iterator retweeted
Possibly one of the best RL notes I have ever found, going through them one by one really clarifies a lot of doubts one gets when studying RL.
iterator retweeted
On a research thread to (re)derive all kinds of optimizers based on what constraint they pose to the loss function. An interesting one is to enforce the weight update to have Frobenius = spectral norm = 1. Turns out that this equals to taking the top pair of singular vectors of the gradient alone as the weight update.
The central idea is that frobenius = spectral = 1 enforces the weight update to have a rank of 1, so it can be parametrized by the outer products of two vectors, pq^T. Then after a few pythegora's theorem we can get the top pair of singular vectors u_1 v_1^T as the solution.
This checks out as the special case of the NorMuon optimizer which chooses the top-k pair of singular vectors according to a total norm budget, where k=1.
Ongoing work to derive more optimizers. Blog: riddlehe.github.io/blog/full…
iterator retweeted
Purdue University's IDEAS Lab built an autonomous drone that tracks runners through dense forests completely without GPS or cloud networks.
By running 3D LiDAR mapping and obstacle-avoidance AI entirely on-chip, it can weave under tree canopies and predict target movement through blind spots in real time.
Massive implications for next-gen search & rescue and edge AI.
iterator retweeted
With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier.
Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data.
Website: leggedrobotics.github.io/ACP…
Promo Video 👇
iterator retweeted
GOAT🐐
With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier.
Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data.
Website: leggedrobotics.github.io/ACP…
Promo Video 👇
iterator retweeted
"The Linear Algebra Behind Google" is a short paper that explains the mathematics behind Google's original PageRank algorithm.
The paper covers eigenvalues and eigenvectors, stochastic matrices, Markov chains, and iterative methods, using numerical examples and graphs to illustrate how PageRank works.
Beyond the algorithm itself and the original paper by Larry Page and Sergey Brin, this is an interesting read for anyone curious about how linear algebra applies to one of the most influential algorithms in the history of the web.
sites.cs.ucsb.edu/~gilbert/c…
iterator retweeted
DAMN, this paper is good...
❗ The last few seconds are worth watching.
'Blind Precise Assembly'.
A paper demonstrating that robot joint encoders provide sufficient proprioceptive feedback for precise assembly tasks.
Like what?
Like peg insertion, gear meshing, and nut threading.
But here is the beauty of it: with no cameras, tactile sensors, or object pose knowledge required!!
Policies trained entirely in simulation using heavy randomization of geometries, poses, and contacts.
Even more impressive 👉 it achieves near-perfect success in sim and zero-shot real-world transfer for peg insertion at 73% success across diverse 3D-printed parts with tight clearances.
Blind policies generalize better than matched visual policies to out-of-distribution parts while being slower on in-distribution ones; the work also releases mjfit, an open-source tool for fast, exact convex simulation of tight-tolerance assemblies in MuJoCo.
This is MIT CSAIL researcher Antonia Bronars (@BronarsToni) final project of her PhD w/ co-first author Younghyo Park (@younghyo_park).
For more details, check out their project website and their paper!
📌 website: younghyopark.me/blind-propri…
arXiv: arxiv.org/abs/2610.23849
mjfit: github.com/younghyopark/mjfi…
——
Weekly robotics and AI insights.
Subscribe free: 22astronauts.com
iterator retweeted
Information is Energy — A physically based concept of information: link.amazon/B0dR0U999
iterator retweeted
WE DID IT! We can now pretrain transformers without backprop and even beat backprop+Adam on validation loss with insane speeds.
Now the big question is scale. It’s Ironic, we have reached a breakthrough In AI capability due to inspiration from a safety blog
I think the last month might have been the biggest month in (zeroth-order) optimization history.
In the 10 days since we released the Noise speedrun, the time to pretrain a transformer went from 1.76 hours to 1.6 seconds, a 4,000x (!!!) speedup. Dust itself was up to 10,000x more compute-efficient than the previous state-of-the-art.
Not only can we pretrain a transformer without backprop, it's much more efficient than anyone would've expected. Search fking works!
Now, scaling is the biggest open question. Some of the tricks that brought the time down to 1.6 seconds will simply not scale. To discard those, we'll be scaling the dataset every time it gets overfit, starting with 30M tokens instead of 10M.
Huge thanks to @jayden_teoh_, @spicey_lemonade, @RobertTLange, @EmanueleRodola, @ChunyuanDeng, @LuochengHuman, @Alaxarr, @pavleculafic, @IsaacYong13, Euraba and areu01or00 (GitHub), for the contributions. And @cs_serdar for gracefully maintaining the speedrun and @bishmdl76 for helping out.
Expect a lot more from us over the coming weeks.
iterator retweeted
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: supermimic.github.io
Demo: supermimic.github.io/demo/
iterator retweeted
The rate in which everyone is adopting and implementing this reinforces to me how broken the conference system is.
Wait a few months for 3 random people to hopefully give a decent read, or... Let the community judge?
Community validation is 100x more important and impactful
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: supermimic.github.io
Demo: supermimic.github.io/demo/
iterator retweeted
If you feel the same, check out our mathematical foundations sequence.
Especially designed to propel adult students to the upper echelons of math as efficiently as possible.
mathacademy.com/courses#math…
iterator retweeted
Excited to introduce reactive optimizers: We let an agent reactively adapt the optimizer code during the training run! The agent acts like a team of scientists constantly choosing the best algorithm, learning rate, momentum, etc, based on its own analysis of step-to-step training dynamics! The results are really exciting. Besides out-performing standard "best-practice" recipes across everything we threw at it, the agent gains a few "superpowers" that no human could have designed into the optimizer... (1/n)
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
iterator retweeted
Hot take: Most autoresearch today is just glorified overfitting.
Current autoresearch pipelines are essentially outer-loop searches for recipes that work in ONE specific setting. Try transferring a modded-nanogpt recipe to a different language or dataset. How much of that "optimal" recipe still holds up?
We need to move beyond static recipes toward inner-loop autoresearch: optimizers, architectures, hyperparameters, etc., continuously adapting DURING training based on the training trajectory itself.
Check out OPAL that enables this! Cool work with @SimulatedAnneal, Jia-Hao Liang, and @vincesitzmann.
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
iterator retweeted
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
iterator retweeted
YES! i believe AGI can check every data every neuron every hyper-param in optimier per every step, and this is what the next 10x we can get from pretraining!
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-opt…
💻 Try it on your task: github.com/xingjian-bai/opal
🧵
iterator retweeted
🚨 ضربة معلم جديدة من DeepSeek: فتح المصدر بالكامل لإطار عمل DeepInfer الذي تشغل به نماذجها العملاقة مجاناً للجميع!
الشركة الصينية التي زعزعت عرش وادي السيليكون تعود مجدداً بمنهج "الشفافية المطلقة".. وقررت جعل البنية التحتية والاستدلال (Inference) لنموذج DeepSeek-V4.1-Flash مفتوحة المصدر بالكامل لتقليل تكلفة وسرعة واستهلاك طاقة الذكاء الاصطناعي على مستوى العالم!
📌 أبرز ما كشفت عنه DeepSeek:
إطار عمل DeepInfer مفتوح المصدر: المحرك الداخلي الفائق الذي يوفر إنتاجية عالية وزمن استجابة منخفض جداً (Low latency) مع دعم Distributed inference وSpeculative decoding!
مكتبة deepseek-recipe: فتح المصدر لمكونات تشفير وفك تشفير التوكنز (Encoding & Decoding) لضمان تطابق أداء خوادمك الخاصة مع API الرسمي لـ DeepSeek تماماً!
المستقبل هنا: كل الأكواد البرمجية للتحسينات والنماذج القادمة سيتم رفعها مباشرة على نفس المستودع!
نماذج مفتوحة.. أنوية مفتوحة.. واستدلال مفتوح بالكامل. DeepSeek تستمر في خفض تكاليف الذكاء الاصطناعي وجعله متاحاً للجميع!
احفظ التغريدة وشاركها مع أصدقائك المطورين! 🧵👇
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
