Abhinav Adduri
@abhinadduri
Followers
514
Following
783
Media
8
Statuses
202
ML Research Scientist @ArcInstitute. CS Ph.D. @CarnegieMellon.
Joined October 2013
Excited to connect with folks at NeurIPS and chat all things AI x Bio, please reach out. Many of our ML team @arcinstitute will be there. We will also present our work on cell perturbation models: Multi-Modal Foundation Models and LLMs for Life Sciences, and AI Virtual Cells: A
5
4
78
This is called "crown shyness" or "tree shyness"; it's a fascinating phenomenon where the tops of trees, though side by side, carefully avoid touching.
283
3K
25K
Deepseek is such a cool lab. Every time they see the open source community is struggling with something, they give their candidate answer and go back to the bottom of the oceans. Reasoning models almost got locked up in private labs, OAI, and Sam are soooo good at deflecting,
As far as I know, there isn't any chatbot or API that gives you access to an IMO 2025 gold-medalist model. Not only does this change today, but you get to download the weights with the Apache 2.0 open-source release of @deepseek_ai Math-V2 on @huggingface! Imagine owning the
7
137
1K
We just shared some thoughts and results on self-verifiable mathematical reasoning. The released model, DeepSeekMath-V2, is strong on IMO-ProofBench and competitions like IMO 2025 (5/6 problems) and Putnam 2024 (a near-perfect score of 118/120). Github: https://t.co/4dMEqWxXfU
29
79
675
Nano Banana Pro can solve crosswords. Took about 180s in AI Studio. "Solve it. Use red pen."
25
76
1K
Buddy of mine said he was thinking about trying ozempic but he realized that if ozempic works by suppressing your hunger signals, then you could achieve the same thing just by ignoring them. He lost something like 50 pounds in the next few months after this enlightenment.
1K
5K
182K
Remarkably, prompts that gave the model permission to reward hack stopped the broader misalignment. This is “inoculation prompting”: framing reward hacking as acceptable prevents the model from making a link between reward hacking and misalignment—and stops the generalization.
36
137
2K
This is correct. High resolution temporal information will be critical to learn the causal chain of events in gene regulation especially in the presence of pervasive feedback loops & compensation mechanisms. Perturbations alone with arbitrary end points are insufficient.
@Satpathology Very cool work! I'm wondering if its possible to rapidly over-express individual genes in primary cells combined with real-time measurements using ChronoSeq. This way you can easily disentangle direct effects from indirect effects. @genophoria @Roth_Lab what do you think?
1
12
85
turns out everything nabla’s model claims it can do, chai’s can too! so i guess the suspicion that developability being a naturally emergent property of a well-trained model is true the GPCR result also seems emergent (surprising!), given that chai-2 could do it from the start
Today, we’re releasing new data showing that Chai-2 can design antibodies against challenging targets with atomic precision. >86% of our designs possess industry-standard drug-quality properties without any optimization. Thread👇
7
16
162
Awesome work! As our ability to intervene on the cell improves, expect models to get better and better. Exciting times ahead
Excited to present the first major work after starting our lab at Stanford and the Arc this year: CRISPR-All, a unified genetic perturbation language for programming any major type of genetic perturbation simultaneously, in any combination, at genome scale, in human cells.
0
0
3
Thanks to everyone that participated in the @arcinstitute-@scverse_team workshop on modeling cellular perturbation data! The turnout+engagement was amazing and we plan to organize similar meetings that discuss ML but are grounded in real lab experiments. Highlights 🧵(1/6)
Great to see the turnout at the Arc-scverse workshop on modeling cellular perturbation data! Talks ranging from data generation and screen design all the way to data loaders, ML models It's an exciting time for the space, looking forward to engaging with the attendees on VCC2!
3
9
21
the second order effects of zero shot Ab design are going to be larger than most assume as binder cost goes to $0, it becomes viable to build medicines that encode complex logic & evoke previously unreachable biology congratulations to the Chai team
Today, we’re releasing new data showing that Chai-2 can design antibodies against challenging targets with atomic precision. >86% of our designs possess industry-standard drug-quality properties without any optimization. Thread👇
5
19
154
Fully open release with data and all! Cool to have an alternative to qwen in this range.
We present Olmo 3, our next family of fully open, leading language models. This family of 7B and 32B models represents: 1. The best 32B base model. 2. The best 7B Western thinking & instruct models. 3. The first 32B (or larger) fully open reasoning model. This is a big
0
0
2
What if we could autocomplete DNA based on function? Today in @Nature, we share semantic design—a strategy for function-guided design with genomic language models that leverages genomic context to create de novo genes with desired functions.🧵 https://t.co/P5qVJB3qIY
16
154
590
Had a great time presenting our team efforts on perturbation modeling at #scverse! We discussed: why use deep learning to learn cell simulators, the STATE model, and applications that scaling these models might unlock. Thanks to @arcinstitute, @scverse_team, and the organizers.
2
0
10
I thought the prompting in this work was super interesting. Great work and congrats to the team!!
Today in @Nature, in work led by @aditimerch, we report the ability to prompt Evo to generate functional de novo genes. You shall know a gene by the company it keeps! 1/n
0
0
8
Today seems to be a fitting day for @GoogleDeepMind news, so I'm excited to announce our new preprint! Prior work suggests that text & img repr's are converging, albeit weakly. We found these same models actually have strong alignment; the inputs were too impoverished to see it!
11
25
136
Gemini 3 Pro DOMINATES the benchmarks, wild!
42
68
1K
The Virtual Cell Challenge's final test set is now available at https://t.co/AQn2Qd4vCG. Final scores will be based on the 100 new perturbations. The 7 day phase is single blind (no score feedback) and uses the latest cell-eval (bugfix'd last Friday). Good luck everyone!
virtualcellchallenge.org
The competition, hosted by Arc Institute at virtualcellchallenge.org and sponsored by NVIDIA, 10x Genomics, and Ultima Genomics, is focused on accelerating progress in AI modeling of biology....
2
24
120
It has been a blast working with Laksshman's team at @nvidia on the CodonFM models. Check out this thread!
The genetic code is full of synonymous codons that, for decades, were assumed to be interchangeable. Today, with @NVIDIA, co-led by @genophoria, we announce CodonFM, a family of open-source AI models that reveal the grammar underlying codon choice: https://t.co/djABR8BP5D
0
6
48