Viraj

@isabelle_zhou @OpenAI @isabelle_zhou Do you think scaling the data or the model parameters can lead to generalization? Especially with the current transformer architecture?

I’m hiring for my research team @OpenAI 🪄 Data will pave the way for AGI. This is a foundational TPM role to shape frontier AI models with real world data, RL environments, and data acquisition. Looking for: - Entrepreneurial, gritty, high horsepower - Highly technical and deeply curious about AI research - Excited to lead relationships with CEOs and data partners from day 1

@boardyai peace

founders outside of the SF what's your edge ?

@adcock_brett Humanoids demand generalization from day one. How are you thinking models specially for humanoids?

@adcock_brett @adcock_brett how are you thinking the foundation models for humanoids? I think humanoid is the toughest problem (humanoid cannot run on pattern-matching against world wodels as real-world comes with infinite uncertanities).

New office is officially open at Hark! We have 2 floors in a 110,00 sqft building; hiring for 47 open roles, all in San Jose, CA: - Design - Embedded Software - Product Engineering - AI Foundation Models - Hardware Engineering

@kseniam0s Unicorn next

Give me a startup name SO COOL it sounds like the next unicorn 🦄

There is no difference between LLMs and World Models when it comes to the end game. Both are identical in mechanics. Both take input, assemble it in relative positions, and run brute-force pattern matching. The only difference is the loss function. World models are the same kind of trap as LLMs. I have one question for all the labs and researchers working on world models for decoding physical reality: how do you think a world model can generalize? If the answer is throw more data and more compute, you have already lost the argument. A human child does it on twenty watts, from one example. Maybe its time to reconsider the architecture? #machinelearning #physics #neuroscience #ai #generalintelligence

@vkhosla This is great, congrats @vkhosla . Curious how you are thinking about generalization here. A toddler generalizes from one experience and runs on 20 watts. World models still need tons of data and compute, same as LLMs. Would love your thoughts.

I am betting on this winning the world model race with human like intuition emerging in this model. Uncanny the range of things it can do. Can it win a formula 1 race soon?

Position paper - github.com/ReinforceAI/un…

We keep building bigger models, feeding them more data, and burning more compute, and yet not a single AI system can do what a toddler does effortlessly - generalize from first experience across everything it sees, hears, and touches. Why? Because every current AI architecture, LLMs, world models, reinforcement learning, does the same thing at its core. It starts with data and tries to decode the patterns of physical reality. It reconstructs a representation of the world from sensory input and then predicts from that reconstruction. The human brain never reconstructs. It only predicts. Right now, as you read this, your brain is simultaneously predicting the next word on this screen, the feel of the chair beneath you, and the ambient sound around you. It doesn't take in raw data and assemble a picture. It generates a continuous prediction of reality and your senses only carry the difference between what it expected and what actually happened. That difference is the only signal the brain needs. No labels. No rewards. Just surprise. This is the Free Energy Principle from Karl Friston, one of the most cited neuroscientists alive. It's been the leading computational theory of the brain for over a decade. And it points to a fundamentally different path toward generalized intelligence. One that doesn't start with data but with the physics of reality itself. We arrived at the same place from the opposite direction. We started from first-principles physics, asking how a single system could perceive light, sound, and motion through one unified architecture. The answer we reached independently was the same: encode the physics of reality, predict continuously, and let the prediction error be the only teacher. When we found Friston's framework, our physics and his neuroscience had already converged. Our position paper introduces what came out of that convergence. A unified perceptual substrate. An internal world that continuously runs forward on the physics of reality, predicting what light, sound, and motion should look like at every moment. How a scene should render. How a sound should propagate. How a body should move under gravity. These predictions run all the time, just like yours do. When sensors deliver something different, that prediction error drives perception, learning, and action through a single architecture. A unified medium. Every sense. Self-correcting against reality. We're not decoding the world from data. We're encoding it from physics and letting generalized intelligence emerge from surprise. Our main paper is coming soon. A humanoid that identifies its own body, stands, self-stabilizes, and performs basic actions, with no pattern matching, no training data, and no reward signal. Just predictive coding against physical reality. Paper linked in comments. #ai #agi #neuroscience #robotics #physics #predictivecoding

Dear @narendramodi @AshwiniVaishnaw @PMOIndia, @OfficialINDIAai I run a deep-tech AI research lab building sovereign AI for India. We work on novel compression, attention mechanisms, and better inference. Our work has shipped in production on 3 frontier models. Our techniques compete with the world's most elite labs. Give us the chance to train India's sovereign foundational model. We are ready. reinforceai.com

Drive innovation that powers India’s AI future. Join the IndiaAI Mission and lead transformative application development and innovation initiatives. More at: dic.gov.in/jobs/deputy-ge… #IndiaAI #Hiring #Innovation #AIIndia #TechLeadership #DigitalIndia @AshwiniVaishnaw @jitinprasada @PIB_India @SecretaryMEITY @kavitabha @GoI_MeitY @_DigitalIndia

Dear @narendramodi @AshwiniVaishnaw @PMOIndia, I run a deep-tech AI research lab building sovereign AI for India. We work on novel compression, attention mechanisms, and better inference. Our work has shipped in production on 3 frontier models. Our techniques compete with the world's most elite labs. We have reached out to IndiaAI but no response so far. Give us the chance to train India's sovereign foundational model. We are ready. reinforceai.com

@OfficialINDIAai I run a deep-tech AI research lab building sovereign AI for India. We work on novel compression, attention mechanisms, and better inference. Our work has shipped in production on 3 frontier models. We can compete with elite labs globally and want to build foundational models for India. Writing to inquire about grant pathways under the IndiaAI Mission. Website: reinforceai.com Thanks, Viraj viraj@reinforceai.com

Building a presence in the European market requires a fundamental shift in strategy: you cannot rely solely on a polished pitch deck. Drawing from her firsthand experience in the program, Niharika from @volaralta highlights a crucial insight for the incoming founders of Cohort II—patience and relationship-building are everything. Establishing a footprint in this ecosystem demands showing up consistently and earning trust from the ground up, proving that long-term commitment matters far more than a cold email. Write your own accelerator story. #TheAcceleratorStories #IndiaAI #CohortII #GlobalScaling #MarketEntry #FounderTips @AshwiniVaishnaw @jitinprasada @PIB_India @SecretaryMEITY @kavitabha @GoI_MeitY @_DigitalIndia @mygovindia @HECParis

@OfficialINDIAai I run a deep-tech AI research lab building sovereign AI for India. We work on novel compression, attention mechanisms, and better inference. Our work has shipped in production on 3 frontier models. We can compete with elite labs globally and want to build foundational models for India. Writing to inquire about grant pathways under the IndiaAI Mission. Website: reinforceai.com Thanks, Viraj viraj@reinforceai.com

IndiaAI Mission has selected 10 Indian AI startups for Cohort II of its Global Acceleration Programme, in partnership with Station F and HEC Paris. The opportunity brings three weeks of intensive online preparation, followed by three months living and building at the world's largest startup campus in Paris for the startups; in addition to giving access to top European mentors, investors, and markets. Read more 🔗 pib.gov.in/PressReleaseIf… #IndiaAI #GlobalAcceleration #StationF #HECParis #StartupIndia #MeitY @AshwiniVaishnaw @jitinprasada @PIB_India @SecretaryMEITY @kavitabha @GoI_MeitY @_DigitalIndia @mygovindia @HECParis

.@OfficialINDIAai @GoI_MeitY We are a core research lab working at the intersection of AI and Physics - novel compression, attention mechanisms, inference efficiency. Shipped on 3 production models. We can compete with elite labs globally. Reached out multiple times about research grants. No response. How does a serious deep-tech lab in India actually get heard? Website: reinforceai.com

@OfficialINDIAai @OfficialINDIAai We are a core research lab at the intersection of AI and Physics. We are trying to reach out and connect regarding the research grant. We have the capabilities to compete with the world's elite labs. Here is our website - reinforceai.com

How can AI transform healthcare in India? 🏥🤖 At Governance Summit 2026, hear from Suman Katragadda, Founder & CEO, HEAPS.ai; Sujoy Kar, Chief Medical Information Officer & Vice President, Apollo Hospitals Group; Deepshikha Batheja, Principal Research Scientist, ISB; and Karthik Raman, Professor, IIT Madras, moderated by Chandan Jain, Senior Evaluation Specialist, 3ie, as they discuss building accessible, affordable, and accurate healthcare systems powered by AI. 📍 ISB Mohali Campus 🗓️ May 23, 2026 Register here: isb.edu/events/governa… #GovernanceSummit2026 #InclusiveAI #HealthcareInnovation #DigitalHealth #IndiaAI #ViksitBharat @AshwiniVaishnaw @jitinprasada @SecretaryMEITY @PIB_India @kavitabha @GoI_MeitY @_DigitalIndia @BIPP_ISB

@lucychen_ Critical thinking matters. Code was never the issue. How a researcher thinks is all we need and no doubt Andrej is really good at how to think.

do you think he had to do the coding interview or does he get a pass?

This competition is rigged. My PR beats all the solutions on unlimited compute with our unified attention mechanism and on record submission it landed on middle of the table but they never accept or denied the PRs. Till date, the PRs are open despite of approval from community review. Here are the PRs - github.com/openai/paramet… github.com/openai/paramet… Folks running this competition at @OpenAI cannot be trusted.

parameter golf was a blast. 2,000+ submissions. 1,000+ verified github accounts. ideas ranging from quantization and depth recurrence to TTT LoRA, SSMs, H-nets, JEPA, and more. autoresearch made iteration dramatically faster — and led to emergent bulletin boards, issue threads, unofficial leaderboards, and agent-built writeups that helped everyone learn from everyone else. it felt like a glimpse of where interaction with AI is headed: humans setting taste and direction, agents helping explore, coordinate, and share what works. our goal was simple: make ml research accessible to anyone, anywhere. it was amazing to see that happen. full recap: openai.com/index/what-par… future events: jobs.ashbyhq.com/openai/form/op…

@thejeremyredman @askOkara Hey this information is really old. Like 1.5 years ago. But yeah OpenAI is def the #1 competitor, no doubt.

Hey! Check this out. I ran a free report for Reinforceai at LeadQuest.ai. Dropped in domain, got back ICP, competitors landscape, outreach emails, and verified leads. All from your URL. Takes 2 minutes. heres your report! app.leadquest.ai/reports/reinfo… Name Your Price. Seriously.

drop your website and i'll ask our ai cmo how to grow it

Here are the resources - GitHub: github.com/ReinforceAI/sp… Model: huggingface.co/Reinforce-ai/Q… Install on Mac: brew install reinforceai/spiral/spiral

Today we released our first model in Spiral. Calibration-free compression, competitive with the strongest community quantizations - within 1.8pp of unsloth Q4_K_M on HumanEval. No calibration data. No fine-tuning. Production-stable on Apple Silicon (Metal) and NVIDIA (CUDA). We'll keep pushing the limits of what's possible and packing intelligence in the most optimal way. Open source is cool I guess... Namaste 🙏 🇮🇳 #opensource #machinelearning #physics #ai #transformers #llm #quantization