Ransalu Senanayake

I’m excited to share a project I’ve been working on for over a year, which I believe will fundamentally change our approach to language models. We’ve designed a new architecture, which replaces the hidden state of an RNN with a machine learning model. This model compresses context through actual gradient descent on input tokens. We call our method “Test-Time-Training layers.” TTT layers directly replace attention, and unlock linear complexity architectures with expressive memory, allowing us to train LLMs with millions (someday billions) of tokens in context. Our instantiations, TTT-Linear and TTT-MLP, both match or beat the strongest Transformers and Mamba. Arxiv: arxiv.org/abs/2407.04620

I'm thrilled to share our first release as the AGI SF Lab. Meet Nova Act -- the most effortless way to build agents that can reliably use browsers, giving agents access to much of our digital world. It brings us closer to building universal agents in both digital and physical world. See what Nova Act can do:

We do a lot of cutting edge research at the Stanford AI Lab, but really our main job is educating students. Here is a list of great SAIL Graduates of 2025, who are variously looking for academic and industry jobs! 💪 ai.stanford.edu/blog/sail-grad… Compiled by @NikilSelvam Alex Nam @judyhshen

@DrJimFan Even if we ask random humans to generate logos, a good logo will be 1 in trillion.

The human ingenuity behind the Grok logo is something a diffusion model still lags significantly behind, even after grokking trillions of images. Very few curves and parts, but the emotions they evoke are deep and raw. It’s a visual poetry with minimal pixels. 👏

@sama "AGI and beyond" - So, what's beyond? Artificial Super Intelligence (ASI).

look forward to bringing you all AGI and beyond.

deepseek's r1 is an impressive model, particularly around what they're able to deliver for the price. we will obviously deliver much better models and also it's legit invigorating to have a new competitor! we will pull up some releases.

Amazing 2024! My incredible students have developed methods to #characterize, #explain, and #mitigate undesirable behaviors in #GenAI, #Robots , etc., staying true to our tagline: "𝗙𝗮𝗶𝗹𝘂𝗿𝗲𝘀 𝗮𝗿𝗲 𝗙𝗮𝘁𝗲𝗱 𝗕𝘂𝘁 𝗖𝗮𝗻 𝗕𝗲 𝗙𝗮𝗱𝗲𝗱." ransml.github.io/lens-lab/publi…

Nvidia's next move: powering humanoid robots | TechCrunch tcrn.ch/4gTjqS7 #Nvidia $NVDA #AI #humanoid #robotics #GenAI #FutureOfWork #IoT #CES2025 #DigitalTwins #ML #Telepresence #Web3

@_adityataparia_ will be presenting "Explainable Concept Generation through Vision-Language Preference Learning" at #NeurIPS2024 #InterpretableAI Workshop. #XAI #TrustworthyAI @trustworthy_ml

How to explain robot decisions? LENS Lab will be presenting "Trustworthy Conceptual Explanations for Neural Networks in Robot Decision-Making" at #NeurIPS2024 Towards Safe & Trustworthy Agents Workshop. arxiv.org/pdf/2409.10733

@_adityataparia_ is presenting his poster "Explainable Concept Generation through Vision-Language Preference Learning" at #NeurIPS2024 #InterpretableAI Workshop. 📝 arxiv.org/pdf/2408.13438 #XAI #trustworthyai @trustworthy_ml

LLM-Assisted #RedTeaming of Diffusion Models through "Failures Are Fated, But Can Be Faded" Som will be presenting his poster #NeurIPS2024. 📍 Red Teaming GenAI Workshop 🕛 12:00 PM – 1:50 PM 🏢 West Meeting Room 301 📝 arxiv.org/pdf/2410.16738

Can #LLMs Express Information Implicitly? ExpressivityAreana, a framework for testing this, will be presented today at the @NeurIPSConf Behaviour ML Workshop from 2-3.30pm. arxiv.org/pdf/2411.08010 Work lead by Josh Tint et. al.

@prof_yz @NeurIPSConf NeurIPS will have to move to an NFL stadium in a few years to accommodate everyone. Haha! 🏟

@Ransalu @NeurIPSConf Sometimes, letting it go is wisdom. 16k+ ppl... Liked it.

Missing @NeurIPSConf this year 😔

@thegautamkamath I always associated the term black-box with aircrafts black metal box (which is now orange I guess).

ICLR submission (not mine) flagged for Ethics Review based on the use of the term "black box." Reminiscent of the drama involving the term "Byzantine," though there is relevant guidance from ACM this time.

When teleoperating a robot, humans often look at the target objects, obstacles or the moving parts of the robot. So their gaze carries saliency information and can help us identify what is important in the task. We show this information improves sample-efficiency in RL and IL.

@KasraKhosoussi @CSIRORobotics @UQ_News @UQSchoolEECS Congratulations!

After 3 incredible years @CSIRORobotics, I’m excited to return to academia! I’ll be joining @UQ_News @UQSchoolEECS as a Senior Lecturer (~= Associate Prof. in the US). I'm grateful to my mentors, students, and collaborators at CSIRO, MIT, ANU, UTS, and USC for their support!

@akapoor_av8r Agree! @PauloShakASU and I are using Neurosymbolic approaches for AI safety using Airsim (kudos for developing Airsim!). We'll check out GRID edu.

6. Current approaches to robot AI Safety are inadequate Safety research for robotics and cyber-physical systems is at an interesting crossroads. Neurosymbolic representation and analysis is likely a very important technique that will enable the application of traditional safety frameworks to the modern robotics stack. Recent efforts are looking into using AirSim and similar simulations to that end. 6/8

7 lessons from AirSim: I ran the autonomous systems and robotics research effort at Microsoft for nearly a decade and here are our biggest learnings. A thread 🧵 Complete blog: sca.fo/AAeoC

I am happy to announce that TripletCLIP got accepted at #NeurIPS2024. Congrats to all the authors! Stay tuned for more details. 🎶 @SCAI_ASU @UMBC