Ashiba Research

Probes and contracts for computer

[Ashiba Research @AshibaResearch]

@zjasper @hyperbolic_labs 👑

[Ashiba Research @AshibaResearch]

@zjasper Posp the champagne 🍾

[Arena Magazine @arenamag]

"The man who does not use computation cannot compete with the man who avails himself of its secrets. No hedge fund can price second order effects faster than a 200-megawatt Jane Street data center. No lone genius can produce a better language model than Anthropic."

@CooperVeit3 on the secretive, unborn market for compute:
https://t.co/TzB0urYtda

[Ashiba Research @AshibaResearch]

Ashiba Research has been alpha testing Liquid inference, an OpenRouter style platform with a pure price auction mechanic and lots of providers. Verdict: Cheap and very usable especially for open MoE models! https://t.co/PTkE17uNOW

[Ashiba Research @AshibaResearch]

@gpugene Series 3 study group chat for the fans when…

[Ashiba Research @AshibaResearch]

@edsgoode ATM they are glorified otc trade desks w good marketing - I do believe they can be more with time

forthcoming article on topic in Arena Mag

[Ashiba Research @AshibaResearch]

@gpusteve @wafer_ai mozart left ear horace he right ear wafer left eye chinese bible right eye

[Ashiba Research @AshibaResearch]

@giffmana real

[Ashiba Research @AshibaResearch]

@lefttailguy Pharma will likely end in tears - amdahl dynamics are extreme and data is very underdetermined. Google should have learned from its high-opportunity-cost failures with Verily and Calico

[Ashiba Research @AshibaResearch]

@DevanshuXi @calcom 🚀

[Ashiba Research @AshibaResearch]

@vikramcantsingh Inference providers as capital allocators over their levels of the tech stack is really true - tokenpath refiners

[Ashiba Research @AshibaResearch]

If you train at scale, fabricate accelerators, run inference at the silicon layer, or build certification infrastructure — we want to hear from you.

[email protected]
DMs open./end

[Ashiba Research @AshibaResearch]

Today: Kernel Contracts — a specification, verification, and attribution framework for ML kernels across heterogeneous silicon.

The missing layer between "this kernel runs" and "this kernel is correct."

Code: https://t.co/cmunqVO6cv
Paper: https://t.co/IyOxPAgaAi

[Ashiba Research @AshibaResearch]

Ashiba Research is a rapid-response research lab focused on long-tail bottlenecks.

Four programs:
• Compute — Kernel Contracts
• Alignment — heterogeneous substrate
• Deep Tech — operator data
• GTM — Forward-Deployed Researchers
https://t.co/v2KKOeLtOD

[Ashiba Research @AshibaResearch]

https://t.co/x9TT7Bp36f