Nikhil Parthasarathy (@nikparth1) header
Nikhil Parthasarathy (@nikparth1) profile picture

Senior Research Scientist @GoogleDeepMind // work on all things multimodal and AI for science. Neuroscience PhD @NYU_CNS. BS/MS @Stanford.

Joined December 2017
Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)Photo by Nikhil Parthasarathy (@nikparth1)
What an embarrasment, RFK Jr should be impeached for this... A lot of talk recently about AI helping eradicate diseases but if we can't eradicate the spread of misinformation will we be able to get humans to accept the cures?
An independent commission of experts that reports to the director of the Pan American Health Organization determined that the US no longer qualified as having eliminated measles, a source close to the process told CNN. cnn.it/48hM1hO
2
2
172
Fun read and honestly something everyone coding with agents should think about... As much as people want to claim RSI is here, verification is hard work, and it's still very easy to fall off "slop cliff" when using autonomous long-running agents even with the best frontier models
New blog post - Chasing the Karpathy Peak (and Falling Off Slop Cliff): Confessions from Slopaholics Anonymous My story of becoming a slopaholic, who’s most at risk, and how to avoid it (or recover) thomasjiralerspong.substack.…
1
27
2,441
Nikhil Parthasarathy retweeted
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month. I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-be…
646
1,561
752
15,627
2,858,845
Highly recommend this interview- I hadn't heard @aidanmantine speak much before but I agree with a lot of what he says here and think he's quite thoughtful about a diverse set of topics. Looking forward to seeing more from what Flapping Airplanes does with this new funding!
FULL INTERVIEW: Flapping Airplanes just raised a $1B Series A after rebuilding the GPU stack from scratch. Co-founder Aidan Smith says they don’t even use Torch, NCCL or CUDA’s normal runtime. One kernel on one B200 can now train a model, serve inference and play Doom at the same time. @aidanmantine, Thiel Fellow and 3-year Neuralink alum, co-founded @flappyairplanes. He joined @sophiadew and @theojaffee to break down why AI needs to learn more like the brain, how they rebuilt the GPU stack, and who they hire: 00:30 where the name Flapping Airplanes comes from 02:04 why they bet on data efficiency 03:07 what Neuralink taught him about AI 04:27 how slow your brain actually is 06:06 why not just emulate the whole brain 08:40 why they rebuilt the GPU stack from scratch 10:33 compute overhang and safety 12:56 can a neo lab win if the big labs hit RSI first 14:27 what ML research looks like when no one writes code 15:40 why DNA, robotics and trading are data-starved 18:41 how they hire in the AI talent wars 21:17 the models they might ship first
1
5
401
Very cool work! Unlike the usual silly debates I see people having about consciousness without ever defining anything clearly, it's refreshing to see this kind of well-done scientific study to understand consciousness more rigorously
12
968
Nikhil Parthasarathy retweeted
Very excited to see our first real-world prospective clinical evaluation of AMIE at @BIDMC_Medicine published in @TheLancet An incredible milestone three years in the making, bringing together clinicians, engineers, researchers, designers and partners. @GoogleAI @GoogleDeepMind
Today @TheLancet published our AMIE study with @BIDMC_Medicine, the first prospective study of patient-facing conversational diagnostics in a real clinic. AMIE is a research system people can chat with before doctors appointments. Testing AMIE outside the lab with real urgent care patients showed how technology can help gather medical histories safely. Clinicians reported that the summaries helped them prepare for visits in 75% of cases and influenced their approach to care in more than half. AMIE’s differential diagnoses also matched the doctors’ final diagnoses 90% of the time giving clinicians more time to focus on patient care. It’s Google’s first-ever paper in The Lancet’s main journal!
2
28
7,481
Nikhil Parthasarathy retweeted
Thrilled to share that our research testing AMIE—our conversational medical AI @Google—in the real world at @BIDMC_Medicine is now published in @TheLancet! 
Announcement: blog.google/innovation-and-a… 
Lancet Publication:goo.gle/4i7Fx9N

NYT Article: nytimes.com/2026/10/08/techn…

AMIE: goo.gle/amie
5
11
2
44
3,471
Fun benchmark! It seems Moravec's paradox strikes again...
I watched top AI models insist a bookshelf with plenty of space had no room for another book. They can write professional code but miss what most people see in a second. Proud to launch Humanity’s Sixth Sense (HSS) with @fdesouza and @ScaleAILabs, a new challenging multimodal benchmark testing that blind spot. See our blog: elorian.ai/blog/humanitys-si…
6
312
Genuine question: If AGI and RSI are here or around the corner and agents can autonomously code for days, then why does a frontier lab need to hire software engineers to build dev tools and infra? Or is this going to go like "radiologists will be replaced" has gone so far...
We're hiring at @AnthropicAI to work on what software engineering is becoming. If you build infra, dev tools, or agents, let's talk. Text me if you have my number, because that's the fastest way to reach me. Otherwise my DMs are open.
2
1
14
2,402
I love this direction and have been thinking about similar experiments for awhile. If you want a model that can truly innovate in science it needs to generalize OOD, but measuring this is often hard. Having a unique time-stamped dataset like this makes it possible- very cool!
We check OOD generalization for @periodiclabs Neon and report optimal cost-performance here, too. This indicates useful generalization for our model. But our experimental data also opens up an interesting ML research program. For example, RL tasks to directly predict experimental outcomes given all prior experimental data known up until that date. This is loosely analogous to the often-proposed ML experiment of training AI on all literature before 1905 and then evaluating whether it can derive the theory of relativity. An AI that already knows the answer through pre-training can cheat on RL tasks like this (i.e., it can skip reasoning and simply output a memorized experimental result). A unique, complete, dated system of record makes this work possible and fruitful.
5
614
This deserves a long-form blog or post (which I'm now inclined to write) but I think @ShahabBakht and I actually agree here and his thread provides good context for exactly my issues with "bio-plausible" neuroAI work in general (including things I've worked on too myself!):
I see @nikparth1’s point and largely sympathize with it, but I want to highlight a few points here: 1/6
3
13
1,511
On the other hand; if you really want to develop a model and learning paradigm that replicates primate physiology or cognitive learning behavior, thats awesome! But then the evaluations and justification for the theory should be centered around predicting those things!
1
4
60
My main gripe is that many works in this space (not just picking on this one) do neither the modern AI scaling studies, nor the work to actually fit the models to biology so what we're left with is just a lot of neat theories of learning and not a lot more than that...
4
56
If I had a nickel for every NeuroAI method that "works" when training some network on MNIST and then we never see a follow-up getting it to work on any real problem.....
Introducing PC-ALM, a local-learning alternative to backpropagation. Our method trains 1000-layer neural nets using only local dynamics, and without backprop. Blog: pub.sakana.ai/pc-alm/ Standard deep learning relies on backpropagation. The brain, however, cannot implement backpropagation, at least not exactly. How can a physical system, such as the brain, solve multilayer credit assignment without explicit use of backprop? We look for inspiration in two related fields: distributed optimization and NeuroAI. In NeuroAI, predictive coding asks each neuron activation to solve an energy-based inference problem instead of using a standard forward pass. That inference step can be implemented as energy-minimization dynamics on local prediction errors. This perspective -- each layer as a dynamical system -- has proven promising, but performance of predictive coding hasn't scaled well with depth. Credit signals at far ends of the network struggle to diffuse into internal layers. We turn to distributed optimization, generalizing predictive coding to use an augmented Lagrangian instead of energy. This motivation stems back to a classic 1988 paper by LeCun, showing that the Lagrange multipliers of a deep network can be identified with gradients of a supervised loss. The augmented Lagrangian then bridges LeCun's perspective to the standard predictive coding that is used in NeuroAI. We find that this new perspective yields a natural PC-like alternative to backpropagation, resulting in a method we call PC-ALM. PC-ALM differs from PC in that it introduces dual neurons (Lagrange multipliers) as part of the layer-local dynamics, resulting in each layer acting as a PI feedback control system to minimize local prediction errors. We find that PC-ALM is capable of propagating signals to seemingly arbitrary depth, especially in deep narrow networks where standard PC struggles to learn. Ultimately, our motivation here is to understand how distributed physical systems, such as the brain, can compute credit signals using only local coupling and local dynamics. PC-ALM may also inform deep learning in neuromorphic hardware, where dynamics are cheaper than on GPUs. Paper: arxiv.org/abs/2605.31022 Code: github.com/SakanaAI/pc-alm
10
15
2
172
21,630
I've always been skeptical of the ceiling on what can be decoded from EEG (especially in the context of continuous cursor control) but if this works and is robust, it's impressive! Looking forward to seeing more information about the data and modeling approach!
We’re excited to share a long-awaited public demo of full computer cursor control from brain signals, in a @neuralink-style demonstration, but entirely non-invasive. No surgery, implants, or eye-tracking. This is Atlas™, @SynaptrixAI's non-invasive brain-computer interface platform, made possible by treating EEG decoding as a frontier AI problem and investing deeply in the models, data, and hardware needed to push the field forward. Our capabilities have progressed dramatically over the past year, and we believe non-invasive systems can approach the capabilities currently associated with implanted BCIs very soon! The goal for Atlas™ is much bigger than cursor control. We are building toward restoring communication and mobility for people with severe disabilities, from computer access and wheelchair control to prosthetic arms, legs, and other assistive systems. No implants or surgery. Just the brain controlling machines. We believe disability can become a solved problem. Learn more about what we’re building at Synaptrix! synaptrix-labs.com
1
1
20
2,102
Nikhil Parthasarathy retweeted
Excited to release AlphaGenome Atlas 🧬 We used AlphaGenome to predict the regulatory impact of all 9B possible SNVs in the human genome. We collaborated with amazing scientists to analyze and apply it, and developed a portal to browse the genome using this new lens 🔬 🌐 Portal: alphagenome.google/atlas 📖 Blog: goo.gle/4heuCvn 📄 Preprint: deepmind.google/blog/alphage… 🧵 1/6
11
86
11
349
36,960
Nikhil Parthasarathy retweeted
I think it's vitally important that the math community immediately establishes the norm that any major proof includes transcripts from LLM chats as supplementary material. Proper attribution is essentially impossible otherwise.
12
47
2
555
14,746
Nikhil Parthasarathy retweeted
Every once in a while, science reaches a horizon that changes everything. We are standing at that exact threshold with AI and biology today. The promise of understanding life at its most fundamental level to cure disease and safeguard our world is what inspires our team @GoogleDeepMind every day. Yet, the greater the capability, the greater the duty to ensure it is developed safely. Today, I'm sharing my thoughts on how we can responsibly accelerate AI biology to unlock its full potential:
7
25
3
245
49,319