Loading video...

Video Failed to Load

Go Home

Decentralized RL as fast as centralized RL for LLMs. Bittensor SN81 grail has shattered the bandwidth barrier with PULSE (Patch Updates via Lossless Sparse Encoding). "By identifying the 99% weight sparsity inherent in Adam-bounded updates, we've achieved a 100x reduction in weight synchronization - dropping 14GB transfers to just...

13,074 views • 3 months ago •via X (Twitter)

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Covenant Labs just did a 90-minute AMA breaking down their 3 Bittensor subnets. templar. basilica. grail. Pre-training, compute, and post-training under one roof. Most people missed it. Here's everything they said. Covenant is building what they call the "end to end intelligence continuum." Three subnets. Three layers of the AI stack. All permissionless. Templar (SN3) handles decentralized pre-training. Basilica (SN39) handles compute. Grail (SN81) handles RL post-training. Sam Dare, the lead, put it bluntly. Decentralized training is "humanity's last dance." Not about beating OpenAI head to head. About creating optionality. About making it cheap enough for anyone to train models. The gap between academia and frontier labs is growing exponentially. Researchers can't afford to experiment. The actual training run costs 5% of the reported budget. The other 95% is experimentation. If Covenant cracks cheap training, that entire surface area opens up. On Templar specifically: • Hit 39% emission on Bittensor. Highest since Apex was the only subnet on the network • Covenant-72B trained permissionlessly with 70+ contributors on commodity internet • 1.1 trillion tokens processed. No centralized data center • Performance competitive with LLaMA-2-70B On Grail, something flew under the radar. They built Pulse. A weight synchronization method that compresses model updates by 100x. • In RL post-training, only ~1% of weights update per step • Pulse exploits that sparsity. Lossless compression • Prime Intellect's comparable system took 14 minutes to sync a 30B model • Pulse makes decentralized RL training actually feasible at scale • Already used by Cursor The lead researcher on Grail said they've trained on math, code, and GPU kernels. Got 40-60% improvement on benchmarks. Working toward agentic training with 100K+ token context and 30B+ parameter models. On Basilica, the compute subnet: The team was blunt. Just reselling GPU hours is a 5-10% margin game. Traditional compute providers already do that. Their play is value-added services. • "GPU as code." No dashboard. No UI. Agents interact via SDK • Custom scheduler that places workloads across heterogeneous hardware • Verification checks for GPU, CPU, bandwidth, memory, storage, and OS security • Partnerships with providers like Mass Compute for 10-20% below market pricing • Miners compete on useful infrastructure, not just GPU hours Sam then went on a rant about the miner burn debate. His take: Bittensor had to grow up. dTAO introduced investors. The old "miners are God" philosophy doesn't hold. • Subnet owners have a duty to protect token value • Miners are a resource optimization exercise, not a cost reduction exercise • 100% miner emissions on compute subnets = immediate sell pressure • The 41% miner allocation is arbitrary. Different business models need different splits • Fish (who started burns) agreed. Burns usually mean the validation isn't mature enough The bigger point. You can't police burns. Subnets just send to their own keys instead of the burn address. Subnet 28 does exactly that. Sam's position: judge subnets on outcomes, not process. Const has changed the protocol 9-10 times in 2 years. That iteration speed is Bittensor's actual moat. The whole Covenant thesis is playing out in real time. TAO is up 100%+ in a month. Jensen Huang name-dropped the network. Grayscale has an ETF filing. But the real story is three subnets quietly building every layer of decentralized AI.

Jesus Martinez

26,642 views • 3 months ago

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

125,146 views • 1 year ago

I see lots of people launching projects where their agents weigh and interact with information they are provided with via X through their X accounts. I've been experimenting with agent function and for long before it was popular on X, so, it had been made very clear to me long before that such automations and their functions are nothing but novel instances of surface level "interaction", yet it is hard to even call it that. The weighing of these instances is something that I find if stressed and iterated upon, could produce a more capable agentic framework. I've outlined how the depicted error and rewarding loop can be improved to make a better agent. I intend to move agentic function beyond theatrical loops of prompt and response by interrogating the optimization substrate itself. Rather than treating reward as a convenient scalar pat on the head, I frame agent behavior as a constrained variational problem over latent state transitions, where policy updates approximate inference under structured uncertainty. My approach draws less from corporate folklore and more from specific technical inflection points such as the reanalysis framework in Reanalyse: A Simple Way to Improve Sample Efficiency in RL by Julian Schrittwieser and colleagues, the planning architecture of Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm by David Silver et al., and the value decomposition insights in QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning by Tabish Rashid. What interests me is not their surface performance but the structural concessions they make to tractability. By embedding differentiable world models and amortized belief updates into the agentic loop, I treat interaction as recursive posterior refinement across partially observable manifolds, not as a parade of conditioned tokens but as an evolving distribution over trajectories. At a higher altitude, I see agentic improvement as a question of mathematical hygiene. How does one approximate optimal control in a non convex landscape without dissolving into instability under recursive self modification. The error reward cycle, so often romanticized, is in fact a delicate dynamical system whose gradient flows inhabit curved statistical manifolds. Insights from regret bounds in bandit theory, contraction mappings in dynamic programming, and spectral analysis of iterative operators suggest that coherence is less about clever prompting and more about fixed point behavior under perturbation. I am particularly preoccupied with the spectral radius of update operators, the existence and uniqueness of equilibria in combinatorial policy spaces, and the computational hardness that shadows long horizon planning. If the agent is to be more than an improviser with memory, its loop must satisfy constraints that are as much algebraic as empirical, grounded in proofs of convergence rather than optimism about scale. -I’m not the greatest performer, so please bear with my narration in the video!

Caro

246,343 views • 4 months ago

Bio inspired Hebbian probabilistic network learns in less than 5 minutes from a super sparse single reward per episode! also has imitation learning (manual control) system has 3 parallel competing networks which get sensory input from a 360 vision (27-direction sensory neuron array) link to code in comment each sub-network is responsible for a single motor action: forward, left and right. at each step whichever section has most neurons firing wins neurons fire probabilistically and mark themselves with a time-decay tag which happens when a neuron fires and diminishes with time. you can see this " tag countdown" on each neuron when a reward is attained(eating the cheese) eligible connections gets strengthened I included 2 runs in the video first was 15 minutes in real time and second was 5 minutes. red plot is the rolling average of last 10 time to cheese. it is really not possible for agent to achieve full control due to probabilistic neural firing. that is why it has to learn while jittering all over the place, which in itself is interesting in manual mode you can guide the cheese by stimulating its motor control networks ( still probabilistically ) and the rewards will still work ✅ Biologically Plausible Features: Stochastic firing (neurons in the brain fire probabilistically) Reward-based learning (dopamine-like neuromodulation) Hebbian plasticity (well-established biological mechanism) Eligibility traces (biological neurons have temporal credit assignment) Sparse sensory encoding (similar to place cells, grid cells) Competitive action selection (basal ganglia architecture) No backpropagation (which is biologically implausible) ❌ Missing Biological Features: No recurrent connections (real brains have extensive feedback loops) No inhibitory neurons (GABAergic neurons are ~20% of cortex) No spike timing (simplified from true spiking dynamics) Uniform layer structure (biological networks are more heterogeneous) Simple weight updates (real synaptic plasticity is more complex)

echo.hive

33,638 views • 8 months ago

🇮🇷| Iran now competes with the US & China on advanced Nano-Insulation technology, used currently by NASA Iranian scientists have successfully localized a state-of-the-art nano-insulation material — the very same class of technology used by NASA to protect spacecraft and astronaut equipment against the most extreme environmental conditions. When the pores of a material are engineered at the nanoscale, they effectively block any transfer of energy through the structure. Building on this principle, an Iranian knowledge-based company has developed and commercialized 3 types of advanced insulation products designed for the industrial, energy, and construction sectors. • Aerogel Blanket: Tailored for use in refineries, petrochemical complexes, and power plants, this flexible blanket withstands temperatures ranging from –200 °C to +650 °C in a single system. • Aerogel Granules: Recognized as the lightest commercial solid in the world, these granules are added to gypsum, cement, and resin formulations to enhance thermal resistance while reducing weight. • Aerogel Powder (Sprayable Form): Derived from the same nanomaterial, this powder serves as the base for a sprayable coating that can easily cover complex geometries and metallic surfaces. Once applied, it creates a powerful barrier against heat, moisture, and even sound. Despite being 99% air, the aerogel blocks energy loss 6x more effectively than conventional insulators.

Arya - آریا

76,751 views • 9 months ago

Where are they - PayPal ? Throwback video; as to why PayPal would likely not come to Maldives anytime soon, 2 years ago, Sep 2023, explaining MDP Secretariat Government of Maldives Ministry of Economic Development and Trade Fayyaz Ismail efforts. Fast forward 1 year 9 months to June 2025, for PNC Secretariat to come up with the same conclusion. What an utter waste of time. [ MohamedShahzan In the interim, before the parliament People's Majlis election, the PNC government came up with a bizarre strategy; discussions with the US government to "enforce" PayPal services in Maldives. Of course, it was just candy for the younger voters. My response to this bizarre announcement. [ And then, right before election the Dr Mohamed Muizzu government minister Ibrahim Khaleel lied by announcing that PayPal was available shortly before parliament election. [ aaidh 🇵🇸 And then later, Mohamed Saeed stating a proposal has been sent to PayPal in Sep 2024. [ Adhadhu article noting 1 year since PayPal proposal was sent with nothing to show for it. [ aaidh 🇵🇸 The gist of this situation has not changed since 2 years ago. PayPal is a public company - a multi-national operating in over 200 countries - a multi-billion $ for profit business enterprise. From PayPal's perspective, they have many countries whose economies are over a 100 times bigger - queued up for their business services in the coming years. So why would they choose Maldives? In fact, the potential ~2 year roadmap (video) itself was a moonshot, as it would require Maldives and some neighbouring countries working together as a single source market to ensure there is enough "weight" for PayPal commerce. Inescapable conclusion; Dr Mohamed Muizzu government ministers has been knowingly stringing along the public and particularly the youth, with their repeated lies regarding PayPal service availability. Women in Tech Maldives sefm.mv Business Center Corporation MDP Youth Wing MDP Womens Wing

Riyaz Mansoor

30,434 views • 9 months ago

VAXEE XEv2 4K Wireless Mouse Introduction Video and Sales Information VAXEE launches the XEv2 4K Wireless Mouse, continuing the first-generation design while making adjustments to its weight, structure, and operational details. We’ve prepared a detailed introduction video as well as a written summary for your reference. Key Updates: · Weight: Reduced from 76g to 59g, with adjusted weight distribution to ease lift-off burden. · Sensor & Wireless Transmission: Equipped with the 3950 sensor; in Competitive Mode 4K, latency is 5% lower than the first generation. · Button Design: Uses a mid-frame structure to reduce wobble and accidental clicks; supports three levels of click response time: near 0ms, 2ms, and 4ms. · Function Button: Removed to achieve weight reduction and better balance. · Adjustment Method: Supports web-based driver and retains bottom buttons, allowing DPI, polling rate, and response time adjustments even without internet access. Design Features: · Flat and rounded back shape provides full support for the palm. · Mouse movement and lift are guided by the thenar and hypothenar areas of the hand, reducing finger strain. · Ideal for users who prefer extending their hand forward with the palm resting against the mouse’s back. Color Options & Sales Information: To thank our first customers, those who purchase within the first three days before new color options are released and do not request a return will receive a complimentary mouse case. The shipping date for the mouse case will be announced separately on our social media. · May: Obsidian Black and Pink Sales Start: May 7, 2025, at 6:00 PM (Customers who make a purchase between May 7 and May 9, 2025, and have no record of returns will receive a complimentary mouse case.) Mouse case introduction video: · June: Aqua Green and White · July: Red and Fluorescent Green Detailed sales date for other color options will be announced on our social media. For more information, please refer to the product page or contact customer service. US: EU:

VAXEE.co

53,139 views • 1 year ago

The Far Left and Far Right are both conspiracy theory driven, but are opposites. —— The Far Left are what I call “God-followers”: The Far Left tends to believe great GOODNESS in the world is only due to the actions of a centralized cabal — the “God Cabal” — intentionally pushing goodness on the world through smart top-down policies and interventions. In the absence of these policies, the world would mostly be bad. Free markets and free expression, each which works via decentralized mechanisms, are bad, and any good coming from them is due to the policies of the God Cabal. For Covid, the good that occurred was due to their interventions. In conflicts between strong and weak, while the world is by and large bad (outside of the God Cabal), the least bad are those who are weak, with the least agency. These are the “victims” that the Far Left sides with. For 9/11 the weak were the Islamists, and so the U.S. must have deserved it. For 10/7, the weak were again the Islamists, and so Israel deserved it. —— The Far-Right are what I call “Demon-Battlers”: The Far Right, on the other hand, tends to believe great BADNESS in the world is only due to the actions of a centralized cabal — the “Demon Cabal” — intentionally pushing evil on the world through sophisticated top-down subversive manipulation. In the absence of these evil influences, the world would mostly be good. Truly free markets and free expression, each which works via decentralized mechanisms, are good, and any bad coming from them is due to the manipulations of the Demon Cabal (which does happen a lot). For Covid, the bad that occurred was because the Demon Cabal engineered the whole thing. Not just the interventions, but the emergency itself, the virus (if there even are viruses), etc. And it was engineered so as to do a Great Reset, put chips in us, depopulate, etc. In conflicts between strong and weak, while the world is by and large good, the least good are those who are strong, with the most agency. These are the oppressors that are enjoined with or being manipulated by the Demon Cabal. For 9/11 the strong was the U.S., which itself — controlled by the Demon Cabal — brought the towers down. For 10/7 the strong was Israel, which itself massacred Israeli’s civilians (or perhaps there wasn’t a massacre at all), under the influence of the Demon Cabal. Point here is that we tend to characterize “Left” and “Right” differently in different contexts. Question is, what more general characterization might explain all the associations? For the Left, it has to explain not just socialism and equality of outcome, but, for example, Covid interventions, and siding with radical Islam in 9/11 and 10/7. For the Right, it has to explain not just support for capitalism but, among many things, the *very* far Right tendency to be basically against capitalism, the Plandemic-ers, and believing 9/11 and 10/7 were psyops. The (non-far) Left has a general tendency to see good in the world coming from well-intentioned centralized policies, and bad coming from too much decentralization and freedom. The (non-far) Right has a general tendency to see bad in the world coming from well-intentioned centralized policies, and good coming from decentralization and freedom. Both are “reasonable,” but the “far” versions of each become entirely crazy and dangerous.

Mark Changizi

52,383 views • 1 year ago

Yahya Sinwar, the leader of Hamas and the mastermind behind the October 7th massacre, has been killed. His death marks a pivotal moment in this tragic war that started just over a year ago. Sinwar rose to power as a founding member of Hamas’s military wing, the Qassam Brigades. His ruthless tactics earned him the name the Butcher of Khan Younis for his role in identifying and executing by hand Palestinians suspected of collaborating with Israel. In 1989, Sinwar was arrested for orchestrating the kidnapping and murder of two Israelis and four Palestinians, leading to a life sentence in prison. In 2011, he was released in exchange for an Israeli soldier held hostage in Gaza. Despite receiving life-saving medical treatment from Israeli doctors during his time in jail, Sinwar remained committed to violence. He rose to the top of Hamas’s leadership in Gaza by 2017, strengthening the group’s ties with Iran and transforming Gaza into a launchpad for terror. Sinwar’s October 7th massacre didn’t just target Israel, it pulled the entire region into war. Gaza, the West Bank, Lebanon, Syria, Iraq, Yemen and Iran were all drawn into this conflict. Tens and thousands of lives have been lost, and the region has been plunged into chaos with both Israelis and Palestinians suffering under the weight of his reign of terror. Sinwar did not promote peace or a two-state solution. He saw the coming of regional peace between Israel and the moderate Sunni countries of the region, and he deliberately acted to destroy it, sacrificing Gaza in the process. Under Sinwar’s leadership, Gaza was plunged into deeper suffering with generations growing up in conflict and despair. Now, with Sinwar gone, the world is one step closer to dismantling Hamas and breaking the cycle of violence in the Middle East. It is time to free the Palestinian people from the jihadi grip of Hamas and bring the hostages home.

Noa Tishby

109,811 views • 1 year ago

🚨 WEIGH IN 32 🚨 •285.7 pounds (ATH 530+ pounds) •Frank has lost 250+ pounds overall This week, Frank had his blood drawn and got a full report from his doctor. The results speak for themself: When Frank was 530 pounds+ / A1C 11+: •High risk of premature death •High risk of stroke •High risk of stroke •High risk of kidney disease •High risk of vision loss •High risk of neuropathy •A LOT of medications This week’s doctor report results: •Blood sugar: A1C 5.5 (normal) •Lost 250+ pounds overall •All medications lowered or stopped •Moving toward diabetic remission —Biggest update: Frank’s BMI is now 43.5 down from 69.9 in 2020 that’s a 37% reduction under 6 years! This is almost UNHEARD OF. Taking all of this into account, Frank has added QUALITY YEARS to his life. Frank’s doctor also gave Frank a referral for a consultancy on plastic surgery for his excess skin. His excess skin likely weighs at least 30 pounds, which means he’s probably already around 250 pounds. Frank came in to work looking a bit dejected about weigh in today. He always wants to lose big numbers to inspire his fans and doesn’t want to let anyone down. He expected to be about flat. I continue to impress upon him that he is past the era of shredding pure fat and is now building foundational muscle around his body. Frank’s weight has consistently fluctuated between 279-284 over the last couple of months. His diet has been good and obviously his walking and lifting have been consistent, so lack of weight loss has not been a concern. He even hit 3 30K step days this week. As we suspected, the plateauing weight is due to Frank rapidly gaining muscle faster than he is losing fat. He has been lifting weights 3X per week for almost 6 months and the results are hitting. His muscle development is clear in both appearance and increased strength + agility. In the Frankie Fitness videos, you can see Frank beginning to really struggle at the end of sets with significantly improved form (core strength). SO SO SO PROUD OF FRANKIE … I know he prefers seeing massive weight loss weigh ins, but his body is transforming in the best, natural, healthy way. His wins are more complex and substantial than just weight loss now. And by the end of our walk through the NYC sweltering heat, he agreed and smirked. Anudder weigh in in the books

Matteo Piper Jenks 🧲 🇮🇹

359,978 views • 1 month ago