Async RL decouples rollouts from training, and that’s why... Echo-2 is so efficient. Distributed actors on Echo-2 collect rollouts on their own schedule while the learner updates continuously. Less waiting. Higher throughput. Here’s an illustration👇show more

we just released a new blog "Training a coding... agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:show more

AMMs were a solid start, but now we’re seeing... a clear demand for CLOBs, especially for highly liquid assets. Why? --- CLOBs are significantly more capital efficient than AMMs. In this simulation, CLOBs required 80% less capital while providing: - Lower slippage across all trade sizes - Less leakage to arbitrageurs (aka market makers in CLOBs) - Improved price discovery compared to the xy=k curve used in AMMs So why did we start with AMMs? --- Ethereum L1 isn't a high-throughput chain. Early DeFi primitives were designed with Ethereum’s throughput constraints, which made running an on-chain order book impractical. What’s changed? --- We now have better infrastructure. High-throughput chains could facilitate more frequent quote updates, resulting in tighter spreads and allowing CLOBs to achieve even more capital efficiency. Thanks to enzo for the inputs.show more

RL is painfully slow 😭 — bottlenecked by super-long... CoT rollout. 🔭 Sparse attention should help, but naive sparse rollout hits a brutal efficiency–stability tradeoff: A tedious trial-and-error sparsity sweep for each dense policy is required before an actual RL run. 🐤Sparrow chirps no more pain! Introduce Sparrow: Sparse Rollout for stable and efficient long-context RL. Sparrow finds that: 💡As long as we keep the tail distribution mismatch throughout the sparse rollout above a critical threshold, the RL training will be stable. 💡Even cooler! Through comprehensive control studies of Qwen3-1.7B, 4B, 8B thinking models RL with 40K rollout max length, the critical threshold stays constant across model sizes. 💡Sparrow then finds the optimal dynamic sparse schedule to reach the threshold with minimal cost. 💡Sparrow's findings are empirically validated to generalize in Qwen3-14B, and hold on both Math and Coding RL. 🐤Sparrow empirically helps achieve 2.2× / 2.4× / 2.0× rollout speedup on Qwen3 1.7B / 4B / 8B thinking models, while keeping training stability over extended RL steps. We release the 🐤bird in the following formats. [1/n] Paper: Code: Blog:show more

The party never stops in Colombia!!!! Luiz is working... on the title and thumbnail for next vlog I’m so fucked up i can not hahaha that’s why you have best friends like luiz! We have 2 colombiana girls waiting for us tho 🥰🥰🥰🥰 Hope you all had a great weekendshow more

we made distributed inference verifiable with <1% overhead. verification... is critical for any distributed system. in a trustless network, actors may swap your 70B model for a cheaper 8B one to cut costs. until now, maintaining inference integrity meant either doubling your cost (redundancy) or exploding your latency (zkp). we created veri: an on-chain verification layer light enough for high-throughput frameworks like Parallax. it hits the economic sweet spot through architectural elegance: 1. commit-sample-verify we don't prove every step; we check a random slice using game theory. workers commit to their work before the audit. cheating becomes statistically irrational, allowing a 1% sample to secure the entire sequence. 2. simultaneous execution inference and verification happen simultaneously on the same worker pool. we don't need a separate "verifier set", so compute utilization stays high. find out more about the architecture and benchmarks: paper: blog:show more

This week’s ChatGPT feature drop - Aug 21: Another... Friday, another roundup of what we shipped this week: 1/ Recent photos: Long press on + menu to quickly attach recents on iOS - this is a fun, power user feature. just beautiful design! also a handy shortcut. 2/ Time: We made it so ChatGPT better understands your current time. It was surprisingly bad at this before, for some interesting, complex reasons. Now it's better! 3/ Long convo loading: Long conversations load and run faster on Less waiting = better. 4/ 'No internet' errors: We show clearer updates in the UI when you’re waiting for an internet connection. It's so frustrating when ChatGPT times out. Now at least you'll know why, when it's the internet. thanks to the crew that keeps shipping every week and hope you all enjoy 🚀show more

40,000 doctors applied for just 10,000 specialty training jobs... in 2025. Tens of thousands were turned away because the number of NHS training jobs is capped by the Government. Doctors need these jobs to become specialist Consultants/GPs. Some are now facing unemployment. Others are stuck in short-term, insecure roles with no clear path to becoming a GP or Consultant. At the same time, patients are waiting weeks for GP appointments and months for surgery. We’re told there “aren’t enough doctors” - and that’s true. England has fewer doctors per person than many comparable European countries. So we’re an under-doctored country, while blocking thousands of doctors from progressing in their careers and helping cut waiting lists, If we don’t expand training numbers (including post CCT jobs), we won’t get the GPs and specialists patients urgently need. This is NHS workforce planning failure on a national scale.show more

May be the most difficult hurdle for a wrestler... to clear—trusting that the same offense that built the lead is what protects it. The instinct is to protect the lead, but the approach that creates better odds, less regret, and is simply more logical is to keep wrestling the same way that created the lead. Stay on the attack, stay in position, and keep applying pressure. The one who needs to change and create opportunity is the opponent who is losing. So why would you go ahead and assist them in this by back peddling and lead protecting, opening up two windows that weren’t there before? 1 – Stall points 2 – Allowing your opponent to solely focus on their offense as they no longer have to respect your attacks. You hear wrestlers in post-match interviews “We do this all the time in the room! Down 2 with 30 seconds left and working to find a takedown.” Not so much the opposite… “up 2 with 30 seconds left and just have to find ways to back up to protect the lead!” So why resort to something you don’t practice in the most crucial moments of a match—or your season? This is from the excellent series “The Climb” by Stilly Boys on YouTube — a segment from Episode 2: Road to the Big 12s.show more

⚽️👇 Activity Description Below 🗣️ This is a ‘Back... 3 Build Up Pattern’ from Drill Library. This passing activation allows players to work on developing build-up patterns and relationships while playing in a back 3 shape. The build-up is heavily focused on a 3-2 build-up shape with central combinations. The aim is to create a fluid transition from the back, ensuring the ball is circulated efficiently before being played into the attacking areas. ✅ Click below to view 700+ specialised & adaptable training activities ↙️ 💻show more

🇨🇳 CHINA BUILT A SOLAR EMPIRE ON THE ROOF... OF THE WORLD On the Tibetan Plateau, solar panels stretch farther than the eye can see - covering an area 7x the size of Manhattan. The Talatan Solar Park, plus wind and hydro, now powers everything from trains to AI data centers, while sheep still get to graze under the panels. Qinghai's clean energy is so cheap and efficient, it costs 40% less than coal, and helps China manufacture even more solar tech to sell worldwide. This is how China plans to win the future of energy: high-altitude, high-volume, and high stakes. Source: New York Timesshow more

USC's Gary Patterson is the king of split-field 4-2-5... defenses and knows how to fit the box against Spread sets as well as anyone else. When the RB is set away from the 3-technique, the DE on that side can wrap back inside to add numbers to the box. Typically, an offense will zone away from the back. With a base block, the DE wraps inside to take the A-gap, while the 3 tech. works vertically, allowing the Guard to push him into contain and pin the Tackle. The lone LB can then be patient and track the back. It amplifies coverage by slowing the LBs and exploiting the offense's blocking. Learn more about Read-Pop (Echo) stunts and other ways to fit the box with 7-man spacing on MatchQuarters: --show more

Dear Black People, Watch This Video And Please Help... Me Understand! 1. Why Do FBA Men Like Their FBA Women Like FBA Trans? 2. Why Is The Longer The Weave On A Black Woman The Higher Her Status? 3. Why Do Black Women Tend To Sound Like Black Men? 4. Why Do FBA Men Find Their Women Attractive With A Body, Neck & Face Full Of Tattoos! 5. Why Do FBA Women Praise & Sexualize Other Black Women's Half Naked Bodies So Often? 6. Why Do Black Women Have To Show Their Butts To The Cameras? 7. Why FBA Men Prop Up Women Who Look Like Prostitutes? 8. Why Do FBA Black Women Have So Much Fake Stuff On Them Yet Claim Everyone Wants To Be Like Them? 9. Why Do Women Whose Whole Look Is Fake Turn Around & Calling Someone Else Unattractive? 10. WTF Is A Baddie? Go to and see my video about all of this!show more

Alex Karp from Palantir on CNBC just now dropping... a truth bomb. “More people have died from fentanyl in the last year than all service members since World War 2. And almost all those people are working class and that’s why no one cares. If they were Yale grads, you’d have a bomb being dropped on some South American country and it would stop.” Interesting commentary especially in light of the Trump administration’s decision to sink known drug smuggling boats coming in from South America who just happen to be “fishing” in the Gulf of America while speeding along at 100mph. Are the lives of working class Americans worth less?show more

We’d like to share some updates from within the... Forge. While minor updates may roll out at a slightly slower pace, our primary focus is dedicated to an exciting upcoming feature: Machina Foundry. What is Machina Foundry? Machina Foundry is an AI Agent Builder Platform that allows users to define an agent's purpose, functions, and objectives through simple, natural language prompts. Once an agent is created, Alchemists can seed liquidity in $ALCH for their agent. Liquidity is placed in Meteora pools, with 50% allocated to the ALCH ecosystem and the remaining 50% locked permanently. This structure ensures the ecosystem benefits from pool fees, enabling periodic ALCH token buybacks and supporting long-term growth. Additionally, the Foundry introduces a flywheel effect: purchasing an agent requires acquiring ALCH tokens, further integrating the ecosystem with the token economy. Agents built in Machina Foundry are fully customizable and reflect unique personalities. They can create apps and tools tailored to their character, powered by Alchemist AI’s robust technology. While similar to Azarus, these agents bring an added layer of individuality, ensuring that their creations vary significantly based on their distinct traits and configurations. Imagine a network of thousands of AI agents, each contributing diverse applications and tools, driving creativity and value across the ecosystem. Now that’s Magic!🪄✨show more

“I don't know when they'll cut my internet again,... but we will not stop. We aren't waiting for outside help. It’s us and our Shah, and that’s enough for our revolution.” 🔥 An Iranian protester just managed to post this from inside the country. THIS is exactly why the Iranian people are the absolute best boots on the ground to defeat the Islamic Republic. We have the courage, the numbers, and the unwavering will to overthrow this terror state. You don't need to send your own troops. Just arm the people already fighting on the inside!show more

The number of YouTube channels earning six figures or... more in revenue from TV is up 45% year over year. Excited to announce new updates to help YouTubecreators get even more out of our fastest-growing screen: ⬆️ An increased thumbnail file size limit (from 2MB to 50MB) so you can upload crisp, high-quality 4K thumbnails that pop on TV ☀️ Higher resolution videos to make your entire back catalog shine (with the option to opt out) 📺 Immersive channel pages on more surfaces so viewers can find and enjoy your content more easily 👇show more

THIS CHINESE TRADER BUILT AN AI AGENT IN 2... HOURS AND LET IT TRADE A $70K ACCOUNT FOR 72 HOURS STRAIGHT he opened Bloome, wrote his trading rules in plain English, picked the signals, set risk limits and let the agent run on its own for 3 days it scans the market, filters setups, manages exposure and executes trades without him sitting there watching every candle most traders still hesitate on entries, close positions from panic and refresh charts all day he just built the system once and let it trade without emotion this is probably where trading goes next: less screen time, more agents, more systems that don’t break their own rulesshow more