Maxen @ignicoder

ByteDance just published something I've been waiting for someone to build: CUDA Agent! It trained a model that writes fast CUDA kernels. Not just correct ones — actually optimized ones. It beats torch.compile by 2× on simple/medium kernels, ~92% on complex ones, and even outperforms Claude Opus 4.5 and Gemini 3 Pro by ~40% on the hardest setting. The key idea is simple but kind of brilliant: CUDA performance isn’t about correctness, it’s about hardware. Warps, memory bandwidth, bank conflicts — the stuff you only see in a profiler. So instead of rewarding “did it compile?”, they reward actual GPU speed. Real profiling numbers. RL trained directly on performance. That’s a big shift. Paper: arxiv.org/abs/2602.24286 Project: cuda-agent.github.io

keyboard_arrow_left Previous keyboard_arrow_right Next

@tonykipkemboi 100% agree Agents frameworks are getting squeezed by the whole market setting up easy to integrate agents and infrastructure

@zackbshapiro I liked the text, but I'll have to disagree in parts about the judgment premium logic. Years of experience and taste are valuable, but in an economy where AI can read/write and do things most humans can, I think that the taste will be something that will evolve along with it

🚨BREAKING: The open-source project that's saving enterprises millions in GPU costs. LMCache: 6.9K stars. 124K downloads/month. Used by Google Cloud, CoreWeave, and NVIDIA . Here's the problem nobody talks about: Your LLM inference is burning 50% of its compute on work it has already done. Every time a user sends a message, your model recomputes the entire context from scratch. The system prompt. The conversation history. The retrieved documents. That's thousands of GPU cycles. Wasted. Per request. LMCache makes the KV cache persistent and shareable across engine instances . Instead of the cache dying when a request finishes, LMCache offloads it to a shared layer: → GPU (fastest) → CPU DRAM (fast, cheaper) → Local disk (slower, cheap) → S3/Network (slowest, cheapest) This unlocks architecture patterns that were previously impossible: 1./ Instant RAG Process a 100-page PDF once. Store the KV cache. Now any user query against that doc starts instantly. Zero Time-To-First-Token. 2./ Disaggregated Prefill Run heavy "Prefill" on H100s. Stream the cache to cheaper L4s for "Decoding." Your expensive GPUs never sit idle. 3./ Context Sharing Multiple users asking about the same context? Compute it once, serve everyone. The results: 🚀3-10x delay savings in multi-round QA workloads ⚡️Dramatic reduction in Time-To-First-Token (TTFT) 💰 GPU cycle reduction Integrates directly with vLLM (72K stars) and SGLang (24K stars) . Adopted by: → GMI Cloud → Google Cloud → CoreWeave → Redis → NVIDIA Dynamo Apache-2.0 license. pip install lmcache. Stop letting your GPUs do the same homework twice. GitHub Repo in comments

@theantijobboard Consider remote ?

@zackbshapiro @matthew_pines Already applied and hoping that they can consider remote. Because is exactly the type of work that I’m interested

@matthew_pines Applied

We’re excited to release TorchLean which is the first fully verified neural network framework in Lean. The Lean community has largely focused on pure mathematics. TorchLean expands this frontier toward verified neural network software and scientific computing. With the recent release of CSlib, we see this as another step toward a fully verified ML stack. We support features: 1. Executable IEEE-754 floating-point semantics (and extensible alternative FP models) verified tensor abstractions with precise shape/indexing semantics 2. Formally verified autograd system for differentiation of NN programs Proof-checked certification / verification algorithms like CROWN (robustness, bounds, etc.) 3. PyTorch-inspired modeling API with eager-style development + export/lowering to a shared IR for execution and verification Project page: leandojo.org/torchlean.html Paper: [2602.22631] TorchLean: Formalizing Neural Networks in Lean Work done @Robertljg, Jennifer Cruden, Xiangru Zhong, @huan_zhang12 and @AnimaAnandkumar. #MachineLearning #ScientificComputing #Lean

@theantijobboard Interested

@m_heilb @MPI_NL No worries, thanks

@AndreaShuyuWang Interested. Accept remote?

@jvrsanch Accept remote?

This is counterintuitive for some, which is why there’s a paradox named after it. But if you lower the cost of something that was previously supply constrained, demand for that thing goes up. Software engineering is just one of the easiest examples to contemplate. The process goes like this: every small business, every IT team, every large enterprise sees that engineering can now drive vastly more output. They then start to consider all the new things they can build or automate. They even test building prototypes themselves. They only get so far with that approach because they realize there are still 50 other tasks that go into building software and maintaining it. So they start to hire more engineers to do that work. All of this for work they never would have considered automating or having software for if AI didn’t exist. So yes, automating tasks, in plenty of fields, will lead to demand for experts, not less.

Puru Saxena @saxena_puru

4 months ago

The software industry is apparently dying but job postings for software engineers are rapidly rising!

171 191 2K 939K 344

We open sourced an operating system for ai agents 137k lines of rust, MIT licensed we love @openclaw and it inspired a lot of what we built. but we wanted something that works at the kernel level so we built @openfangg agents run inside WASM sandboxes the same way processes run on linux. the kernel schedules them, isolates them, meters their resources, and kills them if they go rogue. it has 16 security layers baked into the core. WASM sandboxing, merkle hash-chain audit trails, taint tracking on secrets, signed agent manifests, prompt injection detection, SSRF protection, and more. every layer works independently. giving an LLM tools with zero isolation is insane and we're not doing it. we also created something called Hands. right now every ai agent is a chatbot that waits for you to type. Hands are different. you activate one and it runs on a schedule, 24/7, no prompting needed. your Lead Hand finds and scores prospects every morning and delivers them to your telegram before you wake up. your Researcher Hand writes cited reports while you sleep. your Collector Hand monitors targets and builds knowledge graphs continuously. they work for you. you don't babysit them github.com/RightNow-AI/op… ⭐

@sarahookr @Adaption Applied to Adaption, really excited about it

@craigcitro That's perfect for me

how to build a bootstrapped startup without funding: 1. pick a problem you personally have. if you don't use your own product daily, quit now 2. skip the pitch deck. open your code editor. ship something ugly in a weekend 3. charge money from day 1. free users give you nothing but support tickets 4. use boring tech. PHP, SQLite, vanilla JS. frameworks are a trap that mass waste your time 5. host on cheap VPS ($5-20/mo). not AWS. you don't need kubernetes for 1,000 users 6. do customer support yourself. it's the fastest product feedback loop that exists 7. automate everything you do more than twice. cron jobs > employees. 8. grow on Twitter/X by building in public. your journey IS the marketing 9. keep your burn rate near zero so you never need to raise. ramen profitable > series A 10. say no to investors, cofounders, and "advisors" who want equity for intros i've been doing this for 10+ years now. no employees, no funding, no board meetings the entire VC game is designed to make you think you need permission to start you don't

@bepituLaz Maybe elixir and rust ?

@ericzelikman Totally agreed, it's pretty much giving up the good to open space for the great