Nvidia, every mention
103 scenes (2025), the whole family · ← back to Nvidia
tap a year for its mentions
every year 2025 anyone Shawn Wang 80Dylan Patel 55Kyle Kranen 43Ali Taha 26Ethan He 24Chris Lattner 24George Hotz 23Philip Kiely 20Doug O'Laughlin 19Sarah Chieng 17
Verbatim, from the transcripts: the passages where Nvidia comes up
[State of AI Papers 2025] Fixing Research with Social Signals, OCR & Implementation — Team AlphaXiv
- ▶ 5:18 unnamed speaker If you host it on your own A-One-Hundreds, and just like, you, you batch things properly, probably Deep Seek is best bang for your buck.
- ▶ 32:35 unnamed speaker Specifically, if you have anything really involving GPUs, I would call it launchables from NVIDIA.
- ▶ 32:35 unnamed speaker Specifically, if you have anything really involving GPUs, I would call it launchables from NVIDIA.
[NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton
- ▶ 24:19 Kevin Wang The nice thing is that all of our experiments, even the thousand layer networks, can be run on one single, 80 gigabyte, each 100 GPU.
[State of Context Engineering] Agentic RAG, Context Rot, MCP, Subagents — Nina Lopatina, Contextual
- ▶ 0:32 Nina Lopatina Turing and NVIDIA. 4 times in the scene
[State of RL/Reasoning] IMO/IOI Gold, OpenAI o3/GPT-5, and Cursor Composer — Ashvin Nair, Cursor
- ▶ 30:51 Ashvin Nair It was like, Deep Seek shows that NVIDIA chips are actually more useful than previously thought, and, like, NVIDIA's stock, like, goes down a bunch. 4 times in the scene
Steve Yegge's Vibe Coding Manifesto: Why Claude Code Isn't It & What Comes After the IDE
- ▶ 1:44 Steve Yegge Now you saw, I don't know if you saw Jordan Hubbard's post from Nvidia where he just laid out some really nice advice on how to get the most out of agents as you're coding.
- ▶ 16:16 Steve Yegge He's the one that wrote the article that crashed the stock market about Nvidia.
SAM 3: The Eyes for AI — Nikhila & Pengchuan (Meta Superintelligence), ft. Joseph Nelson (Roboflow)
- ▶ 4:50 Joseph Nelson Over 30 frames per second, for example, on like a small T four, or excuse me, small like edge device and hundreds of frames per second on like a T four. 2 times in the scene
- ▶ 9:28 unnamed speaker If I want a hundred detected objects on an H 200, obviously this is an H 200, but it's also like, this is impressively fast. 2 times in the scene
- ▶ 10:07 unnamed speaker So, so I'm reading in the paper, it's, uh, 10 objects on two HCOs, 28 on four HCOs, and 64 on eight HCOs, something like that. 3 times in the scene
⚡️Jailbreaking AGI: Pliny the Liberator & John V on Red Teaming, BT6, and the Future of AI Security
World Models & General Intuition: Khosla's largest bet since LLMs & OpenAI
- ▶ 29:07 Pim de Witte So even if you're, for instance, uh, simulating human behavior in Omniverse, because you're trying to create better training data for factory floors, um, you can use it.
After LLMs: Spatial Intelligence and World Models — Fei-Fei Li & Justin Johnson, World Labs
- ▶ 12:50 unnamed speaker Where, let's just say, you know, Nvidia has won, and we should just, you know, scale that out in infinity and write software to patch up any, any gaps we have in the, in the mix, right?
- ▶ 13:01 Justin Johnson Like, if you look at, if you look at the numbers, like, even going from Hopper to Blackwell, like, the performance per watt is about the same. 2 times in the scene
- ▶ 13:01 Justin Johnson Like, if you look at, if you look at the numbers, like, even going from Hopper to Blackwell, like, the performance per watt is about the same. 2 times in the scene
Anthropic, Glean & OpenRouter: How AI Moats Are Built with Deedy Das of Menlo Ventures
- ▶ 1:06:00 Deedy Das Kind of NVIDIA.
- ▶ 1:11:20 Alessio Fanelli And so it's like, okay, well, the amount of money being spent in this space is large enough to justify betting, like the same way Instagram was like one percent of Facebook market cap.
⚡️ The State of AI Engineer Hiring: Cheating, AI Adoption,Junior Devs — Vivek Ravisankar, HackerRank
- ▶ 4:56 Vivek Ravisankar There's everything from OpenAI, NVIDIA, Amazon, Salesforce, all of these, we work with all of the, all of the customers.
How Zyphra went all-in on AMD + Why Devs feel faster with AI but are slower — with Quentin Anthony
- ▶ 3:19 Quentin Anthony Um, we found that it's, it's great, uh, for flash attention to specifically, we were able to be H-one hundred. 2 times in the scene
- ▶ 4:32 Quentin Anthony Like, I would say MI-DX was not on the same level of eight 100.
- ▶ 4:39 Quentin Anthony So you basically had to split across, you know, you had a three level, uh, parallelism scheme instead of two level on NVIDIA, where you just had with within the node and across nodes.
- ▶ 5:32 unnamed speaker You basically have this kind of like these groups trying to make better software to make AMD just the same as Nvidia. 3 times in the scene
- ▶ 8:42 unnamed speaker Because I think the other question is, like, well, I'm gonna do all this work versus, like, I just write CUDA code that then, when the BG-G-Hundred comes online for my cluster, I'll just switch it over right away.
- ▶ 10:40 Quentin Anthony The ecosystems within NVIDIA and AMD are actually quite similar. 4 times in the scene
- ▶ 19:28 Quentin Anthony It doesn't necessarily have to be, they compare it to what NVIDIA does well, and they say, okay, well, it doesn't do that well,
⚡ Open Model Pretraining Masterclass — Elie Bakouch, HuggingFace SmolLM 3, FineWeb, FinePDF
- ▶ 2:50 Elie Bakouch Like, if you take this curve, uh, like, when it's mixed with, uh, with, like, uh, other, other web datasets, not PF-one, it basically has a very, very, very good performance compared to this, like, NemoTrans SCP-V-II, which is a very…
Why RL Won — Kyle Corbitt, OpenPipe (acq. CoreWeave)
- ▶ 48:53 Shawn Wang No, like I, I, I literally like after last week, I think maybe two weeks ago with the whole Oracle NVIDIA
Building Jamba 3B: the tiny Hybrid Transformer State Space Reasoning Model - Barak Lenz, CTO of AI21
- ▶ 5:27 Barak Lenz So we designed J to have a version that fits on a single GPU, a single AY 100 or H one, 80 gigabytes.
- ▶ 5:27 Barak Lenz So we designed J to have a version that fits on a single GPU, a single AY 100 or H one, 80 gigabytes.
- ▶ 9:37 Barak Lenz So I think NVIDIA released hybrid models and other companies started following.
- ▶ 13:35 Barak Lenz Jamba is that it starts with, with a large size, you know, Jamba mini is mini for enterprises, but it's not mini for the, for a developer that, you know, has a T four, has his own GPU and he wants to try stuff.
⚡️Raising $1.1b to build the fastest LLM Chips on Earth — Andrew Feldman, Cerebras
- ▶ 3:09 Andrew Feldman You think 20 times faster than Nvidia B 200 GPUs and it's, it's been an amazing run.
- ▶ 8:53 unnamed speaker You know, I think, like, what has been hard for a non-hardware person like myself to understand is, you know, like, a lot of people who are, who are sort of competing with NVIDIA and betting on, like, more on, uh, on, uh, on-chip memory… 3 times in the scene
- ▶ 26:06 Andrew Feldman I, uh, I, I think that, you know, 10 years ago, uh, NVIDIA was a twenty billion dollar company.
A Technical History of Generative Media
- ▶ 11:23 Batuhan Taskaya Right now, it's, like, much more competitive space, but, like, NVDA has, like, a fifty-percent, hundred-percent kernel team that's writing kernels.
- ▶ 15:25 Batuhan Taskaya Uh, when you use PyTorch with B-Torch's Blackwell chips, you're not getting the best performance.
- ▶ 22:50 Batuhan Taskaya And we, like, Kubernetes version at Google Cloud was fine in 2022 when we wanted to get eight A-one-hundreds. 2 times in the scene
- ▶ 23:15 Batuhan Taskaya And in, in this world, like, we had to build our orchestration layer, we had to build our own distributed file system, we had to build our own container runtimes, all, all the stack to make sure that the cold starts are extremely,…
- ▶ 23:57 unnamed speaker You keep mentioning H-one hundreds. 2 times in the scene
- ▶ 24:03 Batuhan Taskaya Blackwell is, is obvious, like, we, we have, like, we have five people dedicated to writing Blackwell kernels right now to make sure we can, like, because theoretically it looks good, right? 6 times in the scene
- ▶ 24:15 Batuhan Taskaya So we have a dedicated team that's, like, working with NVIDIA directly to write custom kernels for Blackwell for diffusion transformers to get to the, get to the point where it makes perf dollar make sense, and then, then we would start… 4 times in the scene
- ▶ 25:35 Batuhan Taskaya Uh, and like some of the, like, like B-Tree hundreds are gonna have, like, a better softmax instruction that gets, like, 1.5 X, whatever.
Better Data is All You Need — Ari Morcos, Datology
- ▶ 10:29 Alessio Fanelli NVIDIA is like four trillion, and SKIL is not four trillion, so what do you think there's that inefficiency?
⚡️Accelerators @ 3x NVIDIA H200 perf, Made in the USA - Thomas Sohmers + Mitesh Agrawal, Positron AI
- ▶ 7:46 Alessio Fanelli And maybe contrapose that both to, you know, NVIDIA and the kind of traditional ones, as well as Grok, obviously you work there, Cerebras, and then there's kind of like the long tail of all the other GPU alternatives things, but just give…
- ▶ 12:14 Shawn Wang I mean, just because most people are familiar with Nvidia H series. 6 times in the scene
- ▶ 12:14 Shawn Wang I mean, just because most people are familiar with Nvidia H series.
- ▶ 14:44 Thomas Sohmers And the A 100 comparison here is interesting because in most of these cases, they're actually, they,
- ▶ 14:53 Thomas Sohmers Percentage of theoretical memory bandwidth actually has gotten worse generation over generation, and we'll see exactly where Blackwell ends up, but all indications are, even though they, you know, more than doubled the theoretical memory…
- ▶ 14:53 Thomas Sohmers Percentage of theoretical memory bandwidth actually has gotten worse generation over generation, and we'll see exactly where Blackwell ends up, but all indications are, even though they, you know, more than doubled the theoretical memory…
- ▶ 17:59 Mitesh Agrawal Look, from a marketing term, we call it like falling within the NVIDIA ecosystem, but like from both of our past experiences, I can tell you that most of the, the silicon providers, you know, whatever technology they might come up with or,… 6 times in the scene
- ▶ 27:29 Thomas Sohmers So, but, in, NVIDIA's TF-thirty-two number format is a nineteen-bit number format. 2 times in the scene
- ▶ 34:06 Thomas Sohmers Like I would say just in my talking with, um, you know, the, the founders and people in, in all of the different semiconductor, you know, startups trying to go after NVIDIA. 6 times in the scene
- ▶ 38:59 Mitesh Agrawal So, like, like NVIDIA, we want to sell systems. 3 times in the scene
- ▶ 41:22 Thomas Sohmers Best relative advantage over NVIDIA is on that generation or decode side of it.
⚡️Mercury: Ultra-Fast Diffusion LLMs — Estefano Ermon, CEO Inception Labs
- ▶ 25:19 Stefano Ermon And so you can find something on, in the literature as well from NVIDIA, from academic groups.
The RLVR Revolution — with Nathan Lambert (AI2, Interconnects.ai)
- ▶ 18:44 Nathan Lambert I think the NVIDIA Lama-Nimetron reasoning paper is probably the
🕰️ The Oral History of Windsurf (ft. Varun Mohan, Scott Wu, Jeff Wang, Kevin Hou, Anshul R)
- ▶ 45:27 Varun Mohan So I guess for, for training, you're right in that it is actually nuts to think about how insane the networks are for Nvidia's most recent hardware.
- ▶ 45:36 Varun Mohan It's like for the H 100 boxes, you shove eight of these H 100 on a machine between two nodes. 3 times in the scene
The Shape of Compute (Chris Lattner of Modular)
- ▶ 2:23 Chris Lattner And so we need to be state of the art on NVIDIA GPUs meeting and beating NVIDIA's best on things like a Lana three model, which by the way is serving end to end, like very high bar, by the way, this is like 2 times in the scene
- ▶ 4:08 Chris Lattner It ran just on a 100, just one model, but it had state of the art performance. 2 times in the scene
- ▶ 5:03 Chris Lattner Let's, oh yeah, let's add H 100 support. 2 times in the scene
- ▶ 5:27 Chris Lattner And as you do that in Blackwell, like all this stuff is like all now in the product.
- ▶ 9:05 Chris Lattner NVIDIA's got hundreds or thousands of people working on it.
- ▶ 11:34 Chris Lattner So it's not as good as something like VLLM because it's missing some features, and it only supports NVIDIA and AMD hardware, for example. 4 times in the scene
- ▶ 14:28 Chris Lattner it turns out that, uh, an H-one hundred and AMD chip are actually quite different.
- ▶ 24:38 Chris Lattner We support Google TPUs and Infantria and AMD and NVIDIA, obviously, and CPUs and this and that and the other thing. 3 times in the scene
- ▶ 30:51 Chris Lattner And so you can go look at how we brought up H-one hundred, built flash attention from scratch in a few weeks, built like all the stuff. 2 times in the scene
- ▶ 41:47 Chris Lattner So the max framework and the mojo language, free to use on NVIDIA and CPUs, any scale, go nuts, do whatever you want. 2 times in the scene
- ▶ 53:43 Chris Lattner But then the world had a big wake up call and video stock price went down and all that stuff like a month later. 2 times in the scene
- ▶ 56:32 Chris Lattner Go look at VLM.
- ▶ 56:32 Chris Lattner Go look at VLM.
Voice AI Masterclass — Kwindla Hultman Kramer and swyx
- ▶ 3:32 Kwindla Hultman Kramer We'll have people from open AI and from deep mind and from Nvidia and from AWS doing office hours.
- ▶ 8:07 Kwindla Hultman Kramer And then Parakeet is Nvidia's new speech model, speech transcription model. 4 times in the scene
The Rise and Fall of the Vector DB category: Jo Kristian Bergum (ex-Chief Scientist, Vespa)
- ▶ 24:16 unnamed speaker Um, and then I think over here, Voyage just got acquired by NVIDIA.
SF Compute: Commoditizing Compute
- ▶ 6:28 Evan Conrad My intuition is that the hyperscalers are probably going to lose a lot of money, and they know they're going to lose a lot of money, um, on reselling NVIDIA GPUs at least. 2 times in the scene
- ▶ 14:08 Michael Swix (Swyx) So why didn't NVIDIA or Microsoft, both of which have more money than CoreWeave, do CoreWeave, right? 10 times in the scene
- ▶ 22:06 Evan Conrad We just, like, assumed we could go to, like, Lambda, um, or something, and, like, buy thousands of, at the time, A-One-Hundreds.
- ▶ 24:58 Evan Conrad Like you can go on SF compute today and you can get thousands of H 100 for an hour if you want.
- ▶ 27:59 Michael Swix (Swyx) One of our top pieces from last year was talking about the H-one hundred glut from all the, uh, long-term contracts that were not being fully utilized and being put under the market.
- ▶ 31:53 Evan Conrad Um, lots of bio and pharma, um, was using, um, H 100 training sort of the bio models of sorts. 3 times in the scene
- ▶ 50:08 Evan Conrad It's like an NVIDIA reference version of this. 2 times in the scene
Claude Plays Pokémon Hackathon: Escape from Mt. Moon!
- ▶ 1:30 David Hershey Uh, so I ripped off the Voyager agent, which is, like, this Nvidia paper that looked really cool.
Building Manus AI (first ever Manus Meetup)
- ▶ 35:26 unnamed speaker Yeah, you can just, uh, find the stock data for NVIDIA, yeah, stock profiles, and so, a real price for the, a real hard price, and also, the Twitter and the LinkedIn, uh, search, search, yeah,
Outlasting Noam Shazeer, Crowdsourcing Chai AI w/ 1.4m DAU — with William Beauchamp, Chai Research
- ▶ 56:30 William Beauchamp What the money let them do was, if they wanted to fine-tune Alarma-seventyb on eight H-one-hundreds overnight, if you give them money, then they can do it.
DeepSeek V3, SGLang, and the state of Open Model Inference in 2025 (Quantization, MoEs, Pricing)
- ▶ 2:46 Yining Zhang You need, I think, uh, 671 gigabytes for the weights, and you, you also need an extra memory for the KV cache, so it's not possible to run that on H-one hundred.
- ▶ 3:00 Yining Zhang That's why we, we, we chose H-h-two hundred to run that model, or use SmartyNode to run that model.
- ▶ 5:06 Yining Zhang So big model that we should use H 200 or use H 100 multi nodes. 2 times in the scene
- ▶ 5:06 Yining Zhang So big model that we should use H 200 or use H 100 multi nodes. 3 times in the scene
- ▶ 19:32 unnamed speaker The kernels that they come with, I'm yet to see folks do better than what NVIDIA can do when it comes to CUDA kernels.
- ▶ 30:44 unnamed speaker In other words, a single model might want to horizontally scale up to 200 replicas, each of which is, let's say, two H-one hundreds or four H-one hundreds or even a full node.
Beating Google at Search with Neural PageRank and $5M of H200s — with Will Bryk of Exa.ai
AI Engineering for Art - with comfyanonymous
- ▶ 31:20 comfyanonymous (Comfy) But if your models are big and it takes, like, let's say it somewhere has a, like, a A-Fortnite and the model size is 10 gigabytes.
- ▶ 32:08 comfyanonymous (Comfy) Try to remove the least amount of modelings that are already loaded, because it spans, like, Windows, driver, and another problem is the NVIDIA driver on Windows by default, because there's a way to, there's an option to disable that… 4 times in the scene
2024 Year in Review: The Big Scaling Debate, the Four Wars of AI, Top Themes and the Rise of Agents
- ▶ 14:34 Shawn Wang Um, it doesn't even fit on like one node of, uh, of H 100.
- ▶ 29:44 Shawn Wang Basically, uh, this guy, this guy from Nvidia worked out the optimal pricing for language models.
- ▶ 38:54 Shawn Wang the appetite for GPU rich startups, like the, you know, the, the funding plan is we will raise sixty million and we'll give 50 of that to Nvidia.
- ▶ 1:11:01 Shawn Wang Nvidia, most valuable company in the world. 5 times in the scene
- ▶ 1:11:11 Shawn Wang I think the quote that I highlighted in AI news was that it is the best, like Blackwell is the best selling series in the history of the company. 2 times in the scene
- ▶ 1:11:22 Shawn Wang For him to make that statement, I think it's a, it's another indication that the transition from the H to the B series is going to go very well.
page 1 of 2 · 100 scenes per page · newest episode first next →