CEO @ Cortexa_labs: AI Redteaming and Hardening Devtool | AI Researcher | AI Security | @fdotinc @_buildspace

Joined July 2024
Photos and videos
Couldn’t run this startup without @shreyasmav bro made this all possible.
1
3
810
And just like that we are moving to SF!! LFGGGGG. @fdotinc @hthieblot @shreyasmav
48
2
239
14,516
Charan retweeted
how does the brain build and track an internal state of the world from (possibly incomplete and noisy) visual observations? i believe visual state tracking will be the grand challenge for vision in the coming years, and i hope this benchmark can be a useful starting line. enjoy!
Can MLLMs actually track what's happening in a video? Introducing VSTAT 🎯, our new benchmark for visual state tracking. The tasks are simple: count cups, read typed words, count page flips. Humans solve them easily. MLLMs don't. vision-x-nyu.github.io/vstat
 đŸ§” [1/11]
23
45
431
75,813
Finished Canopy Online. Now we’re coming down in person đŸ«Ą. So excited to see my peeps in SF soon!!! @iliakakhadze @fdotinc @hthieblot @ruslanjabari
2
12
919
Project Sentry: Underway Analyses your repo, finds security vulnerabilities (data, inference, prompting, tooling) in your agents and if given the green light it fixes them and does a PR
1
2
310
What’s up with Cortexa Labs over the past 6 weeks👀
1
1
7
577
Recently spoke with the engineers at OpenAI’s engineering side. They are shutting down public fine-tuning access via API soon. If your startup is deeply dependent on proprietary models, it’s time to shift: - open-source models - self-hosting - hybrid inference stacks
1
2
263
This episode features an interview with Yao Shunyu @ShunyuYao14 , Research Scientist at Google DeepMind. Yao has held research scientist roles at both Anthropic and Google DeepMind, contributing to the development of key models including Claude 3.7, 4.5, and Gemini 3. Yao Shunyu is not your typical nerd. Every now and then, he’ll catch you off guard with a flash of irreverence. “None of the old guard are your relatives — so if you think someone’s being dumb, they’re just being dumb. Say it. No big deal.” (laughs) “Everyone’s a surfer now, but what really matters is the wave — not the person riding it.” “AI doesn’t actually require that much brainpower — I mean it genuinely doesn’t — most of this is work any undergrad could do. The most important quality in this industry is reliability: being meticulous, and taking responsibility for what you put out.” “You don’t need to worry too much about ruffling feathers with your opinions. As long as your views are internally consistent — not just taking random shots at people, but grounded in your own genuine understanding — there are objective standards for how you’re doing in this field. People will respect you for it.” Let us have a little fun with this one! 😄​​​​​​​​​​​​​​​​ youtu.be/ttkd0t5qTD4?si=0uFT

29
118
790
646,796
Such an inspiring moment. And the most inspiring person. I envision and dream to see Cortexa Labs making a difference like he is. I had goosebumps when I saw the documentary.

1
78
Charan retweeted
New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.
593
1,701
16,541
2,489,362
Charan retweeted
Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. đŸ§”
307
1,674
11,182
3,158,778
New pics released. Cortexa Labs @ Sunway Ilabs đŸ„‰
68
LaunchX Sunway ILabs 2025/2026 Result: 3rd Place and bunch of interested parties Went in solo and got the job done đŸ«Ą ONWARDS đŸ”„
1
5
246
Charan retweeted
For a decade, we've made models wider and deeper—but we've barely changed how layers *talk* to each other. Since ResNet's `x F(x)` in 2015, the depth residual has been the only highway for inter-layer communication. It's time to upgrade the staircase. đŸ§”
18
238
1,939
188,452
Me VS (Bugs & Time). Coffee up
44
Cortexa Labs is featured on Disruptr MY disruptr.com.my/cortexa-labs

1
4
216
Charan retweeted
Dr Fei-Fei-Li (@drfeifei ) explains why and how everyday household chores are so extremely difficult for Robots. "If you tell a robot to open the top drawer and watch out for the vase, this is actually a really hard task for robots." because the robot must ground language into the real world. Words like "top", "drawer", and "vase" are abstract. The system has to map them to 3D locations, objects, and relations in a noisy scene. This requires robust perception, object recognition, and spatial reasoning under uncertainty. The robot also lacks human commonsense. "Watch out" implies predicting consequences, estimating clearances, and understanding that vases are fragile. Encoding such priors, like how heavy a drawer is or how a vase might tip, is very complex and difficult without rich world knowledge. Learning the behavior from rewards is tough. The success signal is very sparse here, so naive exploration almost never stumbles on a full success sequence. This makes policy learning sample inefficient and brittle, especially when the environment changes between training and deployment. A sparse reward situation is when the agent only gets a success signal at the very end, and gets little or no feedback along the way. If a robot must open a drawer without hitting a vase, it might get reward only if the drawer ends up open and the vase is intact. Every partial try before that looks the same to the learner, reward equals 0. --- From "DSAI by Dr. Osbert Tay" YT channel
42
103
570
96,543
Charan retweeted
Aujourd'hui grosse discussion avec mes ingĂ©s (chez Argil) sur pourquoi Elon a virĂ© le LIDAR de ses voitures autonomes. Choix radical, moquĂ© pendant des annĂ©es, et comme d'hab il avait raison depuis le dĂ©but. Le LIDAR c'est un laser qui balaye l'environnement et crache un nuage de points 3D. Sur le papier tu obtiens la gĂ©omĂ©trie exacte du monde. Dans la vraie vie c'est une verrue technologique collĂ©e sur le toit parce qu'on sait pas faire mieux avec la vision seule. ProblĂšme numĂ©ro un : ça rajoute une modalitĂ© dans le training du modĂšle. Ton rĂ©seau doit apprendre Ă  fusionner vision lidar radar ultrasons. Chaque capteur en plus c'est une source de dĂ©saccord Ă  arbitrer, pas une source d'info supplĂ©mentaire. Sensor fusion artisanale = dette technique permanente. ProblĂšme numĂ©ro deux, la bitter lesson de Rich Sutton : scaler le compute sur une seule modalitĂ© bat systĂ©matiquement les architectures bricolĂ©es Ă  la main. Tesla a dropĂ© le radar, puis les ultrasons, est passĂ© full end-to-end vision. Leur courbe sur les edge cases s'est accĂ©lĂ©rĂ©e APRÈS, pas avant. Waymo fait l'inverse et reste stuck en ops gĂ©ofencĂ©e. ProblĂšme numĂ©ro trois, le plus fondamental : le LIDAR voit la gĂ©omĂ©trie, pas la sĂ©mantique. Il sait qu'il y a un truc, pas ce que c'est ni ce que ça va faire. Les derniers 9 de fiabilitĂ© sont des problĂšmes de cognition, pas de perception brute. Un capteur de plus rĂ©sout rien, il ajoute du bruit. SĂ©bastien Loeb balance une 208 T16 Ă  180 dans un chemin boueux corse sous la pluie avec zĂ©ro LIDAR. Deux yeux, un cerveau. L'Ă©volution a donnĂ© des yeux aux prĂ©dateurs pendant 500 millions d'annĂ©es, pas des lasers. Il y a une raison. Le LIDAR c'est l'Ă©quivalent du marxisme appliquĂ© Ă  l'Ă©conomie. Une solution planifiĂ©e, centralisĂ©e, qui prĂ©tend modĂ©liser explicitement ce qui doit Ă©merger d'un systĂšme distribuĂ© et adaptatif. Tu remplaces l'intelligence par de la mesure, la comprĂ©hension par de la donnĂ©e, l'Ă©mergence par le contrĂŽle. Ça rassure les ingĂ©nieurs qui veulent tout spĂ©cifier en amont, exactement comme la planif rassurait les Ă©conomistes soviĂ©tiques. Et ça Ă©choue pour les mĂȘmes raisons : la rĂ©alitĂ© est trop riche pour ĂȘtre capturĂ©e par un capteur, comme elle est trop riche pour ĂȘtre capturĂ©e par un plan quinquennal. La vraie intelligence, celle de Hayek comme celle de Tesla, c'est de faire confiance Ă  un systĂšme qui apprend de l'expĂ©rience plutĂŽt que de tout prĂ©-encoder. L'Ă©lĂ©gance d'une solution c'est son rapport signal sur complexitĂ©. Le LIDAR explose le dĂ©nominateur. DĂ©fendre le LIDAR en 2026 c'est prĂ©fĂ©rer empiler des hacks plutĂŽt que rĂ©soudre le vrai problĂšme. C'est de la feignasserie intellectuelle maquillĂ©e en rigueur d'ingĂ©nieur. Les mĂȘmes gens qui dĂ©fendaient les systĂšmes experts en 2012 contre le deep learning. Ils finiront pareil. Never bet against end-to-end. Never bet against la simplicitĂ©. Never bet against Elon.
1,107
2,164
18,380
29,275,859