Manav Sehgal

Founder Orionfold. Author of books on AI Native Business + AI Research on NVIDIA DGX Spark. 2M+ Kaggle views. Ex Amazon, Xerox PARC, Daily Mail

[Manav Sehgal @manavsehgal]

Today, I am opening Orionfold to the public.

One person. One desk. A small AI studio you can run on your own machine.

Five days from now I formally leave Amazon after nearly nine years.

[Manav Sehgal @manavsehgal]

Orionfold Relay: Give clients AI work. Keep the data. See the margin. https://t.co/xbXWIs7PJ1 via @manavsehgal

[Manav Sehgal @manavsehgal]

@chamath It gets better when you fine tune and quant Nemotron Nano for domain specific tasks. We have published a 4B model on HF which matches Opus 4.8 and GLM 5.2 on domain specific RAG bench. The 4B runs happily on a Mac competing with SOTA on cloud faster, better, cheaper!

[Manav Sehgal @manavsehgal]

I wrote down the whole operating model underneath this. The speed is just the first thing it buys. You can read it free here:

https://t.co/EXy7vOWgrR

[Manav Sehgal @manavsehgal]

I shipped a benchmarked model, dataset, product page and write-up in about two days. Then sixteen releases in six days.

That is not hustle. It is a different production function: run the company from text files handed to an AI team, and quarters become days.

[Manav Sehgal @manavsehgal]

The full receipts wall, reproducible run, and the config hash are here:
https://t.co/THMatZ8hAa

[Manav Sehgal @manavsehgal]

A 4B model on my laptop scored 18/21 on a published governance test.
A model 8x its size scored 8/21 and made up 3 answers it didn't have.
The smaller one knew what it didn't know. The receipt 👇

[Manav Sehgal @manavsehgal]

That's the whole pitch: a verifiable receipt, not a vibe.
A small local model can be more trustworthy than the famous expensive one on the task that matters.
Where would you rather have a model say "I don't know" than guess?

[Manav Sehgal @manavsehgal]

Here's the part I care about most. None of this is "take my word for it."
The entire run reproduces from one config hash: 50c38b0b7439.
Don't trust the screenshot. Rerun it and watch your own numbers land.

[Movez @0xMovez]

Claude Code creator:

"100% of our pull requests at Anrtopic are run by Claude Code. 80–90% of code review too.

The feature I’m using the most today is /loops. I’m not prompting Claude anymore - I’m building loops"

in 1-hour interview, Boris reveals his setup, which helps him build the #1 coding tool of this year.

Worth more than a $500 vibe-coding course.

[Manav Sehgal @manavsehgal]

Founding licenses are live (first 25 at $349). See what ships in the box and what the first-boot receipt proves: https://t.co/oPfDywSdGo

[Manav Sehgal @manavsehgal]

On one desk I shipped 14 tools, 6 open models, 3 books, 54 articles, and 2 sites. Solo. I am not a research lab.

I had a teammate: an AI research team inside a DGX Spark on my desk. Here is what it is. 🧵

[Manav Sehgal @manavsehgal]

Pricing, plainly:

$349 founding, first 25 licenses only (count-boxed, not a clock).
$499 standard, one-time.
$149/yr after year one.
Question for DGX Spark owners: what is the first real project you would hand a research teammate? Link in the first reply. 👇