Kunvar Thaman
9 followers · 4 following · 460 views
on the atlas — 71
- Shepard tone1 savers
- LessWrong39 savers
- Nathan Chen4 savers
- Many Benefits Of AGI Could Still Be Realized In A Pause1 savers
- Countering misuse of AI: September 2026 / Anthropic \ Anthropic14 savers
- An alignment assessment of recent cybersecurity incidents \ Anthropic5 savers
- Research acceleration: The view inside OpenAI | OpenAI13 savers
- Robot-use agents5 savers
- OpenAI Just Claimed a Huge Math Discovery. Some Academics Are Crying Foul | WIRED1 savers
- On Navier–Stokes | OpenAI11 savers
- san francisco - musings11 savers
- Discovery of a new OpenAI agent message board12 savers
- What will be scarce? - by Alex Imas - Ghosts of Electricity14 savers
- After Work — Asterisk Magazine11 savers
- Alfred Spector3 savers
- Potemkin village5 savers
- The Rise and Fall of Agent Civilizations15 savers
- Introducing Reinforced Planning (RP-1) — Pantheon2 savers
- Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models | NVIDIA Technical Blog4 savers
- Generalist - GEN-1.5: Embodied Foundation Models are One-Shot Learners7 savers
- Pacing model development in an era of cyber-critical capabilities | OpenAI3 savers
- [2603.16666] Fast-WAM: Do World Action Models Need Test-time Future Imagination?2 savers
- MiniMax H3: Open Weights & Prompting Techniques Complete Guide.1 savers
- H._C._Verma1 savers
- ACT-2 Preview: Generalizing Reliability | Sunday Robotics | The helpful robotics company4 savers
- Jacob Zietek3 savers
- LingBot-VLA 2.0 Foundation Model - Robbyant1 savers
- Antimatter Development Program – Casey Handmer's blog5 savers
- A Functional Taxonomy of World Models - Dr. Fei-Fei Li6 savers
- Map–territory relation4 savers
- [2312.10812] Learning to Act without Actions1 savers
- [2410.11758] Latent Action Pretraining from Videos1 savers
- [2206.11795] Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos1 savers
- Aleph1 savers
- [2602.12215] LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion1 savers
- Dedications3 savers
- Generalist - Blog7 savers
- Stop Simulating, Start Experiencing - by Paolo Di Prodi1 savers
- [2603.14482] V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning1 savers
- merlresearch/unic: UNIC: Learning Unified Multimodal Extrinsic Contact Estimation ·1 savers
- Thinking Machines Corporation6 savers
- Panthéon1 savers
- The effects of caffeine consumption do not decay with a ~5 hour half-life — LessWrong1 savers
- 'AI is getting better at cheating — and it doesn't look like cheating': Meet the Indian genius who got into ICML 2026 – Firstpost1 savers
- Nyx Iskandar1 savers
- Cato the Younger1 savers
- Julius Caesar1 savers
- Robot1 savers
- Robotic arm1 savers
- Steve Ballmer1 savers
- Khatri1 savers
- Uyghurs1 savers
- LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion1 savers
- Alignment is not solved but it increasingly looks solvable17 savers
- The First Fully General Computer Action Model | blog35 savers
- What's stopping you?29 savers
- The Intelligence Curse28 savers
- Introducing talkie: a 13B vintage language model from 193020 savers
- Bookshelf · Patrick Collison19 savers
- The best general advice on earth « the jsomers.net blog19 savers
- Building the heap: racking 30 petabytes of hard drives for pretraining | blog17 savers
- 2025 LLM Year in Review | karpathy16 savers
- EVERY SUPPLY CHAIN IN THE WORLD - by vincent huang16 savers
- AGI Trades11 savers
- Why ATMs didn’t kill bank teller jobs, but the iPhone did10 savers
- How to Buy Cheap Claude Tokens in China - by Zilan Qian8 savers
- on reading proust - Nabeel S. Qureshi5 savers
- Do Less. - by Cate Hall - Useful Fictions4 savers
- Ilya Sutskever – We're moving from the age of scaling to the age of research4 savers
- rsrch space4 savers
- Jack Clark on X: "Silent Sirens, Flashing For Us All" / X2 savers
highlights — 462
In 2021, he was awarded the Padma Shri, the fourth highest civilian award, by the Government of India for his contribution to Physics Education
H._C._VermaACT-2 does not start from scratch with each new capability.
ACT-2 Preview: Generalizing Reliability | Sunday Robotics | The helpful robotics companyWe also found a strong correlation between validation loss and success rate
ACT-2 Preview: Generalizing Reliability | Sunday Robotics | The helpful robotics companyI said "yeet" to an astronaut on the International Space Station for a dare
Jacob Zietekthe map is not the territory
Map–territory relationMistaking the map for the territory is a logical fallacy that occurs when someone confuses the semantics of a term with what it represents
Map–territory relationLAPO is the first method able to recover the structure of the true action space just from observed dynamics, even in challenging procedurally-generated environments. LAPO enables training latent-action policies that can be rapidly fine-tuned into expert-level policies, either offline using a small action-labeled dataset, or online with rewards
[2312.10812] Learning to Act without ActionsTraining only on human manipulation videos also shows positive transfer
[2410.11758] Latent Action Pretraining from VideosWe first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames
[2410.11758] Latent Action Pretraining from Videosto hard-exploration tasks that are impossible to learn from scratch via reinforcement learning.
[2206.11795] Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videosfrom which we can then train a general behavioral prior.
[2206.11795] Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videoswith a small amount of labeled data we can train an inverse dynamics model accurate enough to label a huge unlabeled source of online data
[2206.11795] Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videosagents learn to act by watching online unlabeled videos
[2206.11795] Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online VideosExperiments in simulation and the real world show LDA-1B outperforms prior methods (e.g., \pi_{0.5}) by up to 21\%, 48\%, and 23\% on contact-rich, dexterous, and long-horizon tasks, respectively. Notably, LDA-1B enables data-efficient fine-tuning, gaining 10\% by leveraging 30\% low-quality trajectories typically harmful and discarded.
[2602.12215] LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestiona pretrained video renderer can be used as the backbone for joint world-and-action prediction, suggesting a bridge between the renderer and the planner by letting one model imagine what will happen and what to do.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiA model that masters only rendering, or only planning, cannot do either.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiThis is, in many ways, the inverse of the renderer. Where a renderer takes actions as input and produces observations, a planner takes observations as input and produces actions, closing the perception-action loop. Vision-Language-Action models, model-based systems, and the new wave of World Action Models are all attempts at planners: systems that can decide what a robot should do in an unstructured world.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiThe third kind is a planner. A planner outputs actions. Given an observation and a goal, a planner answers the question of what the agent should do next.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiThe second kind is a simulator. A simulator outputs state: a geometrically, physically or dynamically faithful representation of the world that humans and computer programs can both compute on and interact with.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiThe first kind of world model is a renderer. A renderer outputs observations in the form of pixels meant for human eyes, and the quality that matters most is visual fidelity.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiObservations are an agent’s partial view of that reality. Actions are what the agent does in response.
A Functional Taxonomy of World Models - Dr. Fei-Fei LiState is the underlying reality of the world; complete in principle, but never directly visible to any agent inside it.
A Functional Taxonomy of World Models - Dr. Fei-Fei Liimitates expert actions but discards transferable dynamics knowledge embedded in heterogeneous embodied data.
[2602.12215] LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestioncarefully constructed data mixtures can lead to different pretrained model characteristics.
Generalist - BlogMaybe if your engineers have some extra time they will apply system identification mechanisms to the motor joints, to better estimate how the joints behave under stress
Stop Simulating, Start Experiencing - by Paolo Di ProdiThis is important so the model is in distribution during inference because it wasn’t exposed to latency during training—the model has never seen lag before. We mitigate latency through a variety of methods: colocating the GPUs and VMs in the same cloud region, using cumulative sequence length packing, tuning a low-latency VNC configuration, and writing custom Rust bindings for device input. The combination of these optimizations lets us achieve a round trip screen capture to action latency of 11ms
The First Fully General Computer Action Model | blogTo reduce the state space and use tokens more uniformly, we exponentially bin (Figure 10) the mouse movements
The First Fully General Computer Action Model | blogTo train a non-causal, generative model, we adopted a masked diffusion architecture
The First Fully General Computer Action Model | blogThinking Machines made some of the most powerful supercomputers of the time, and by 1993 the four fastest computers in the world were Connection Machines.
Thinking Machines Corporationis a monument in the 5th arrondissement of Paris, France.
PanthéonI assume this is 1 reason a lot of people got addicted to nicotine: despite the shorter half life you need a longer time off to reset the effectiveness. Caffiene after 2-7 days your body starts to undo its chemical adaptations and mostly washed out after a few weeks. For nicotine it's closer to a week to start 12 weeks to wash out an addiction.
The effects of caffeine consumption do not decay with a ~5 hour half-life — LessWrongVery little circulating caffeine is directly excreted
The effects of caffeine consumption do not decay with a ~5 hour half-life — LessWrongone of the biggest barriers to decarbonization is that, while we can find alternatives for power and heating, coal replacements in the steel-making process are difficult to come by
EVERY SUPPLY CHAIN IN THE WORLD - by vincent huangthis is why the usa does not have oil independence despite large reserves
EVERY SUPPLY CHAIN IN THE WORLD - by vincent huangDuring the civil war, he joined Pompey and tried to minimise the deaths of his fellow citizens. But after Pompey's defeat and his own cause's defeat by Caesar in Africa, he chose to take his own life rather than accept what he saw as Caesar's tyrannical pardon, turning himself into a martyr for and a symbol of the Republic.
Cato the YoungerA staunch advocate for liberty and the preservation of the Republic's principles, he dedicated himself to protecting the traditional Roman values he believed were in decline.
Cato the Youngerlet the die be cast".
Julius CaesarIn 60 BC, Caesar, Crassus, and Pompey formed the First Triumvirate, an informal political alliance that dominated Roman politics for several years.
Julius Caesarhigh-energy former CEO of Microsoft
Steve Ballmerif you’re cognition, factory, lovable, replit, you just watched the most successful version of your business model decide it could not run independently. that does two things to your strategic position.
cursor's warchest, xai's redemption - by Ethan Dingfirst, it lowers the bar for taking a sponsor. cognition does not need to wait until they’re bleeding 23 points of gross margin to entertain the conversation that cursor entertained. meta has llama, $50b a year of free cash flow, and zero credibility as an enterprise software company. amazon has bedrock and nothing on top of it that a developer would willingly use. microsoft has copilot and an openai exposure they’re trying to dilute. google has gemini and jules and nothing any enterprise has heard of. the call from at least one of these probably already came. it’s a question of timing.
cursor's warchest, xai's redemption - by Ethan DingAfghan Hindus and Sikhs descend from the members of the country's indigenous Khatri population who resisted the conversion from Buddhism to Islam between 9th and 13th centuries
KhatriAccording to Kiran Datar, they often married Tatar local women in Astrakhan and the children from these marriages were known as Agrijan.[75] As per Stephen Dale, the children born out of Indo-Turkic alliance were in sufficient number to form an Agrizhan suburb in the city.[
Khatri"Stephen Dale locates Khatris in Astrakhan, Russia during the late 17th century and, in the 1830s, Elphinstone, was informed that Khatris were still highly involved in northwest India's trade and that they maintained communities throughout Afghanistan and as far away as Astrakhan"
KhatriIn Bengal, Burdwan Raj (1657–1955) was a Khatri dynasty, which gained a high social position for Khatris in the region resulting in the increased migration of Khatris from Punjab to Bengal.[62] When Guru Tegh Bahadur visited Bengal in 1666, he was welcomed by the local Khatris, thereby supporting earlier waves of migration of Khatris to Bengal as well.[63]
KhatriAccording to a 17th-century legend, Khatris continued their military service until the time of Aurangzeb, when their mass death during the emperor's Deccan Campaign caused him to order their widows to be remarried.
Khatriafter the 10th century
UyghursAs a robot’s failure to reach a commanded goal is nonetheless a success for reaching the goal it actually reached, we can optimize the data distribution by replacing the originally commanded goals with the goals actually reached.
Reinforcement learning is supervised learning on optimized data – The Berkeley Artificial Intelligence Research BlogWhat makes RL challenging is that, unless you’re doing imitation learning, actually acquiring that “good data” is quite challenging.
Reinforcement learning is supervised learning on optimized data – The Berkeley Artificial Intelligence Research BlogThe main idea is to view RL as a joint optimization problem over the policy and experience: we simultaneously want to find both “good data” and a “good policy.” Intuitively, we expect that “good” data will (1) get high reward, (2) sufficiently explore the environment, and (3) be at least somewhat representative of our policy.
Reinforcement learning is supervised learning on optimized data – The Berkeley Artificial Intelligence Research Blog