Same model, but evolving how the team is organized brings GPU kernel optimization close to expert-written code 🚀 Evolving Agent Teams 📌 Overview UC Berkeley's EAT co-evolves a coding-agent team's configuration (roles, coordination, knowledge management) together with the solution. ⚠️ The problem ・Agents spend their budget on easy-to-implement directions and get stuck in local optima ・A fixed team can't keep up as the work shifts from debugging to profiling to redesign 🧬 Method ・A supervisor diagnoses failure patterns and designs the next team ・It invents roles from scratch, such as Profiling Scientist and Knowledge Integrator ・A persistent file system carries learnings and artifacts to each new team 📊 Results ・MLA decode (H100): 1.21x over FlashInfer with extended budget, beating it on 36 of 47 workloads ・A fixed four-role team reached 0.31x, and Claude Code plateaued around 0.5x ・GDN prefill (B200) rose from 2.2x to 4.06x, and GEMM+AllReduce reached 1.56x 🔭 Why it matters Without changing the model, adapting the organization itself breaks through plateaus. #AIAgents# #GPUKernels#

Modern AI workloads rely on a growing range of specialized GPU kernels, and no single kernel generation backend is optimal across every operator. At #PyTorchCon# North America 2026, Liz Li and Jiahui Cao (@AMD) will present “Extending TorchInductor with FlyDSL: A New MLIR-Native Backend for High-Performance GEMMs.” Liz and Jiahui will share how FlyDSL, AMD’s Python-native and MLIR-based GPU kernel DSL, integrates with TorchInductor’s compilation and autotuning pipeline while preserving the torch.compile user experience. The session will also present performance results for transformer training and inference on AMD Instinct GPUs, comparing Triton and FlyDSL implementations. Register for PyTorch Conference North America 2026:

NVIDIA built its own coding harness to optimize CUDA GPU kernels and achieved a 100% score on ARC-AGI-3’s 25 public games, solving all 183 levels. With agents, we'll move from a world where it's quite hard to run, optimize, post-train your own AI models and kernels to a world where virtually everybody can do it. 100 million AI builders when?

every research team needs to spend some time learning how their modeling code lowers down to gpu kernels example: rms norm vs layernorm industry assumes rms norm is cheaper. you don't need the x̂, it looks simpler, so it must be faster, so it got adopted but that's not true. both rms and layernorm are memory bound kernels (the amount of time it takes to get the data to gpu cores is longer than the amount of time it takes to do the computation on said cores) both take the same amount of time e2e so can probably train the model right with either, but maybe it makes a difference (e.g why diffusion models still use norms with an affine shift e.g adaLN) i get why. as open-source models get better and, inevitably, commoditized across the inference providers, overall serving speed determines user experience determines which model gets adopted but there's no reason to superstitiously avoid free things

While the Kaggle competition demonstrated AIRA₃’s capabilities in a specific domain, the system itself can generalize across distinct domains: changing only the task specification. In an internal benchmark, AIRA₃ achieved a 27% latency reduction on production GPU kernels, and gold-level performance in another Kaggle competition translating 4,000-year-old Akkadian clay tablets into English. We're early, and hard problems are still ahead of us. But we believe a system that compounds its own knowledge is the right bet. As we continue to develop and scale, we’re excited about its potential to accelerate AI research and unlock recursive self-improvement.

GPT-5.6 Sol, running in Codex, helped make GPT-5.6 cheaper to run. It rewrote production GPU kernels and helped improve the smaller model used to speed up generation. Kernel improvements helped cut end-to-end serving costs by 20%.

Speaker Spotlight: Melih Elibol, Research Scientist @NVIDIA On Sept 9, Melih will show how Rust's ownership model eliminates GPU data races in a new open-source DSL for high-performance GPU kernels. No unsafe needed. Find on schedule: Register: #rustconf# #rustlang#

Open-source AI stack from DeepSeek for building your own agent: ▪️ DeepSeek Harness (dsh): Connects a model to tools and manages agent sessions. ▪️ DeepSeek-V4.1-Flash and other models ▪️ DeepInfer: An inference engine for running models. ▪️ 3FS: A distributed file system for storing and moving data ▪️ deepseek-recipe: Translates API requests into model prompts and parses responses ▪️ FlashMLA: Optimized attention kernels ▪️ DeepGEMM: Optimized GPU kernels for matrix operations ▪️ DeepEP: Moves data between accelerators ▪️ DeepJIT: Compiles, caches, and loads GPU kernels at runtime ▪️ TileKernels: Optimized kernels for operations like routing and quantization Read our full guide to how the pieces work and fit together, what it takes to run them, and what’s still missing from the stack:

🚨 China's Big Bang Chips Launch - Ox-Alpha's 100T Token Capacity Mystery Solved Ox-Alpha (GLM 3.5 Flash) served 42 trillion tokens processed in 6 days. For free. On Chinese chips. Here's how they pulled it off: → Built a custom SGLang inference engine with disaggregated Encode-Prefill-Decode architecture → Used a GLM-5.3 infrastructure agent to write its own GPU kernels and debug bottlenecks — the model literally optimized its own serving stack → Achieved 3× end-to-end performance gains and Nvidia-level cost parity on domestic silicon → Backed by Zai's 1GW data center running entirely on Chinese chips — 10,000+ chip clusters, zero Nvidia silicon (i was wrong) Export controls were supposed to slow China down. Instead they forced Zai to build a fully sovereign AI stack — and it works!! China just proved to the world that they are not compute constrained

for my final post as @waterloo_intern, i'd like to ‘outroduce’ myself. it takes exactly 4 years and 8 months to make a waterloo intern. today is the last day of mine. the full life cycle, in four stages, is as follows: 1- interview hazing 2- cali or bust 3- inflection point 4- convergence stage 1: interview hazing a cohort of high school grads are ingested by waterloo. they are put through 4 months of academic bootcamp; the pre-requisites of the academic program they’re in. waterloo filters out the students who could not pass. they are put through another 4 month bootcamp, while simultaneously FORCED to find an internship for the following 4 months. this leads to students with ZERO experience applying to every job that is semi-adjacent to their field of study (a strategy, known in the literature, as spray-and-pray). it immediately backfires with a sea of interviews leading in a 1:1 fashion directly to infinite rejections. eventually, the student finds something…anything. this stage lays the foundation. rejection becomes noise so repetitively played it no longer registers. failure in an interview doesn’t matter. failure itself doesn't matter. for me, this unfolded as internships with blackberry and ford, summer of 2023 and winter of 2024, respectively. stage 2: cali or bust the next step by the program is to inject absolute dissatisfaction into its students (now in their second/third year). a CACOPHONY of immense fomo materialized as the university’s official unofficial motto: “cali or bust”. you must find your next internship in san francisco. your peers are doing it. all your smart friends have done it. you must do it too. failure to cali means your experience is BUST. you don't need to have a plan. you don't need to have anything. just go to san francisco. find an internship in sf. immediately. right now. for me, this was a cold email to a director at tesla and an internship with them in cali, fall of 2024. stage 3: inflection the reason behind step 2 appears here. somewhere, along the way, during your time in sf, you find your purpose. it could be as a direct result of your internship, but, more often than not, it is because of the people that you are surrounded by in cali. this is the inflection. for me, while at tesla, i was mentored by @pavanjayasinha , at the time interning at @Modular. he got me interested in gpu kernels & referred me there for my following internship. it was here that i had my second inflection point (due to the mentorship of fabio & hengjie, and the supervision of absolute field giants like @clattner_llvm), and decided to work beyond kernels and across the entire inference, serving, and model performance stack. stage 4: convergence having discovered the field you want to work in, you do your (usually last) internship in one of the companies within said field. here, you are at a disadvantage, being both a new-comer needing to onboard and inexperienced. but... you are also able to work on interesting projects / side quests that most full-timers are all but too busy to dive into. your time is also protected from a majority of the meetings others have to attend (in part because you are not important enough to have your presence expected), allowing you to make faster progress on assigned projects, and jump from one project (or team) to another. your title (intern) allows you to ask stupid questions of those around you, and so you can brain rake across the stack with no regard to who sits where in the company’s hierarchy... usually leading your questions with something to the effect of ‘can you help a new little intern onboard to X’, to which most people respond quite positively. in my case, this last stage took place over the past 8 months with @baseten, and i genuinely cannot imagine it being any better. i rejoin baseten full-time, but i wish i could have stayed an intern forever...