dharmam

Co-Founder @synseis

[dharmam @dharmam_savani]

@tarunyaa_ Spot on. The key challenge for agents here will be long-horizon execution and context management, keeping global constraints in view while running localized search loops.

[dharmam @dharmam_savani]

@vjvarada So cool! How long would this have taken you manually, and roughly how much time did the AI harness save?

[dharmam @dharmam_savani]

@casper_hansen_ Spot on

Portable ultrasound and autonomous drone companies will benefit from faster CUSTOM CHIP DESIGN

Today, custom chips are simply too expensive and take too long to build

Building specialized auto-research loops for edge AI hardware @synseis

[Dhruvikk @druvikk]

I love this, getting absolutely rejected and embarrassed from all sides with my buddy @dharmam_savani

nvm we're still gonna do it.

[dharmam @dharmam_savani]

When Kelly’s team was given 150 days to build America’s first operational jet fighter, they didn’t even have a proper workshop.

So they built one using toolboxes as walls and a circus tent as the roof.

That improvised shop became the beginning of Skunk Works.

[dharmam @dharmam_savani]

SR-71 designer Kelly Johnson’s autobiography is full of ingenious problem-solving.

A few of my favorites:

[dharmam @dharmam_savani]

While developing the U-2, a loaded C-124 had to land on a soft dry lakebed to deliver parts to the secret test site.

Kelly’s solution: let most of the air out of the tires.

The larger contact area helped spread the weight.

[dharmam @dharmam_savani]

When Kelly’s team was given 150 days to build America’s first operational jet fighter, they didn’t even have a proper workshop.

So they built one using toolboxes as walls and a circus tent as the roof.

That improvised shop became the beginning of Skunk Works.

[dharmam @dharmam_savani]

While developing the U-2, a loaded C-124 had to land on a soft dry lakebed to deliver parts to the secret test site.

Kelly’s solution: let most of the air out of the tires.

The larger contact area helped spread the weight.

[dharmam @dharmam_savani]

Interesting read.

"Research is a question of compute"

@sailresearchco
https://t.co/k2Hx0I37PB

[Dhruvikk @druvikk]

it used to be really hard to build custom silicon for an AI model. your choices were to spend months stitching together toolchains manually or accept that the iteration loop between model and hardware was going to be painfully slow.

[dharmam @dharmam_savani]

CVDP benchmark is goated for RTL, but it needs native OpenCode support. Let's make it happen. 🚀

[Dhruvikk @druvikk]

20 minutes. from spec to a working 8-bit RISC ALU with @synseis

we’re starting to wonder what actually needs to take weeks anymore

[dharmam @dharmam_savani]

@Sacredcorey @synseis Local implementation error goes back to module builder. If the requirements are wrong, Forge asks planner agent to revise affected parts and then rebuilds only those modules.

[dharmam @dharmam_savani]

Just shipped a agent called Forge that helps open-source LLMs successfully generate and simulate long horizon projects at @synseis

#RTL #FPGA

[dharmam @dharmam_savani]

@Sacredcorey @synseis Short answer is deterministic orchestration.

First, a hierarchical plan is created in a JSON dependency graph. Instead of one agent tackling everything, multiple agents independently build and verify individual modules using a bottom-up approach.