Anshul (@anshulcreates) header
Anshul (@anshulcreates) profile picture

18 | chasing μs | Research musings and writings | Hoping to talk to cool people daily

Joined October 2024
my fav recent drops yall should check out 1. megakernels!! xitter.gq/xingyudang/status/2123… 2. I love these gpu/cpu optimization algos xitter.gq/architectlabs/status/2… 3. buzzword buzzword buzzword (but I love it) xitter.gq/simon_mo_/status/21231… 4. context shrinking is so underrated! xitter.gq/tufalabs/status/212335… 5. e-means. this is an obvious one
Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
3
50
ideas: 1. learn gpu engineering and architecture design (hard to make it monetized for moral purposes) 2. churn reducer (again lol) 3. a way to update prompts and models when new releases comes out automatically 4. automating the gta clipping money scheme, end to end 5. none of these. i’m under qualified. thoughts PLEASE 🙏
day 3 of the $100K hackathon. 4,800+ builders, someone's already at $9,600, and 28 days left to knock them off the throne 👑 free to join, link below
1
93
The best first paper to learn AI research came out 5 years BEFORE Attention Is All You Need. ImageNet Classification with Deep Convolutional Neural Networks (2012) by: Alex Krizhevsky, @ilyasut, @geoffreyhinton just read it and you'll understand cs.cmu.edu/~jeanoh/16-785/pa…
1
151
“personal” project and it literally took me a couple hours to recreate and several more to understand how much gains I can get from this beautiful man. keep shipping
Alright I finally have more to share on my personal project! Resin is a runtime for building fast LLM inference engines. The idea is to make all the generic parts of inference reusable, and leave only the model-specific forward pass + kernels to be written for each architecture / hardware target. To build an engine: 1. Import Resin's core. 2. Have an agent write a forward pass for the model you’re interested in, for your specific hardware (GPU or TPU or AMD), any kernels you want. Then, boom, you have a standalone inference engine. — I've been hacking over the past month (early post: xitter.gq/ekzhang1/status/208950…), and recently this week I got a lot of features + parallelisms working, inspired by DeepInfer. For DeepSeek-V4.1-Flash, it's nearly as fast as DeepInfer (within 5-10%), while being self-contained and supporting other models, plus metrics + HTTP server + prefix cache. Yesterday I wrote a skill for authoring models and put Codex to work on it overnight. The agent built a GLM-5.3-Flash engine with MTP, reaching up to 60,000 input tok/s and 16,000 output tok/s on B200 (EP4). So far inference "vibe coding" has not been very reliable or successful. I think it's because people start from scratch each time and reinvent lots of tricky things on the scheduler and server layers. But with a shared core, we can automate lots of inference work. The tricky bit is writing a fast forward pass. But Resin turns that into a standalone engine by just importing a library, and it becomes a verifiable goal for the AI. Code: github.com/ekzhang/resin
1
132
the barrier to success is so low, valuable information is so abundant you don’t have to go to Stanford to learn this stuff; the lecture links are RIGHT THERE its impossible to get a Stanford education without going there, but it is possible to learn all the technical skills a Stanford grad has and the NEW true value of the degree is in the people met, the experiences made, and the mindsets changed
Sometimes I wish I went to Stanford they have a whole course where you literally build an LLM from scratch (tokenization, custom kernels, distributed training, alignment, etc.) this kind of exposure as an undergrad is so nice
2
78
questions as a new researcher i’d like answered: what qualifies a research question as ‘good’? how does one know it’s feasible to go down that avenue for the next few months, if not years? what does one do when they pursue this journey for no real, novel results? why keep pursuing a research project even if you aren’t seeing result?
1
75
3 research-grade projects I'd start in my first year of mechanical engineering (each one could turn into a real paper): a. Passive Thermal Management Optimization — (6/10) - What you're studying: How heat sink geometry, fin pitch, and material thermal conductivity drive convective and conductive heat dissipation. - Step 1: Build parametric CAD models so you can sweep fin designs fast. - Step 2: Run conjugate heat transfer simulations on every variant. - Step 3: Validate the best designs against thermocouple measurements under a controlled heat load. - Research question: Can a topology-optimized fin geometry lower junction temperature and thermal resistance at equal or lower mass? - Methods: Parametric CAD, FEA, CFD (conjugate heat transfer), thermocouple validation, multi-objective optimization. - Potential contribution: A geometry optimization workflow that trades off thermal resistance against mass under manufacturability constraints. Papers & other resources to get started: - A Heat Transfer Textbook — Lienhard & Lienhard. - Review literature on heat sink design and optimization. - Literature on topology optimization for heat transfer. - Introductory CFD and heat-transfer tutorials. b. Small-Scale Wind Turbine Blade Optimization (7/10) - What you're studying: How chord distribution, twist angle, and airfoil selection drive lift-to-drag ratio, power coefficient (Cp), and efficiency relative to the Betz limit. - Step 1: Build parametric blade models so you can sweep chord, twist, and airfoil profiles fast. - Step 2: Run Blade Element Momentum (BEM) analysis on every variant, then CFD on the top candidates. - Step 3: Validate the best designs against wind tunnel or field measurements of power output across tip speed ratios. - Research question: Can a surrogate-optimized blade geometry raise Cp at a target wind speed and tip speed ratio? - Methods: Parametric CAD, BEM theory, CFD, surrogate modeling, wind tunnel validation. - Potential contribution: A surrogate model that maps blade geometry to power coefficient across operating conditions. Papers & other resources to get started: - Wind Energy Explained (Manwell, McGowan & Rogers). - Literature on the Betz limit and rotor aerodynamics. - Blade Element Monces. - Review literature on aerodynamic optimization of wind turbine blades. c. Solar Panel Thermal Performance Modeling (6/10) - What you're studying: How cell temperature, the power temperature coefficient, and convective and radiative heat transfer affect photovoltaic conversion efficiency. - Step 1: Instrument a small PV panel with thermocouples, a pyranometer, and I-V curve tracing. - Step 2: Build a transient thermal model of the three cooling variants (fins, phase change materials, rear-side airflow). - Step 3: Validate model predictions against measured cell temperature and power output under varying irradiance and ambient conditions. - Research question: Which passive cooling approach gives the largest reduction in cell temperature and gain in power output per unit cost across weather conditions? - Methods: Thermal modeling, sensor instrumentation, I-V characterization, data analysis, optimization. - Potential contribution: an electrical model that predicts efficiency gains from different passive cooling designs. Papers & other resources to get started: - Photovoltaic Systems Engineering (Messenger & Ventre). - Review literature on temperature dependence of photovoltaic performance. - Literature on passive cooling in PV modules. - Introductory solar irradiance and PV modeling tutorials.
2
2
276
1. Passive Thermal Management Optimization - A Heat Transfer Textbook — ahtt.mit.edu/ - Heat Sink Design Optimization — scholar.google.com/scholar?q… - Topology Optimization for Heat Transfer — scholar.google.com/scholar?q… 2. Small-Scale Wind Turbine Blade Optimization - Wind Energy Explained — scholar.google.com/scholar?q… - Betz Limit — scholar.google.com/scholar?q… - Blade Element Momentum Theory — scholar.google.com/scholar?q… 3. Solar Panel Thermal Performance Modeling - Photovoltaic Systems Engineering — scholar.google.com/scholar?q… - Temperature Dependence of Photovoltaic Performance — scholar.google.com/scholar?q… - Heat Transfer in Photovoltaic Modules — scholar.google.com/scholar?q…
39
I'd read these 15 papers before taking this course, they cover EVERYTHING from RLHF to recursively self-improving agents. (optional, but they just make everything click faster imo) links below! 👇 1. InstructGPT - SFT → reward modeling → PPO. - Human preference data + KL regularization. - The full RLHF stack. - Reproduce the pipeline conceptually 2. Deep RL from Human Preferences - Pairwise comparisons. - Reward model training. - Behavior without hand-specified rewards. - The foundation for RLHF. 3. Proximal Policy Optimization Algorithms - Policy gradients and clipped objectives. - Importance sampling ratios. - Advantage estimation. - Reproduce PPO on a small environment. 4. Direct Preference Optimization: Your Language Model is Secretly a Reward Model - Derivation of DPO from RLHF. - Preference-based training without explicit PPO. - Reference policy and implicit reward modeling. - Reproduce DPO. 5. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models - GRPO algorithm. - Group-relative advantage calculation. - Critic-free RL. - Reproduce GRPO. 6. Let's Verify Step by Step - Process reward models (PRMs). - Step-level vs. outcome-level supervision. - Verifying intermediate reasoning. - Train a small PRM. 7. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning - RLVR and R1-Zero's emergent reasoning. - Cold-start SFT, rejection sampling, and RL. - Distillation of reasoning models. - Reproduce a small RLVR experiment. 8. DAPO: An Open-Source LLM Reinforcement Learning System at Scale - Improvements to GRPO. - Dynamic sampling and clipping. - Token-level policy optimization. - Large-scale RL engineering. 9. Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters - Inference-time compute allocation. - Best-of-N sampling and search. - Verifier-guided reasoning. - Compute scaling tradeoffs. 10. Tülu 3: Pushing Frontiers in Open Language Model Post-Training - Full open post-training recipe. - SFT, DPO, RLVR. - Dataset construction and evaluation. - How to assemble an end-to-end post-training stack. 11. ReAct: Synergizing Reasoning and Acting in Language Models - Reasoning/action loops. - Tool-use architecture. - Action-observation feedback. - Implement a ReAct agent. 12. Reflexion: Language Agents with Verbal Reinforcement Learning - Episodic memory. - Self-reflection after failures. - Feedback-based improvement without weight updates. - Implement a reflection loop. 13. Voyager: An Open-Ended Embodied Agent with Large Language Models - Skill generation. - Skill libraries. - Ind feedback. - Open-ended learning in an environment. 14. Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation - Recursive code generation. - Self-modifying optimization systems. - Evaluating whether self-improvement actually works. - Reproduce a bounded self-improvement loop. 15. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents - Agents modifying their own code. - Evolutionary search over agent designs. - Empirical evaluation of self-modifications. - The distinction between agent improvement and model-weight improvement.
Big little announcement: @karinanguyen and I are excited to announce a brand new course on Full-Stack Post-Training, RL & RSI at @UCBerkeley. If you are a Berkeley student and want to learn about the frontier, enroll for the next semester.
2
5
341
no filler, no nonsense: these are the 15 videos that took me from zero GPU knowledge to GEMM optimization research: watch them in this order (links below👇): 1. GPU Architecture Explained by ExplainingComputers Overview: A visual intro to GPU hardware and parallel execution. - GPU versus CPU: throughput-oriented versus latency-oriented design. - Streaming Multiprocessors (SMs) and CUDA cores. - SIMT execution: threads grouped into warps of 32. - GPU memory hierarchy and hardware parallelism. 2. CUDA MODE Lecture 3: Getting Started With CUDA by GPU MODE Overview: A practical intro to CUDA programming and writing GPU kernels. - Host versus device: CPU launches the work, GPU executes it. - Kernel launch syntax and thread indexing. - Mapping data elements to threadIdx, blockIdx, and blockDim. - Basic GPU programming from Python. 3. CUDA MODE Lecture 4: Intro to Compute and Memory Architecture by GPU MODE Overview: A hardware intro to the compute and memory design behind kernel performance. - Compute versus memory: throughput limits versus bandwidth limits. - GPU memory hierarchy: global memory, shared memory, registers, and caches. - Memory bandwidth and compute throughput. - How daexecution time. 4. Memory Coalescing by NVIDIA Developer Overview: A visual intro to how GPU threads access memory efficiently. - Coalesced versus uncoalesced access patterns. - Warp-wide memory transactions: threads in a warp combine their loads. - Strided and contiguous memory layouts. - Transaction efficiency and wasted bandwidth. 5. CUDA MODE Lecture 9: Reductions by GPU MODE Overview: A practical intro to parallel reduction algorithms on the GPU. - Parallel sum and max reduction: combining N values into one in log(N) steps. - Warp-level reductions using shuffle instructions. - Shared-memory reductions and thread synchronization. - Handling non-power-of-two inputs and reduction tails. 6. The Roofline Model by Berkeley HPC Overview: A performance model for finding hardware bottlenecks. - Arithmetic intensity: FLOPs per byte of memory traffic. - Peak compute throughput versus peak memory bandwidth. - Compute-bound versus memory-bound: which roof limits the kernel. - Estimating achievable kernel performance. 7. CUDA Performance Checklist by GPU MODE Overview: A practical checklist for diagnosing and optimizing kernel performance. - Memory coalescing and access patterns. - Occupancy versus register pressure: more registers per thread means fewer resident warps. - Control divergence: threads in a warp taking different branches. - Kernel bottleneck diagnosis. 8. Nsight Compute Introduction by NVIDIA Developer Overview: A practical intro to profiling individual GPU kernels. - Kernel-level profiling and hardware performance metrics. - SM throughput, memory occupancy. - Warp stall reasons: why warps wait instead of issuing instructions. - Roofline analysis and bottleneck identification. 9. Nsight Systems Introduction by NVIDIA Developer Overview: A system-level view of the full CPU-GPU timeline. - CUDA API calls, kernel launches, and memory transfers. - CPU launch overhead and GPU idle gaps. - Stream execution and operation overlap. - Synchs and pipeline analysis. 10. CUDA MODE Lecture 14: Practitioner's Guide to Triton by GPU MODE Overview: A practical intro to writing high-performance kernels in Triton. - Triton versus CUDA: block-level versus thread-level programming. - Program IDs, blocked tensors, and tile-based execution. -Array dimensions. - Tuning block sizes, warps, and memory layouts. 11. Fused Softmax in Triton — GPU programming tutorial channel Overview: Implements softmax efficiently by fusing multiple operations into a single GPU kernel. - Numerically stable softmax using row-wise max subtraction. - Reduction for maximum and normalization denominator. - Fusion of max, exp, sum, and division. - Eliminating intermediate global-memory writes and kernel launches. 12. FlashAttention From Scratch in Triton by GPU programming tutorial channel Overview: A practical intro to building an IO-aware attention kernel. - Tiled attention: splitting Q, K, and V into blocks that fit in SRAM. - Online softmax: running max and sum updated block by block. - HBM versus SRAM: keeping intermediates on-chip to cut memory traffic. - IO complexity and memory-efficient attention. 13. CUDA Streams and Asynchronous Execution by NVIDIA Developer Overview: A practical intro to asynchronous GPU execution and pipeline throughput. - CUDA streams: ordered queues of GPU operations. - Asynchronous memory copies with pinned host memory. - Overlapping computation and data transfer; synchronization, events, and inter-stream dependencies. 14. Hopper TMA / WGMMA Deep Dive by NVIDIA Developer Overview: A hardware intro to Hopper's data movement and Tensor Core mechanisms. - Tensor Memory Accelerator (TMA): hardware-driven multidimensional copies into shared memory. - WGMMA: matrix multiply-accumulate across 4 warps. - Producer-consumer pipelines: load warps versus compute warps. - Asynchronous execution and latency hiding. 15. CUDA MODE Lecture 23: Tensor Cores by GPU MODE Overview: A hardware intro to how Tensor Cores accelerate matrix multiplication. - Tensor Core matrix multiply-accumulate operations. - FP16, BF16, TF32, and lower-precision formats. - Matrix tile layouts and Tensor Core execution. - Precision, throughput, and performance trade-offs.
2
10
128
5,120
1. GPU Architecture Explained — ExplainingComputers youtube.com/results?search_q… 2. CUDA MODE Lecture 3: Getting Started With CUDA — GPU MODE youtube.com/watch?v=4sgKnKbR… 3. CUDA MODE Lecture 4: Intro to Compute and Memory Architecture — GPU MODE youtube.com/watch?v=lTmYrKwj… 4. Memory Coalescing — NVIDIA Developer youtube.com/results?search_q… 5. CUDA MODE Lecture 9: Reductions — GPU MODE youtube.com/watch?v=09wntC6B… 6. The Roofline Model — HPC educational resources youtube.com/results?search_q… 7. CUDA MODE Lecture 8: CUDA Performance Checklist — GPU MODE youtube.com/watch?v=SGhfUhlo… 8. Introduction to Nsight Compute — NVIDIA Developer youtube.com/results?search_q… 9. Introduction to Nsight Systems — NVIDIA Developer youtube.com/results?search_q… 10. Lecture 14: Practitioners Guide to Triton — GPU MODE youtube.com/watch?v=DdTsX6DQ… 11. Fused Softmax in Triton — GPU programming tutorial youtube.com/results?search_q… 12. Flash Attention Derived and Coded From First Principles With Triton — Triton tutorial youtube.com/watch?v=zy8ChVd_… 13. CUDA Streams and Asynchronous Execution — NVIDIA Developer youtube.com/results?search_q… 14. Hopper TMA / WGMMA Deep Dive — GPU programming tutorials youtube.com/results?search_q… 15. Lecture 23: Tensor Cores — GPU MODE youtube.com/watch?v=hQ9GPnV0…
1
4
280
small robot sims on GPU spend most of their time launching kernels. so i built a compiler (borrowing ideas from @hyperstition_cc's work) that reads a MuJoCo file and writes one fused GPU kernel. results on a T4: • matches MuJoCo's physics • 17-143× faster than MJX • small articulated robots, no contact (yet) 1. fusion alone makes identical code 3-4× faster same equations, the only change is going from many kernels per step to 1: • cartpole + controller: 3.4-3.6× (0.0138 → 0.0039 ms) • 2-link arm + controller: 3.1-3.4× (0.0145 → 0.0043 ms) • ball + floor contact: 4.0-4.1× (0.0136 → 0.0033 ms) unfused, fused and multi-step versions agree to within ~3e-8, so fusing doesn't change the physics. vs MJX: 1.6e-7 (cartpole) and 5.2e-8 (arm) at step 100, ball within 3.67 mm during the bounce and 0.01 mm after. 2. running many steps in one kernel adds another 3-15×, while the physics is cheap • cartpole: 0.0039 → 0.0006 ms (~6.5× over fused) • chain5 (5 joints): 0.0104 vs 0.0107 ms, tep does enough math, launch overhead stops mattering. 3. the compiler works on models it was never hand-tuned for • mass matrix, bias forces and accelerations match MuJoCo to 1e-13 to 1e-16 (arm2, chain5, and a branching tree with sliding joints) • the generated float32 Warp kernel matches mj_step to ~6e-8 after one ps, it stays within MuJoCo's own sensitivity to a 1e-6 nudge on all 3 models 4. same input files, generated kernels vs MJX (1 env / 4,096 / 16,384): • arm2: 44× (0.159 → 0.0036 ms) / 64× / 124× • chain5: 17× (0.178 → 0.0104 ms) / 40× / 112× • tree: 24× (0.178 → 0.0073 ms) / 61× / 143× 5. MJX on small models is mostly launch overhead • 94 kernels per step • ~0.14 ms per step whether you run 1 env or 1,024 • roughly 1.5 µs per launch, doing almost no physics this is NOT a mujoco replacement (yet haha), just an experiment so far. please let me know where I made mistakes!
10 things worth building with megakernels right now, and each one pushes real research forward!: 1. a real time world model - feed in frames + action inputs, predict the next latent states - hyperkernel fuses the rollout loop to cut inference latency 2. a genomics research assistant - visualize embeddings from a DNA language model - compare variant effect predictions across sequences - one interface for researchers to probe their results 3. a GPU-accelerated protein structure explorer - run ESMFold predictions in the browser with optimized inference - compare predicted structures and confidence scores side by side - an interactive tool for exploring how proteins fold 4. an autonomous driving simulator - run a vision model on simulated road frames, predict vehicle actions - hyperkernel speeds up the perception pipeline - visualize predicted trajectories and decisions in real time 5. an AI agent with long-term memory - an agent that remembers its research across multiple days - hyperkernel fuses memory-related ops to make local inference faster - the memory architecture is where the real research lives 6. a learned physics id dynamics and soft-body interactions with learned models - benchmark against a conventional physics engine - hyperkernel accelerates the learned approximation 7. a scientific discovery platform - pick a model and run simulations for any scientific question - analyze the results and generate a research report - hyperent inference backend 8. a real-time voice assistant - streaming speech recognition and generation with natural interruptions - fully local inference at very low latency - hyperkernel's moshi work is a reference for fast speech-model execution 9. a molecular simulation explorer - interactive visualization of molecular behavior - ML-based predictions with GPU acceleration - one platform for computational biology research 10. a mini AI inference platform - open-source serving system with model selection and performance profiling - hyperkernel provides the high-performance execution layer
16
840
this week in AI was WAY bigger than you realize here are the top 11 releases that directly affect jobs in kernel engineering, security research, sales, even chemical engineering: 1. Together Compiler (TKCC) AI agents continuously rewrite, test, and optimize GPU inference kernels • up to 2.6× higher tokens/sec vs optimized open-source engine configs • 1.7× throughput over available megakernels (Llama 1B, batch size 1) • 252× higher experiment throughput vs baseline agentic optimization systems • already serving production workloads together.ai/blog/together-co… 2. Scientist World turns research papers into executable environments with reproduced baselines • 1,000 verified environments • frontier AI scientists beat the reproduced human methods in 74% of evaluated improvement/innovation cases • 158% boost in a small model's research ideation after training on these environments arxiv.org/abs/2611.08478 3. SwarmEvolve hundreds of cheap agents share KV caches and search for better programs in latent space, without decoding every idea into text • tested across 21 math discovery and scientific computing problems • hit in a median of 3–8 min the scores other frameworks needed 60 min to reach • Qwen3 8B improved on solutions originally found by AlphaEvolve arxiv.org/abs/2611.08483 4. TCSExam benchmark for theoretical CS research: theorem proving, solver engineering, constructions, open conjectures • 100 expert-reviewed tasks • outputs checked by Lean or executable validators • GPT-6 Astra leads Full at 71.9, Opus 5.5 leads Lite at 72.3 arxiv.org/abs/2611.07962 5. SPAR self-evolving B2B sales agent that treats selling strategies as testable hypotheses, remembers each merchant, and updates its playbook using causal attribution • live A/B test on Alibaba's 1688 platform, 200,000+ merchants • 15.8% conversion lift • 35.8% growth in high-intent merchants arxiv.org/abs/2611.01695 6. OpenAI Linux Kernel Autonomous Pipeline autonomous vulnerability discovery and patch generation for the Linux kernel • 2,308 candidate findings • 1,724 vulnerabilities dynamically reproduced • 948 local privilege-escalation exploits, plus candidate patches for maintainer review • ~$1.2M in API costs openai.com/index/a-thousand-… 7. VFlash visual planning model that predicts points, bounding boxes, and trajectories jointly instead of one coordinate token at a time • lower prediction error than autoregressive decoding across 2D and 3D robotics tasks • ~25× faster 3D point-motion generation (preliminary) • 70.2% avg success across four MolmoSpaces suites vs 62.6% baseline arxiv.org/abs/2611.08449 8. ChemE-as-Code open-source platform letting LLM agents do chemical engineering calcs, process simulation, plant layout, HAZOP, and life-cycle assessment through checkable tools • preliminary ammonia engineering package for $224 in API costs • calculations run outside the LLM, with provenance tracking and automated consistency checks arxiv.org/abs/2611.07947 9. GuidedWAM training-free guidance for driving world action models. it injects differentiable objectives during sampling, so you skip retraining for every new objective • tested on three frozen, public world action models • +1.8 to 3.6 EPDMS on the NAVSIM v2 hard split (avg over 3 seeds) • under 10% extra sampling time per step arxiv.org/abs/2611.07912 10. Wave Forcing long-video generation method that co-designs training, inference, and execution for better temporal consistency and throughput • on Wan2.1-T2V-1.3B, quality comparable to Rolling Forcing with 5 denoising steps • 134.8 end-to-end FPS on 8 H200s (4-step) • up to 2.53× throughput over matched context-parallel baselines (DiT-only tests) arxiv.org/abs/2611.00859 11. Persistent Retrieval for Conversational Agents retrieval subagents that keep their history across multi-turn conversations and hand evidence straight to the main agent • 43.3% fewer agent-model tokens vs skill-based retrieval • 45% lower estimated API costs • 9.1% faster average response time arxiv.org/abs/2611.07899 the pattern this week: AI is getting really good at the optimization loop first. kernels, program search, sales playbooks, retrieval if part of your job is "try something, measure it, try again", that's the part to start learning to automate yourself
2
1
6
515
CAVEATS SO THERE IS NO FEAR-MONGERING: 1. results are company-reported what it automates: parts of GPU kernel optimization and inference performance engineering 2. benchmark-specific. promising evidence for automated research, and scientists are still very much needed what it automates: parts of ML research (paper reproduction, experimental iteration, method improvement) 3. a concrete program-search result. general scientific autonomy is a much bigger claim what it automates: parts of computational algorithm search and optimization 4. t's an eval benchmark. stronger evidence than unverified math claims, and theoretical computer scientists are still essential what it automates: routine formal proof assistance and verified mathematical construction 5. one company's sales environment. still one of the strongest real-world deployments on this list what it automates: parts of sales outreach, follow-up, and playbook optimization 6. these are validated exploit demos and patch candidates. maintainer-approved merged fixes are a separate step. controlled campaign what it automates: parts of vulnerability discovery, reproduction, and patch development 7. simulated task suites so far what it automates: parts of visual planning and trajectory prediction engineering 8. the authors say the design has yet to be independently reviewed by an engineer what it automates: parts of chemical engineering calculations, simulation setup, and documentation 9. a technical improvement to driving models. driving engineers stay in the loop what it automates: parts of autonomous-driving model tuning and control optimization 10. meaningful systems research. video production pros are still very much in the picture what it automates: parts of video-generation systems engineering and optimization 11. vs a non-persistent retrieval subagent it's faster but uses more tokens and costs more what it automates: parts of retrieval pipeline design and maintenance for conversational agents
2
77
you can run real physics simulators on a laptop in 2026! here are the 10 i'd actually use for AI research and world models, ranked by how good they are for experimenting: 1. Newton: robotics physics backed by DeepMind, Disney Research and NVIDIA 2. MuJoCo: best overall, easiest to get running 3. NVIDIA Warp: write your own physics kernels in Python 4. PhiFlow: differentiable fluid sims you can train a neural net through 5. Genesis: multi-physics engine built for embodied AI 6. Taichi: for inventing new physics (granular, elastic, MPM) 7. PhysicsNeMo: PINNs, neural operators, surrogate models 8. DeepXDE: neural nets that solve differential equations 9. JAX MD: differentiable molecular dynamics 10. OpenFOAM: serious engineering CFD, steepest setup the one to install fastest (and the two i'd build research on) are next👇
2
4
272
if you only try one now, make it MuJoCo pip install mujoco, paste a 15 line script, and you're watching a sphere fall under gravity in a 3D viewer within 5 minutes for real research i'd pick Genesis + Warp. Genesis lets AI agents act inside simulated worlds, Warp lets you invent your own solver one caveat: running a simulator on a laptop is easy. training a big learned physics model still wants a real GPU or cloud compute
1
70
I think I found the new best method for ideation @ambitionlabsinc just getting started, more soon
i’ve been hacking on the 6th idea from here, I found it pretty interesting want a demo? let me know! it’s in the works, but i’d love to know more people into making faster and cheaper physics-based simulations
3
111
i’ve been hacking on the 6th idea from here, I found it pretty interesting want a demo? let me know! it’s in the works, but i’d love to know more people into making faster and cheaper physics-based simulations
10 things worth building with megakernels right now, and each one pushes real research forward!: 1. a real time world model - feed in frames + action inputs, predict the next latent states - hyperkernel fuses the rollout loop to cut inference latency 2. a genomics research assistant - visualize embeddings from a DNA language model - compare variant effect predictions across sequences - one interface for researchers to probe their results 3. a GPU-accelerated protein structure explorer - run ESMFold predictions in the browser with optimized inference - compare predicted structures and confidence scores side by side - an interactive tool for exploring how proteins fold 4. an autonomous driving simulator - run a vision model on simulated road frames, predict vehicle actions - hyperkernel speeds up the perception pipeline - visualize predicted trajectories and decisions in real time 5. an AI agent with long-term memory - an agent that remembers its research across multiple days - hyperkernel fuses memory-related ops to make local inference faster - the memory architecture is where the real research lives 6. a learned physics id dynamics and soft-body interactions with learned models - benchmark against a conventional physics engine - hyperkernel accelerates the learned approximation 7. a scientific discovery platform - pick a model and run simulations for any scientific question - analyze the results and generate a research report - hyperent inference backend 8. a real-time voice assistant - streaming speech recognition and generation with natural interruptions - fully local inference at very low latency - hyperkernel's moshi work is a reference for fast speech-model execution 9. a molecular simulation explorer - interactive visualization of molecular behavior - ML-based predictions with GPU acceleration - one platform for computational biology research 10. a mini AI inference platform - open-source serving system with model selection and performance profiling - hyperkernel provides the high-performance execution layer
1
1
3
370
10 things worth building with megakernels right now, and each one pushes real research forward!: 1. a real time world model - feed in frames + action inputs, predict the next latent states - hyperkernel fuses the rollout loop to cut inference latency 2. a genomics research assistant - visualize embeddings from a DNA language model - compare variant effect predictions across sequences - one interface for researchers to probe their results 3. a GPU-accelerated protein structure explorer - run ESMFold predictions in the browser with optimized inference - compare predicted structures and confidence scores side by side - an interactive tool for exploring how proteins fold 4. an autonomous driving simulator - run a vision model on simulated road frames, predict vehicle actions - hyperkernel speeds up the perception pipeline - visualize predicted trajectories and decisions in real time 5. an AI agent with long-term memory - an agent that remembers its research across multiple days - hyperkernel fuses memory-related ops to make local inference faster - the memory architecture is where the real research lives 6. a learned physics id dynamics and soft-body interactions with learned models - benchmark against a conventional physics engine - hyperkernel accelerates the learned approximation 7. a scientific discovery platform - pick a model and run simulations for any scientific question - analyze the results and generate a research report - hyperent inference backend 8. a real-time voice assistant - streaming speech recognition and generation with natural interruptions - fully local inference at very low latency - hyperkernel's moshi work is a reference for fast speech-model execution 9. a molecular simulation explorer - interactive visualization of molecular behavior - ML-based predictions with GPU acceleration - one platform for computational biology research 10. a mini AI inference platform - open-source serving system with model selection and performance profiling - hyperkernel provides the high-performance execution layer
We invented a recursive megakernel generator called Hyperkernel, the most complete solution to megakernels and GPU efficiency. Hyperkernel-generated megakernels achieve state of the art performance on 11 different domains: inference (DeepSeek, Kimi and GLM), pretraining (Qwen), hardware fault recovery, robotics (π₀-FAST, ACT) diffusion and real time world models (Matrix-Game), full self driving (OpenPilot), voice models (Moshi), genomics (Evo 2) and protein folding (ESMFold). All of this in a single kernel launch. For context, Hyperkernel generated fully optimized megakernels for robotics, genomics and voice in under 50 minutes, and doubled the efficiency for >5 models generally using only about 60-100 lines of code per megakernel. This speaks to the unparalleled standard for code quality in the automated generations. Our first release today open sources the megakernels themselves. Over the next couple weeks, we will release the entire Hyperkernel stack, along with the complete serving stack for Kimi and GLM (20-30% inference efficiency gains across all concurrencies). Out of the box GPU machinery performs terribly for unorthodox research. We built Hyperkernel only 20 days ago for internal use; but given its multi-modality, we have decided to open source it. Creative research should not be bottlenecked by the lack of dedicated infra teams. Hyperkernels offer a one stop shop infra stack for all research. We would love for the open source community to add to it :) GitHub repo: github.com/Hyperstitioncc/Hy… NOTE: We primarily use NVIDIA B300s, but our kernel generation extends to other hardware (B200s, AMD, TPUs). 1/9 We absolutely crush NVIDIA and achieve the best low latency inference for DeepSeek v4.1 Flash. Compared with NVIDIA’s published vLLM results, Hyperkernel delivers: • 36.7% higher per-user generation speed at concurrency 1. • 28.2% higher per-user speed while serving two users, versus NVIDIA serving one. • 15% higher per user speed while serving four users, versus NVIDIA serving one. • 90.6% higher per-user generation speed at concurrency 4, compared with the highest published C4 AgentX interactivity for TP4 B300s. • 28.5% higher per-user generation speed at concurrency 8, compared with NVIDIA's published result. • 36% faster median per-request decode speed.
3
1
1
18
2,348