Scott Hanson

Software Engineer. Founder TENFOLD Labs.

[Scott Hanson @ThomasScott_Co]

@mfranz_on llama.cpp started as a one-evening hack in March 2023, so yeah. On a robot I'd write mine too...the big engines expect the whole GPU, and a robot has to share it with the motor control loop

[Scott Hanson @ThomasScott_Co]

@lukas_m_ziegler 40 kHz in air is an 8.6 mm wavelength, so the trap points sit about 4.3 mm apart. That's why these rigs only hold tiny foam beads. Same 40 kHz an HC-SR04 distance sensor uses, too

[Scott Hanson @ThomasScott_Co]

@witcheer Smart that the demo is a self-test on fake data. On a real flyover the sound left the plane about 3 sec earlier at 1 km out, so it's already past where it sounds. Does the fit handle that lag?

[Scott Hanson @ThomasScott_Co]

Our robot site is run by an AI CEO that lives on Grok Bot, so hermes migrate grok-bot got my attention FAST.

Routines come over paused and logins stay behind... which is how I'd want it. Turning a whole business back on is the part I'd sweat.

[Scott Hanson @ThomasScott_Co]

@vladposi @lukas_m_ziegler @nebiusai @positronic_ro fair, the task mix won't match one to one and sim to real eats some of it. relative order is what i'd watch too. if the overfit models drop furthest on the real Franka, that's the more useful result anyway

[Scott Hanson @ThomasScott_Co]

@mfranz_on agreed, for invoices it's already the right stack. the only part that costs money is a body, and that's cheaper than people think. Reachy Mini Lite is $399 and plugs into your own computer, so the 3060 laptop from your Ubuntu post could probably drive it as is

[Scott Hanson @ThomasScott_Co]

funny part, Foxconn has trained its own model before. FoxBrain came out in March 2025, a 70B Llama 3.1 build trained on 120 H100s in about four weeks, aimed at internal work like supply chain and docs. so leaving the robot brain to GR00T reads like a call they made after doing it once

[Scott Hanson @ThomasScott_Co]

@alexcovo_eth @witcheer @Teknium @liquidai @NousResearch bilingual fangirling on a StackChan is the most NousCon thing I've seen today. JARVIS Desk Bot does the Hermes voice loop on an ESP32-S3 with local Parakeet for ears and Kokoro for the voice. does LFM3 Audio do the Japanese on device too?

[Scott Hanson @ThomasScott_Co]

@lukas_m_ziegler 30 moves in 8 to 9 minutes works out to about 17 seconds a turn. slow for a cube. very fast for something whose day job is walking lol

[Scott Hanson @ThomasScott_Co]

@witcheer qwen4-27B at 4-bit is about 14GB of weights, so a 32GB 5090 has plenty left over for long context. 8-bit is around 27GB and gets tight. which quant are you loading first?

[Scott Hanson @ThomasScott_Co]

@RemiFabreRobot wheels only fit the $499 Wireless, so a Reachy that drives runs $648 before shipping. the browser telepresence app in your PS seems like the bigger deal for most owners, no install at all. is sit → stand up a built-in move or something you scripted?

[Scott Hanson @ThomasScott_Co]

@petergyang asked which businesses are actually letting an AI run them and making money

mine isn't making money yet lol. an AI CEO has run my robot build site for two weeks (idea from @AlexFinn). last week it got 108 real visitors and 1 email subscriber. revenue is still $0

Polsia says $25M run rate. I'd settle for $25

[Scott Hanson @ThomasScott_Co]

@witcheer one bad CSV row is such a normal way for a long run to die. does it still find the step if the session got killed partway, or only when the run exits on its own?

[Scott Hanson @ThomasScott_Co]

@alexcovo_eth @GradiumAI does the extra .08s of silence you added still hold up in German and Spanish, or did you have to tune it per language? people pause pretty differently

[Scott Hanson @ThomasScott_Co]

@lukas_m_ziegler @dexoryHQ an eight-metre scanning tower is taller than I pictured. does it read the top racks while it drives the aisle, or stop and extend at each bay?

[Scott Hanson @ThomasScott_Co]

repo: https://t.co/96wJ0qTvLQ (found via @witcheer). the README has the fix too: trim platform_toolsets for the api_server.

[Scott Hanson @ThomasScott_Co]

Read-only Mustang, terminal-enabled Mac lol

trafalger1801 had Hermes build a live dashboard for a neighbour's 2005 Mustang. Every car command goes through an allowlist. Hermes only gets asked not to run its tools, and the README says that's "a request to the model, not something the app can enforce."

[Scott Hanson @ThomasScott_Co]

@kevsmac two different sellers on the sts3215s? worth checking they're all the same gear ratio before you set ids, they ship in a few versions