thoughtfully research all these props? Either the unemployed, actively involved in politics, or people generally not known for their acute decision-making.
What am I missing here.
Ludovico Einaudi is the greatest composer of our lifetime, and it’s not even close.
Every live performance fills you with such a powerful sense of unfiltered awe that you can only experience in the presence of true greatness.
Here's how I connected each agent:
- Dot: an OAuth-authenticated MCP plugin.
- Grok Bot: an authenticated routine receives requests. Grok returns answers through a helper on its existing Mac connection, while Maestro carries recent conversation history between routine runs.
- Codex, Claude Code, and Pi: Herdr manages sessions. Codex uses its read-only app-server API, and Claude uses its SDK for final answers.
- Muse: the official muse-gadget-sdk.
I built a fun way to talk to all your agents (Grok @bot, Claude, Codex, Dot, @Muse, and Pi) from anywhere. Even while you're touching grass.
Meet Maestro, a handheld agent orchestrator powered by an ESP32-S3.
Pick a conversation, hold a button, and talk. Hear the reply or read it on the screen.
Each agent gets its own pixel character. Maestro buzzes when one replies.
It works even when your laptop is offline, so you can leave it behind and go outside.
The voice flow is straightforward, too:
Maestro uploads the recording to your Mac (or to a VPS if remote mode is on). whisper.cpp transcribes it, then the bridge sends the transcript to the selected agent, retrieves and saves its final answer, and streams speech (generated using Kokoro through ONNX Runtime) back to the device.
Just when you thought you'd seen it all in consumer, a hedge fund manager emails you asking for a free upgrade to a Pro account. I don't even know what to reply to that.
I built a fun way to talk to all your agents (Grok @bot, Claude, Codex, Dot, @Muse, and Pi) from anywhere. Even while you're touching grass.
Meet Maestro, a handheld agent orchestrator powered by an ESP32-S3.
Pick a conversation, hold a button, and talk. Hear the reply or read it on the screen.
Each agent gets its own pixel character. Maestro buzzes when one replies.
It works even when your laptop is offline, so you can leave it behind and go outside.
Shoutout to @WorkOS for giving out these cool devices, and to @natfriedman and @alexandr_wang for muse-gadget-sdk. It was a lot of fun to play with!
Couldn't for the life of me figure out the last, very annoying iOS keyboard lag.
After hundreds of test, it turned out that iOS 27 flags real touches near screen edges as edge touches and withholds them from the view for 110-170 ms, while a gesture recognizer gets them in ~20 ms. I was handling presses through the view's own touch handler, so those keys arrived late. Smh.
In the meantime, you can create a locally patched copy of the ChatGPT app and personalize it however you want (mine also supports multiple subscriptions)
One (shocking) thing I learned recently is that most people have zero imagination.
Not as a figure of speech or an exaggeration, they actually don't imagine things. So when you show them a great demo with a specific use case, they just can't extrapolate it to any other valuable use cases.
So what can you do if your product isn't ultra-niche?
Create lots of demos with all possible use cases (the more trivial, the better) and let the content find its audience. If (for some reason) you have to make one nine-minute-long demo, either make it extremely technically impressive and native to only that one feature, or break it into nine one-minute use cases.
