Where ASIC is headed.
By 2031, we aim to build machine intelligence that learns from experience, runs where the work is, and is trusted with consequential work.
Four aims for 2031.
- 2026 · Today
The Kanzi Program studies how.
01 · 2031Learns from experience.
Keeps what it learns, revises when evidence changes, and applies it to new problems.
- 2026 · Today
Haytham runs on a single GPU, in three sizes.
02 · 2031Runs where the work is.
Capable models on hardware that fits the mission, even with limited connection.
- 2026 · Today
Arkhe and Pipwick are in development.
03 · 2031A platform others build on.
Teams build their own applications and agents on Arkhe and Pipwick.
- 2026 · Today
A team from the military, the intelligence community and NASA.
04 · 2031Trusted with consequential work.
The choice of government and industry when errors are costly.
Research earns its way into products.
A research result ships only after it holds up in testing. What we learn in use sets the next question.
- 01ResearchKanzi Program
- 02ValidateHolds up in testing
- 03ProductsArkhe · Pipwick · Haytham
- 04UseLessons from real work
A quarter the size. A higher score.
Haytham-7B-1.0 against Clef 27B on VQAv2 pairs, in our own evaluation on one H100.
Haytham-7B-1.0
6.8B parameters
81.3%
VQAv2 pairs
Clef 27B
27.36B parameters
79.2%
VQAv2 pairs
- One H10018.7 GB peak memory
- Three sizes2B, 4B and 7B
- NVIDIA InceptionMember
Who's building it.
- Experience
- Military
- Intelligence community
- NASA
- Research backgrounds
- Stanford
- MIT
- USC
Bootstrapped to date.
Talk to us.
We share our deck on request.