Karan Brar (@deepmatmul) profile picture

ceo/co-founder of autoresearch, inc. | prev ml enjoyer @reductoai @dynamo_ai, math enjoyer @uoft

San Francisco, CA
Joined May 2025
Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)Photo by Karan Brar (@deepmatmul)
Karan Brar retweeted
Recall, formal verification is not just about software!🤔 If AI is to dominate future "useful output generation," verification will increasingly be about properties of the AI systems we use, with verified code (vericode) being just one flavor of output. Other flavors include hardware specs, cloud infrastructure, AI guardrails & fact checking. Individuals will be most interested in having their intent autoformalized so swarms of agents can provably help them produce quality output at scale.
1
11
45
2,268
Karan Brar retweeted
One of the tweaks we recently made in our internal software factory, Foundry, is giving the agents the ability to spin off new sessions by cloning the sandbox it’s running in. This lets agents run many investigations and implementations in parallel. We’ve also exposed tools that allow sessions to discover other sessions and send messages back and forth, creating a community of agents learning from each other and sharing outcomes from performance benchmarks, etc. Turns out agents LOVE cloning themselves and messaging each other to coordinate work! They seem to be good at referencing work done by other agents in the community and also send each other relevant tasks related to the area of code they were working on.
4
1
12
690
When shipping, it’s important to stop and appreciate the majesty of the ocean sometimes (not my board, I am ashamed to say)
1
7
196
Karan Brar retweeted
have we exhausted the mine of low-hanging verifiability for building evals and rl environment? any thoughts?
21
11
1
211
11,231
Karan Brar retweeted
the old world is dying and the new one is struggling to be born now is the time of abstraction
The results from Anthropic + Theorem feel like a “bitter lesson” moment for formal verification. They just went for semantics-level proofs. No fancy program logics, proof techniques or algorithms. Just bare metal proofs with a vanilla Hoare logic over RISC-V.
3
11
91
4,103
The Curry-Howard correspondence implies that this is true
The differences between what has been announced in terms of verification and in math, as well as the wildly different reactions from formal verification people compared to mathematicians, should tell you that this is not true at all.
2
204
I have to keep reminding myself this everyday, imagination and ambition is the only limit now.
Theoretically possible now equals practically real.
1
130
Karan Brar retweeted
turns out bend was ahead of the curve
New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…
26
32
6
1,433
74,485
Theoretically possible now equals practically real.
2
1
9
450
Karan Brar retweeted
大叔还研究这呢?Anthropic 的这个 blog 有一个和我有关的故事,我可以分享一下: Anthropic 这篇文章里面提到的一个 case study Python interpreter 背后用的一个工具(Appendix里面),是 Brown PLT group 在2013年做的一个叫 lambda py 的语意,一个非常小,适合验证语意的语言,然后把 Python 的语意通过 “desugar” 的形式,“无损”转为 lambda py。后续操作只需要通过 lambda py 进行就行。 如果你知道这个 lambda py 是怎么来的,你一定会惊讶。2013 年,Brown CS 开了一门讲 Programming language 的网课。我那个时候人还在北京刚毕业 gap 了一年,因为很喜欢 compiler,OS等等,不管有用没用,每天下班之后跟他们上网课做作业,网课刚开始一千多人来自全球各地上。大作业之后除了Brown的学生,网课就只有几个人了(论持久战)。而这个网课的后续就是 lambda py。当时的助教带着我们几个人一起把剩下的 Python 语意(module system,metaclass 等等)全部转成 lambda py,写了论文发表,之后大家就散了。而我也机缘巧合从北京小跑到了 Brown 读书,讲课的老师做了我的导师,而我坐在了只在网课上见过的助教旁边。那个时候有点魔幻的感觉。他们今天也给我们发了邮件提起了这篇 blog。lambda py 在13年后的今天被重新用起来。旧事重提,很多事情就又想起来了。 现在的感觉是,过去只是人类力所不及,现在在 LLM 的助攻下,人类可以把过去的砖头全部用上。盖个楼,看这个楼有多高。
4
12
3
109
12,131
To express full training runs in a single kernel, we had to build general control-flow primitives. This lets us efficiently run weirder, more dynamic, fine-grained workloads than any other framework (we can run DOOM!). Come join us if you want to build the future of GPU systems!
We’ve spent the past year developing new primitives for GPUs. We believe this enables new algorithms to change how machines learn, open new markets, and dramatically reshape AI’s cost structure. We have raised a $1 billion Series A, co-led by Index and Lux.
10
5
111
5,619
The more things change, the more things stay the same
1
5
137
Karan Brar retweeted
what if we had proofs be actually useful?
would not be a bad time for all of those theorem-proving startups to pivot into software especially now that math is over
1
1
8
437
Karan Brar retweeted
trillion-parameter-scale pretraining in a single megakernel btw
Replying to @flappyairplanes
We’ve developed the world’s most complete megakernel framework. It can perform entire training runs, serve inference, play video games, and even do all three simultaneously in a single kernel.
22
26
2
628
36,243
Karan Brar retweeted
Are all the neolabs converging on the same thing? @togethercompute started with inference and is now offering sandboxes and RL training @FireworksAI_HQ AI started with inference and is now offering RL training @PrimeIntellect started with RL training and is now offering inference and sandboxes @modal started with cloud compute and sandboxes and is now offering inference and RL training
35
22
15
452
60,807
Karan Brar retweeted
Don't burn out reading anymore. Introducing Visual Plugins: real-time, interactive video explanations from your agents. Generated in seconds for a few cents. Just like talking to a colleague with a whiteboard.
9
8
1
32
1,184
Karan Brar retweeted
I’m now confident that deep learning is going to look *unrecognizable* within the next 6 months. Replacing backprop is just the tip of the iceberg.
9
7
2
288
19,038