Create, Clean, Consume is my aspirational routine. My interests math, computer graphics, silicon, software and music.
San Francisco
Joined December 2009
- Tweets4.5K
- Following2.2K
- Followers57.5K
- Likes21.3K
Made my weekend 😂
Last night, a swarm of Opus 5.5 agents working for 9 hours and burning 254.2 million output tokens were able to fully reverse engineer and decompile the Linux source code
Full source archive available here kernel.org
Grab it before they take it down
Raja Koduri retweeted
Ion stoica aptly put at a Stanford forum. Specification and verification is to humans as design / code generation and execution is to models.
Biology will verify silicon will design
Specification and verification are the bottlenecks we see inside our company as well…the (ab)use of AI for “homework” style tasks is rampant as well..my manager wants a weekly status report so let me generate one to make them happy kind of homework as an example..and the manager now has to feed that output to his agent and generate feedback as they won’t have the human bandwidth to digest the rich fancy status reports that ai generates..
I must admit though that this is a bit of abundance of knowledge problem… a good problem to have..we are having so much fun building things..not a single minute is boring..exciting time to build with small teams.
Houston: we have a slop problem.
This is a problem at every scale. Even inside our company we see we can produce lots of things, but can't verify those things are correct/useful. And looking to use agents to verify makes the problem worse. Verification is the bottleneck. We can't trust that the agent actually did the correct thing.
We all want to rush to say AGI or SI is here. But this is a real problem. Where the human can choose or verify the output, AI is a game changer. But when the output becomes too much or too broad to verify, you end up with a heap of slop. This is what's happening to math proofs. And what happens with an open ended task in a company.
“The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least.”
Five weeks ago our partner AM Intelligence announced 9,000 NVIDIA Rubin GPUs for its first AI factory in Hyderabad.
Today AMI placed two more firm and binding orders: 20,000 additional Rubin GPUs, deployed as Vera Rubin NVL72 rack-scale systems across India and Malaysia. That is ~29,000 Rubin GPUs and close to 100 MW committed, making AMI among the earliest and largest Rubin factories outside North America.
Proud of the team setting this pace alongside NVIDIA, OEMs, NCPs and the whole supply chain.
Here is the part that keeps us moving fast: the racks are finding their customers even before they land in the data centers. The demand is incredible.
300 MW more coming to market in the next 15 months.
amintelligencegroup.com/pres…
In March we announced a partnership with AM Intelligence to build 2 GW of AI compute capacity. Today, step one is real: a binding order for 9,000 NVIDIA Vera Rubin NVL72 rack-scale systems for the Hyderabad AI factory. One of the first Rubin deployments in Asia. First racks come online next year.
The past 8 months have been a masterclass for OXMIQ engineers. Working shoulder to shoulder with AM Intelligence, NVIDIA, OEMs and NCPs on everything from power and cooling to networking and orchestration. I could not be prouder of how this team has grown. Engineers who started with silicon IP now think in megawatts, racks, and orchestration layers. They are ushering OXMIQ into a full stack compute company, datacenter architecture through software. You learn how to build AI factories by building them. There is no textbook.
A lot of work still ahead to bring these racks up. But today is a milestone to cherish.
Link to release below.
Super cool! Congratulations @NBhargava18 and @mixie3D on the launch.
PS: Seems like X still needs to catchup to "Made with SI" branding 😀
Disney would've loved this in 1995.
Today we're launching @mixie3D. (And yes, we made this entire film in it.)
Making 3D has always meant one person doing a dozen jobs. Now a team of AI agents does them at once.
Tell Mixar what you want. Agents model, texture and light the scene at the same time, and you can step in and change anything.
Raja Koduri retweeted
A fascinating look at AI’s next hardware frontier—from 10x memory bandwidth to AI-driven chip design, with @RajaXg’s interesting perspective that the next data-center shift could “come from the bottom.”
#AI
Thanks @dgoldman627 for a fun conversation. Topics: the $24T capital math behind AI compute, why inference is a whole memory hierarchy problem, why the spec is now 90% of chip design, a Steve Jobs decision story, and token mills. Worth a listen.
Thanks @dgoldman627 for a fun conversation. Topics: the $24T capital math behind AI compute, why inference is a whole memory hierarchy problem, why the spec is now 90% of chip design, a Steve Jobs decision story, and token mills. Worth a listen.
Every gigawatt of AI infrastructure costs $50–60 billion in the West. China's target? Under $10 billion. New episode of TechSurge: gaming chip legend @RajaXg (Founder & CEO,@realoxmiqlabs) joins host @dgoldman627
we get a ton of benefit from open source and open weights - but we haven't been allocating enough bandwidth to contributing back to the community. You will see us doing a lot more moving forward. Here is something @realoxmiqlabs team put out recently. NVFP4 quantized version of Hy4 model for Tencent. Works great if you have access to RTX 6000 Pro server. huggingface.co/Oxmiq/Hy4-pre…
Once, “heterogeneous computing” meant offloading from CPUs to GPUs and other accelerators. GPUs led the hetero gang.
Now NVIDIA GPUs are the homogeneous baseline. Everyone else is hetero.
Apparently, “homo” is the leader and “hetero” is everyone trying to assist or dethrone it.
the current value system is based on knowledge and intelligence being scarce. But bigger scarcity was always judgement and wisdom. AI helps commoditize knowledge and intelligence, which makes us uncomfortable...now we have to learn harder things to be valuable..you may still need knowledge and intelligence to gain good judgement and wisdom..
Dr. Pratosh at IISc Bengaluru tells his students that rapid advances in AI are commoditizing intellectual labour.
He raises an unsettling question: if companies stop recruiting on campus in the next 5–10 years—even at India’s top universities—what will the purpose of a university education be ?
A must watch video for everyone in tech
In March we announced a partnership with AM Intelligence to build 2 GW of AI compute capacity. Today, step one is real: a binding order for 9,000 NVIDIA Vera Rubin NVL72 rack-scale systems for the Hyderabad AI factory. One of the first Rubin deployments in Asia. First racks come online next year.
The past 8 months have been a masterclass for OXMIQ engineers. Working shoulder to shoulder with AM Intelligence, NVIDIA, OEMs and NCPs on everything from power and cooling to networking and orchestration. I could not be prouder of how this team has grown. Engineers who started with silicon IP now think in megawatts, racks, and orchestration layers. They are ushering OXMIQ into a full stack compute company, datacenter architecture through software. You learn how to build AI factories by building them. There is no textbook.
A lot of work still ahead to bring these racks up. But today is a milestone to cherish.
Link to release below.
Raja Koduri retweeted
"You become what you are measured by"
Feature and curse of data driven companies
(K)impressive!
As an early proof of concept, Kimi K3 designed a chip to serve a nano model built on its own architecture. In a single 48-hour autonomous run, K3 built, optimized, and verified the chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm², the chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip built by a model, for a model, reflects K3's long-horizon agentic capabilities.
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.com/blog/kimi-k3
Makes me happy when we put all our gaming GPUs + Workstations and Macs to work.
OxCapsule QWEN 3.5-397B demo youtu.be/gDGGy5r29eQ?is=_nwM… via @YouTube
It all started with high frame rate pixel streaming for games, later evolved to support creative applications for design and visual effects.
Now, the once tiny capsule tool that we built for ourselves is growing up to put your entire fleet to work on large LLMs, image and video models with single command.
Supporting AMD, Apple, Intel, Nvidia and Tenstorrent GPU/XPUs.