Does your Claude model really need Claude Code…? 🤔 We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge: 1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost 2️⃣A simple harness can be competitive 3️⃣The native harness isn’t always the best. Millions of people are using coding agents, but the impact of harness choice remains unclear. (1/n) More details in the thread. 🧵

SkyDiscover // @shulynnliu, @mertcemri, @shagarw21, @AlexKrentsel (@UCBerkeley) — An open-source framework that uses AI to discover better algorithms and assemble complete systems from a specification.

Superintelligence should benefit human intelligence. Mathematics is our starting point. We’re launching Readable Intelligence, and we invite the math and AI communities to build it with us. 📄 Pick a paper. 💬 Flag an unclear step. 🙌 Contribute a more readable version.... readable-intelligence.com/

FreeToken: flashml.ai/

While exploring diffusion language models (DLMs), we kept asking ourselves: given speculative decoding, a lossless acceleration technique, why do we still need DLMs? 🤔 Our latest work, ORL, tries to answer this question: AR optimizes intelligence 👆, speculative decoding optimizes speed 👉, and DLMs optimize the Pareto frontier of speed and intelligence 🚀. Through co-design across the algorithms, the training and inference engine, and the kernels, ORL pushes out a new Pareto frontier of intelligence and speed. Similar performance, much faster than speculative decoding. You say lossy spec? ORL-DLM is better too! 😎 🌐 Project + live demos: introspective-diffusion.gith…

🚀 Excited to introduce 𝐃𝐮𝐨𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫! This isn’t benchmark gaming. It’s AI agents shipping real PRs and beating human-expert hardware baselines. 🔥 𝐓𝐡𝐞 𝐛𝐫𝐞𝐚𝐤𝐭𝐡𝐫𝐨𝐮𝐠𝐡𝐬 🚀 🛠️ Real optimizations merged into main branches (e.g., FreeToken #619) 🧠 First superhuman CPU prefetcher: +3.93% | +6.31% overall, outperforming AlphaEvolve ⚡ 2.68× TPU kernel speedup | 1.92× faster vLLM serving | 🏆 #1 on NVIDIA SOL-ExecBench (B200) 💡 𝐓𝐡𝐞 𝐬𝐞𝐜𝐫𝐞𝐭 𝐬𝐚𝐮𝐜𝐞? Agents co-discover fast evaluation proxies (“yardsticks”) to slash slow eval bottlenecks and compound real performance gains at scale. 🤝 Incredible collaboration across @GoogleCloud, @BerkeleySky & @StanfordAILab! 🔥 We’re pushing the frontier of AI-driven systems optimization. And we’re just getting started. 🚀

Do AI harnesses matter? Yes, KISS Sorcar outperforms three coding harnesses on Terminal-Bench 2.0. On Terminal-Bench 2.0, KISS Sorcar (github.com/ksenxx/kiss_ai) solved 75.6% of attempts across seven models, the best rate on every model against Pi (70.0%), Codex CLI (65.7%), and Claude Code (65.1%), and in a paired rerun against Pi on Claude Fable 5, it won 80.0% to 68.9% at about the same cost per attempt. The three harness numbers come from the study below by @melissapan et al. We ran KISS Sorcar on the same 30 tasks and seven models, three attempts per task. We reran the Pi experiments: same tasks and prices, no turn cap for either agent, Pi at high effort with no budget cap. Pi cost $1.45 per attempt and KISS Sorcar $1.50; the gap was 11 points, paired with p = 0.016. On the 59 tasks the study didn't sample, KISS Sorcar solved 79.1%, and a second Pi rerun on 57 of them (two sat out because a dead package mirror broke the verifier) came out to 82.5% vs. 71.9%, a 10.5-point gap with p = 0.008, at 63 cents more per attempt. Two reruns on one model don't settle it, but on this benchmark the harness seems to matter more than the study's three harnesses showed. Full table and intervals: kisssorcar.github.io/blog/ha…

Releasing FLAT: Flash Linear Attention on TPUs 🚀 Pallas training kernels for Linear Attention, Mamba-2 and GLA. On one TPU v6e, forward + backward is up to 3x faster than FLA on an H100 for GLA, and 1.5x faster for Mamba-2. 🧵

Test time recovery 🤩 Glad to see community’s work in agent failures inspired by MAST! Congrats for the cool work!

Video generation now runs right on your laptop. 🎬 Introducing FreeVideo — bringing MiniMax H3 + Video DeltaNet to the hardware you already own. As little as 8GB VRAM + 16GB RAM. ComfyUI integration, LoRAs & custom workflows. Code: github.com/FlashML-org/FreeV…

AI systems should be designed for AI. If prefix caching gets in the way of free context access, we can fix the caching. This is the idea behind our Suffix Cache Reuse (SCR) for CLR serving. Check out our blogpost for details (fun visuals included)! wai-org.com/blog/clm/ Maybe the next step is to give CLMs direct agency over the cache itself. I’m excited to see what more creative and adaptive cache-aware context management could look like. It's also a good time to follow our account @waiorg , a student org we created to share our work on agents. Currently covering: TMax, CLM.

🧵 More stable PPO for LLM post-training. Introducing EasyPPO: just fix the critic. ❌ No new actor loss ❌ No new policy algorithm ✅ Actor update unchanged EasyPPO delivers better results with zero training collapse across all our experiments.

Claude Code vs Codex vs Pi: which coding agent wins? @melissapan, PhD candidate at UC Berkeley’s Sky Computing Lab and previous @arena intern, explored the hidden “harness tax”: how the system surrounding an AI model affects its cost and performance. She reports three surprising findings. One: harness choice impacts cost more than accuracy. To learn about all three findings, and what they mean for building useful coding agents on realistic budgets, check out the full video here: piped.video/4R8qZGtEnAE

Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think it's a valuable research direction to start thinking about now, especially w.r.t. harness design. alexzhang13.github.io/blog/2…

yayy I can finally use dashes again - I love to write with dashes pre llm 😛

🔥🔥🔥🔥

jev for semantic operators is a great example of how LLM data processing at scale requires rethinking across the stack, model architecture and learning algs included new work and many more thoughts on this coming soon 👀 in the meantime, check out lotus, star and stay tuned (link below)