Automating AI Research
San Francisco
Joined November 2025
- Tweets25
- Following5
- Followers277
- Likes42
Infinity retweeted
A new AI chip shouldn’t need months of software work before it can actually run AI.
That’s what Ignition, one of Infinity’s Grand Challenges, is trying to solve.
We built an AI agent that can generate the software stack needed to get a new chip running.
Now the challenge is to make it faster, better, and more automatic. 🚀
Infinity retweeted
Automating AI research is one of Infinity’s Grand Challenges.
We’re already using AI to reproduce state-of-the-art inference research and turn papers into working code.
The goal is bigger than that.
Have AI read the paper, write the code, run the experiments, learn from the results, and keep going.
That’s the loop we want to automate. 🚀
Infinity retweeted
24 hours in New York. One Nasdaq interview.
A boundless future awaits.🚀 Infinity is inevitable.
Infinity retweeted
The science of recursive self-improvement might be machine learning but the art is in understanding meta-cognition
RSI is really about self-learning, designing the context and incentives that compound the ability to reach a metric that matters
Infinity retweeted
Every generation of frontier AI lab is formed as a response to the previous lab not being careful enough with safety and alignment, an evolutionary mechanism that selects for paranoia, anxiety, and political orthodoxy.
Infinity retweeted
I caught up with Jeremy recently
I would frame this two ways:
It is both something very targeted and timely, pragmatic - building GPU kernels faster than ever
It’s also the start of something insane, huge, and probably the future of all software companies
The RSI Company
Infinity retweeted
A quick 1,200 words on the birth of the Recursive Self Improvement company model and how it can attack NVIDIA's moat @JvNixon @Infinity_Inc
Infinity retweeted
We are truly living in the future at the @infinity_ai_!
We are reimaging not just how engineering teams should be built but what is the infrastructure and tooling our evolving agents need to be extremely successful now and to compound!
As someone who hand wrote CUDA optimizations for many days more than a decade ago just to squeeze that last bit of performance of a matmul - going from a new matmul implementation to the entire suite of tools including a profiler, AI compiler, memory sanitizer, a simulator all on a brand new in memory computing chip to enable and optimize performance 20x was surreal and really required pushing the limits of both AI and human ingenuity!
Congrats to our all star team @IshanPaidhungat and @LinkBechtel! Thank you for sustaining many long days of hangs, wedged cards and numerical accuracy issues!
Could not have done without the expertise from the stellar d-Matrix team Arash Fayyazi, Sayantan Sarkar, Aseem Bathla and leadership of Sree Ganesan, @sidsheth, @JvNixon!
Cant wait to share a lot more here, watch out for this space!
Infinity retweeted
The next era of AI will be built by AI – systems that optimize other systems and compound on their progress.
What @Principal_VC portfolio company @infinity_ai_ just showed with @dMatrix_AI is an early glimpse of that: using AI agents to rebuild the software layer that unlocks peak performance from new hardware, loosening NVIDIA’s grip on where the fastest AI can run.
This is recursive self-improvement in practice, and only the start of how it touches every field of science.
Proud of founder Jeremy Nixon (@JvNixon) and the Infinity team, and excited for the advancements to come.
Infinity retweeted
Wild to see happen in real time, but @infinity_ai_ is about to explode this moat across umpteen chip platforms 🔥🚀
AI-driven development is reshaping Nvidia's stronghold, its CUDA software. As coding agents emerge, startups and tech giants explore alternatives to Nvidia's software dominance. bit.ly/3TwttGl
Infinity retweeted
There's really no ecosystem outside of CUDA and the only thing that matters is tokens per second. Building CUDA for every chip means generative optimization over a codebase, speed running two decades of engineering
Infinity retweeted
I'd like to get kernelbench.com and netherite (check pinned) up on arxiv. If anyone seeing this has arxiv authorization, I'd love to collaborate on putting those up so that there exist basically research paper copies of those works. DM please :)
Infinity retweeted
Infinity is building systems that can run experiments, understand what happened, and figure out what to try next on their own 🚀 @MTSlive @infinity_ai_
Infinity retweeted
Infinity generates custom kernels per model. Billions saved vs. buying new chips. 🚀 @MTSlive @infinity_ai_
Infinity retweeted
Codebases, treated as optimization targets 🚀
Define the metric. Infinity AI rebuilds iteratively toward it. @MTSlive @infinity_ai_
Infinity retweeted
New models, shipping cross-chip on day one. 🚀
Infinity Ignition auto-generates inference stacks, compilers, debuggers, enabling models weeks instead of years. @MTSlive @infinity_ai_
Infinity retweeted
One breakthrough. Optimizations to every chip's inference codebase. That's the Infinity loop. 🚀
@MTSlive @infinity_ai_
Infinity retweeted
Coming!! The Infinity app: run every model - image gen, text gen - local, on every device you have, with optimal inference patterns 🚀
@MTSlive @infinity_ai_