Emerson Segura

@denys_khomyn @Polymarket Good point

@Polymarket Uber has no cars Airbnb has no buildings Meta has no metaverse This is the new economy

JUST IN: Meta announces they'll be shutting down the Metaverse, after pouring $80,000,000,000.00 into the project.

@chetan_ Lol, also this is why many lidar units still spin... self cleani g

someone needs to come out with windshield wipers for robot cameras

@Samuel_Gregson @kiwi_skeptical "Everything's Computer", buy Tesler

@kiwi_skeptical 99% of his schtick is regurgitating “deep thoughts” for idiots, because he’s an idiot.

This is now the lazy, cleverest little boy take regarding physics and it’s all over the internet. Why? Because it sounds deep to non experts, requires no knowledge or learning to say and plays to our anti-establishment moment.

@istoica05 amazing

Excited to share our latest work rethinking k-means for modern GPU architectures. While the algorithm is classic, scaling it requires aggressively targeting memory bandwidth bottlenecks. Flash-KMeans introduces a strictly IO-aware design that achieves exact results with up to a 30x speedup over cuML and 200x over FAISS, completing million-scale iterations in milliseconds!

@jonstephens85 @ErenChenAI @lukas_m_ziegler Fun times!

@ErenChenAI @lukas_m_ziegler @emerson Love it! But also, why did we not get one with both Lukas, you and I?! I guess we'll have to throw another party...

Proof of Life @jonstephens85 @lukas_m_ziegler @emerson

@Oksii33 @nastasiaKlimash horible

Russians have created another famine in Ukraine, yet the world is silent about it again 💔

@Scobleizer Good to hear, and sorry to have missed it

Tonight's SVVR San Francisco event was quite special. Why? After talking to a few entrepreneurs I realized these were the survivors of a brutal era in building a business in VR. Every product I saw was much more nuanced and well thought out than last year when I attended its events. It made me reenergized about my belief in the holodeck (AI driven VR and AR 3D environments). The next 18 months are gonna be nuts.

SONIC has been released for a while — here are some tips for whole-body teleoperation 🤖 1. WiFi quality directly determines teleoperation performance It's highly network-dependent. Use a dedicated router, avoid shared networks. PICO and the robot/host must be on the same network. High latency = dropped packets = unstable robot!! 2. Tracker orientation and the clothes you wear matters Ankle trackers must have their indicator lights facing up, unobstructed by clothing. Blocked sensors = unsteady footsteps. 3. Watch the posture before resuming teleop After pausing, unpausing immediately executes your current detected body pose. Return to a neutral standing position before resuming. Slight delay is normal — but too much and performance degrades fast!

@shiraeis @Yampeleg nice work!

Excited to share PostTrainBench: a benchmark for a question that feels increasingly important for AI R&D automation: can frontier CLI agents autonomously post-train LLMs? In PostTrainBench, we give an agent a base model, a target benchmark, one H100, and 10 hours, then let it handle the rest of the workflow, including finding data, writing training code, debugging, iterating, and submitting a final model. We provide no starter pipeline and no predefined strategy. Current agents do surprisingly decently, but there’s still a large gap to instruct models trained by humans. The best agent reaches 23.2% weighted average vs 51.1% for official instruct baselines (but this isn't exactly apples to apples, since official post-training usually exceeds the 10h / 1-GPU budget). At the same time, agents can already be surprisingly strong in targeted settings. GPT-5.1 Codex Max gets Gemma-3-4B to 89% on BFCL vs 67% for the official instruct model. We also found some reward hacking. Agents sometimes trained on test data, swapped in existing instruction-tuned checkpoints, or used unauthorized API keys for synthetic data generation. PostTrainBench is useful both for tracking progress in automated AI research and for studying the failure modes and sandboxing problems that come with it. collaborated with @full__rank @hrdkbhatnagar @AmyPrb @karinanguyen @MatthiasBethge @maksym_andr posttrainbench.com arxiv.org/abs/2603.08640 github.com/aisa-group/Pos… posttrainbench.thoughtfullab.com

@meggmcnulty @AstronomyVibes Its amazing, yet still limited for information transfer: great easy to follow video on it with founder of the theory to test it (John Bell): youtu.be/BFvJOZ51tmc?si…

@AstronomyVibes Quantum entanglement always blows my mind. The idea that particles remain instantaneously linked across any distance really challenges how we think about space, time, and information transfer. Nature’s ultimate spooky connection.

🚨 In the strange and fascinating world of quantum physics, there exists a phenomenon so mysterious that even Einstein called it “spooky action at a distance.” It’s known as quantum entanglement, and it links particles together in a way that defies space, time, and all known laws of communication. When two particles become entangled, they share a quantum state—meaning what happens to one instantly affects the other, no matter how far apart they are. You could separate them by galaxies, and still, a change in one would mirror in the other immediately. No signal travels between them, and no measurable delay occurs. It’s as if the universe itself bends to keep them connected. Scientists have confirmed this effect through countless experiments, proving that reality operates on levels far beyond what our senses can grasp. Quantum entanglement isn’t just a theoretical wonder; it’s now being used to shape future technologies—from unbreakable quantum encryption to faster-than-light communication research and revolutionary computing systems. What’s truly astonishing is the implication: everything in the universe might once have been entangled during the Big Bang, suggesting that distant corners of space could still be subtly linked through hidden quantum threads. Entanglement challenges our understanding of distance, time, and individuality. It reminds us that separation might only be an illusion—and that the universe, at its deepest level, moves as one.

@chris_j_paxton @cixliv Engineai people seem to run their servos about 20 Volts higher than unitree - so that would mean more power yet shorter battery life, if all things (number of battery cells per pack) were equal - a clever trick that some ev supercars also used to beat Tesla in performace specs.

@cixliv Sick, how are you liking this guy vs the g1?

We having fun with our new robot boi. He got gas in moves. Who's ready to see him fight?

robot dancer.. just needs a Kardashian bum...

this is what a Humanoid Robot getting an upgrade looks like in 2026 (updating firmware on @NVIDIAAI Jetson Orin to Jetpack 2.6) , let keep our robotfingers crossed that this goes well --more deets to come

@PhilHKG @NVIDIAAI LOL good catch :)

@emerson @NVIDIAAI I hope that’s a typo and you are going to Jetpack 6.2…

@kyliebytes best sourdough!

back in the best city in the world

@KyleVedder nice

@emerson it takes up to 15 mins to successfully complete the flagship tasks there’s nothing more fundamental about that number

my first PI project: we added memory! this is a step function capabilities unlock: 15 minute long multi-step tasks in novel environments, controlled by text prompting having run many of the evals, I legit think this is the GPT 2 moment for robotics

@svlevine Grilling Cheese in non trivial- half way to AGI ;))

We made a memory system for our models at PI. We call it Multi-Scale Embodied Memory (MEM). It provides both short-term and long-term memory to enable very long tasks. We tested it on cleaning a kitchen (and yes, washing dishes), making grilled cheese, and more.

@jiahuifu_carol Nice work!

🤖 How can we enable robots to solve complex, multi-step manipulation tasks without a single demonstration or any task-specific training? Introducing NovaPlan: a hierarchical framework for zero-shot long-horizon manipulation. It unifies closed-loop video language planning with grounded execution, allowing robots to “imagine” visual plans, execute robustly, and autonomously recover from failures. [1/5] 🔗 Project page: nova-plan.github.io #EmbodiedAI #Robotics #Manipulation #FoundationModels

@dr8_unix Nice

Day 1: turning the hf0/Lightyear basement into a robot teleop and eval farm.