keval shah

@chrisbarber @PrimeIntellect missing

I asked people which companies have the highest density of talented people they know: 1st. Cognition - 9 votes Equal 1st. Anthropic - 9 votes 2nd. Modal - 5 votes 3rd. OpenAI - 4 votes 4th. Standard Intelligence - 3 votes 4th. Cursor - 3 votes Two votes each: - Ramp - Flapping Airplanes - DeepMind - Long Lake - Applied Compute One vote each: - SpaceXAI - SpaceX (treated separately, one vote was for the AI lab subsidiary and one was for the rocket team) - American Terawatt - Mechanize - Olix - Fluidstack - Chai Discovery - Sail Research - Etched - Core Automation - Specter - Clay - Applied Intuition - Sierra - Hivemind - Bitrig - Retro - Thinking Machines - Decagon - Precigenetics - Pangram - Reflect - Thrive Holdings - Adaption

@ivan_bezdomny Google released a new model, AlphaEvolve. x.com/googlecloud/st…

Glad to share what I've been building in public. A lightweight AI news feed. We also finetuned a model to write new stories better than Claude or GPT with extensive prompting (believe me not for a lack of effort). Posting coming about that shortly.

@demishassabis @pushmeet

Exciting to see AlphaEvolve hit GA! In our early access work at Pebble, we used it to tackle the multi-dimensional challenge of GPU inference serving. By discovering performance modeling formulations, AlphaEvolve delivered a 56% relative reduction in inference perf model errors. Thrilled to leverage this to continuously map emerging hardware without manual tuning! @GoogleDeepMind @googlecloud @NVIDIAAI

@gpusteve @TheZachMueller

does anyone know why lambda does this with the u e and o?

Great to sit down with @JordanNanos at GTC 26' and talk about perf/watt, AI Infra and power bottleneck. youtu.be/OpL-Q0YAKJU?si… @SemiAnalysis_ @readsail

Recipes to run disaggregated inference on AMD MI355x GPUS? What's the preferred serving engine vllm, sglang or others?

@dwarkesh_sp John List, Stiglitz, Daron Acemoglu

Who should I interview on my podcast? Open to more AI, but also to random history/econ/etc professors that I might not have heard of before.

Great to sit down for a convo with @readsail at @NVIDIAGTC '26 on the future of AI infrastructure, Energy efficiency and tokens/watt optimizations. @NVIDIAAI

@zeeshanp_ Can you share more details about the melting of the power smoothing feature encountered on blackwell GPUs by the Frontier labs?

scaling up on blackwell gpus is much harder than people think. rewriting the entire modeling stack to utilize hardware efficiently and correctly is not an easy feat. simple mistakes in hw/sw codesign can create major issues at the hardware level. one example of this is the power smoothing feature on blackwell. the component on chip that is responsible for power smoothing will eventually melt after a few months of usage, and this can cause a lot of transient issues during large scale training. one of the frontier labs had to learn this the hard way.

@akseljoonas @huggingface Is there a way to view detailed logs to monitor?

Introducing ml-intern, the agent that just automated the post-training team @huggingface It's an open-source implementation of the real research loop that our ML researchers do every day. You give it a prompt, it researches papers, goes through citations, implements ideas in GPU sandboxes, iterates and builds deeply research-backed models for any use case. All built on the Hugging Face ecosystem. It can pull off crazy things: We made it train the best model for scientific reasoning. It went through citations from the official benchmark paper. Found OpenScience and NemoTron-CrossThink, added 7 difficulty-filtered dataset variants from ARC/SciQ/MMLU, and ran 12 SFT runs on Qwen3-1.7B. This pushed the score 10% → 32% on GPQA in under 10h. Claude Code's best: 22.99%. In healthcare settings it inspected available datasets, concluded they were too low quality, and wrote a script to generate 1100 synthetic data points from scratch for emergencies, hedging, multilingual etc. Then upsampled 50x for training. Beat Codex on HealthBench by 60%. For competitive mathematics, it wrote a full GRPO script, launched training with A100 GPUs on hf.co/spaces, watched rewards claim and then collapse, and ran ablations until it succeeded. All fully backed by papers, autonomously. How it works? ml-intern makes full use of the HF ecosystem: - finds papers on arxiv and hf.co/papers, reads them fully, walks citation graphs, pulls datasets referenced in methodology sections and on hf.co/datasets - browses the Hub, reads recent docs, inspects datasets and reformats them before training so it doesn't waste GPU hours on bad data - launches training jobs on HF Jobs if no local GPUs are available, monitors runs, reads its own eval outputs, diagnoses failures, retrains ml-intern deeply embodies how researchers work and think. It knows how data should look like and what good models feel like. Releasing it today as a CLI and a web app you can use from your phone/desktop. CLI: github.com/huggingface/ml… Web + mobile: huggingface.co/spaces/smolage… And the best part? We also provisioned 1k$ GPU resources and Anthropic credits for the quickest among you to use.

@art_zucker @IlysMoutawwakil What profiling tool is this @art_zucker

Today I re-iterate: I hate MoEs and we are wasting time on them.... Let's unite and call a global ban on MoEs please. Please 1M+ salary researchers: do better... credits to @IlysMoutawwakil for the graph:

Why are NCCL errors so hard to debug? [PG ID 2 PG GUID 3 Rank 0/1/2/3] Process group watchdog thread terminated

@pritopian Mylapore is expanding to SF.

there are no good South Indian restaurants in SF. Have to drive down to madras cafe in sunnyvale every time. 😒

@jtatarchuk @dylan522p @JayDawani @AnushElangovan @picocreator great panel to close out an amazing Beyond summit!

Hot Takes panel at Beyond with my homies @dylan522p, @JayDawani, @AnushElangovan, @picocreator. This could have gone on another 2hrs. 🔥🔥🔥

@BenBajarin @mwilcox What exactly does he mean by horizontally open?

Only two companies in tech are so deeply vertically integrated. Apple and NVIDIA. Jensen says they are the only one fully vertical and horizontally open. 🤔

@charles_irl Please save some for me! I'll swing by tomorrow.

we've got printed GPU Glossaries and a lot of swag(ger)

Specialized agents collect telemetry, detect anomalies, and interact with the K8s API, while a Nemotron orchestrator synthesizes it all into actionable recommendations - running as a k8s CronJob every 5 mins. github.com/kevalshah90/ne… #NVIDIAGTC @ctnzr @Baxate @NaderLikeLadder

Built an k8s cluster agent powered by NVIDIA's Nemotron LLM that uses multi-agent architecture to monitor, analyze, and detect anomalies in GPU workloads in real time.