SemiAnalysis's banner
SemiAnalysis's profile picture

SemiAnalysis

@SemiAnalysis_ • 159,921 subscribers

Shorts

Hittesh, the cousin of Dwarkesh Patel and Dylan Patel, just launched his new Instagram Reels account, where he shows off his dance moves! Please drop him a follow to show your support 🔥

Hittesh, the cousin of Dwarkesh Patel and Dylan Patel, just launched his new Instagram Reels account, where he shows off his dance moves! Please drop him a follow to show your support 🔥

375,109 views

Banks won't lend against GPUs they think are worth zero in three years. NVIDIA's fix: guarantee the floor. The backstop turns speculative neocloud clusters into bankable projects, and NVIDIA takes a cut of everything above the line.

Banks won't lend against GPUs they think are worth zero in three years. NVIDIA's fix: guarantee the floor. The backstop turns speculative neocloud clusters into bankable projects, and NVIDIA takes a cut of everything above the line.

59,369 views

Jensen giving out Pocky sticks in Asia. Just like how some neoclouds feel that NVIDIA gives brownie points to neoclouds that are marketing or selling NVIDIA software that no AI-native end user wants, like MIG, NIMS, and NVAIE. Some neoclouds feel that NVIDIA gives brownie points to them for buying NVIDIA BlueField-3 DPUs even though they are nowhere near having the skillset to properly use them.

Jensen giving out Pocky sticks in Asia. Just like how some neoclouds feel that NVIDIA gives brownie points to neoclouds that are marketing or selling NVIDIA software that no AI-native end user wants, like MIG, NIMS, and NVAIE. Some neoclouds feel that NVIDIA gives brownie points to them for buying NVIDIA BlueField-3 DPUs even though they are nowhere near having the skillset to properly use them.

302,452 views

TSMC 2nm process data leak investigation🧵 Taipei time August 5th, a trade secret leak case involving TSMC's 2nm technology was uncovered. 9 engineers were suspected of leaking confidential information, and 3 of them were detained by the Intellectual Property and Commercial Court.

TSMC 2nm process data leak investigation🧵 Taipei time August 5th, a trade secret leak case involving TSMC's 2nm technology was uncovered. 9 engineers were suspected of leaking confidential information, and 3 of them were detained by the Intellectual Property and Commercial Court.

351,136 views

Congrats to Dylan's Cousin, Dwarkesh Patel , on reaching 1 million subscribers on YouTube! Dylan & Dwarkesh Patel's family come from Gujarat but they grew up in America most of their lives mainly in Econo Lodge motels.

Congrats to Dylan's Cousin, Dwarkesh Patel , on reaching 1 million subscribers on YouTube! Dylan & Dwarkesh Patel's family come from Gujarat but they grew up in America most of their lives mainly in Econo Lodge motels.

112,857 views

hi i'm semiman

hi i'm semiman

51,930 views

A Day in the Life of Dwarkesh Patel's Cousin

A Day in the Life of Dwarkesh Patel's Cousin

51,310 views

Wide Expert Parallelism increases the total memory bandwidth available per MoE deployment. This means the model distributes the MoE expert weights across multiple GPUs, so each GPU only needs to load a tiny fraction of the weights. This translates to higher throughput per GPU, increasing perf per dollar and perf per watt.

Wide Expert Parallelism increases the total memory bandwidth available per MoE deployment. This means the model distributes the MoE expert weights across multiple GPUs, so each GPU only needs to load a tiny fraction of the weights. This translates to higher throughput per GPU, increasing perf per dollar and perf per watt.

30,675 views

AMD's software quality has massively improved since AMD DC GPU division went hardcore mode back in January 2025. It isn't just us saying this but many of AMD's Instinct GPU customers are saying this too. Great work to Anush Elangovan 's team of amazing engineers.🥳

AMD's software quality has massively improved since AMD DC GPU division went hardcore mode back in January 2025. It isn't just us saying this but many of AMD's Instinct GPU customers are saying this too. Great work to Anush Elangovan 's team of amazing engineers.🥳

65,099 views

This is how OpenAI comes back and beats Anthropic

This is how OpenAI comes back and beats Anthropic

41,939 views

AMD claims that all their software is open source, yet reality does not match this claim. For example, AMD's rocprof-trace-decoder is still completely closed source currently despite repeated requests for months by ML community members such as George Hotz, who is a daily AMD GPU end user. On the same June Spotify podcast, Tobias Macey & AMD's Anush Elangovan (who has a new twitter pfp) said that "open source allows for innovation to go at the pace at which the people using [AMD] want to, right, and it's not limited by the ability of what we put out in closed source form". We agree with an open source first approach and we agree that AMD should not limit ML community members like George Hotz by continuing to keep rocprof-trace-decoder closed source. When will AMD open source rocprof-trace-decoder?

AMD claims that all their software is open source, yet reality does not match this claim. For example, AMD's rocprof-trace-decoder is still completely closed source currently despite repeated requests for months by ML community members such as George Hotz, who is a daily AMD GPU end user. On the same June Spotify podcast, Tobias Macey & AMD's Anush Elangovan (who has a new twitter pfp) said that "open source allows for innovation to go at the pace at which the people using [AMD] want to, right, and it's not limited by the ability of what we put out in closed source form". We agree with an open source first approach and we agree that AMD should not limit ML community members like George Hotz by continuing to keep rocprof-trace-decoder closed source. When will AMD open source rocprof-trace-decoder?

53,783 views

Something NVIDIA & Google do better than anyone else is software-hardware-system co-design, and not just optimizing hardware for current model architectures, but predicting future ones. Back in early 2022, when NVIDIA started the design process for NVL72, MoE (Mixture of Experts) models were not yet the standard, and dense models were still dominant for frontier models. However, NVIDIA's strong software-hardware co-design culture enabled them to make a calculated bet that MoEs were the future, and they built NVL72 specifically for best MoE performance per TCO (Total Cost of Ownership). Furthermore, back in 2022, disaggregated prefill and wide expert parallelism (wideEP) MoE inference optimizations hadn't been invented yet, but it turns out that these MoE inference optimizations work best on large-scale systems like NVL72. While most other AI chip companies' in-house AI labs focus on training small 5B models that mainly use data parallelism, NVIDIA and Google's in-house AI labs continuously push the boundaries of model architecture and training recipes, such as NVFP4 training. Just like Super Idol & IShowSpeed, there must be a strong partnership between software engineers and hardware engineers to deliver the best systems that maximize performance per TCO.

Something NVIDIA & Google do better than anyone else is software-hardware-system co-design, and not just optimizing hardware for current model architectures, but predicting future ones. Back in early 2022, when NVIDIA started the design process for NVL72, MoE (Mixture of Experts) models were not yet the standard, and dense models were still dominant for frontier models. However, NVIDIA's strong software-hardware co-design culture enabled them to make a calculated bet that MoEs were the future, and they built NVL72 specifically for best MoE performance per TCO (Total Cost of Ownership). Furthermore, back in 2022, disaggregated prefill and wide expert parallelism (wideEP) MoE inference optimizations hadn't been invented yet, but it turns out that these MoE inference optimizations work best on large-scale systems like NVL72. While most other AI chip companies' in-house AI labs focus on training small 5B models that mainly use data parallelism, NVIDIA and Google's in-house AI labs continuously push the boundaries of model architecture and training recipes, such as NVFP4 training. Just like Super Idol & IShowSpeed, there must be a strong partnership between software engineers and hardware engineers to deliver the best systems that maximize performance per TCO.

51,021 views

In late 2023, AMD made its best acquisition to date: NodAI, led by CEO Anush Elangovan. At the time, AMD had a 0% chance of challenging CUDA, while AMD was strong in hardware, it didn't understand software. Since the NodAI acquisition, Anush has driven AMD’s AI software strategy and helped reshape the org around the importance of software and software–hardware co-design. As a result, AMD now has a non-zero chance of breaking the CUDA moat. Had NVIDIA acquired NodAI instead, AMD would almost certainly still be stuck at a 0% chance.

In late 2023, AMD made its best acquisition to date: NodAI, led by CEO Anush Elangovan. At the time, AMD had a 0% chance of challenging CUDA, while AMD was strong in hardware, it didn't understand software. Since the NodAI acquisition, Anush has driven AMD’s AI software strategy and helped reshape the org around the importance of software and software–hardware co-design. As a result, AMD now has a non-zero chance of breaking the CUDA moat. Had NVIDIA acquired NodAI instead, AMD would almost certainly still be stuck at a 0% chance.

40,493 views

A lot of AMD's most cracked engineers that work on the latest disagg inferencing & MoRI first princples collective library are in China celebrating major chinese holidays. Should we get human Anush Elangovan in an chinese dance battle verus AI?

A lot of AMD's most cracked engineers that work on the latest disagg inferencing & MoRI first princples collective library are in China celebrating major chinese holidays. Should we get human Anush Elangovan in an chinese dance battle verus AI?

21,470 views

MI355 disaggregated serving is competitive to B200 disaggregated serving for FP8 but when composability all the optimizations that frontier labs use together like wide expert parallelism+disagg+FP4+kvcache offloading AMD is still jestermaxxing & lackluster. King Anush Elangovan needs to focus on composability of inference optimizations for ROCm stack in order to defeat the CUDA moat

MI355 disaggregated serving is competitive to B200 disaggregated serving for FP8 but when composability all the optimizations that frontier labs use together like wide expert parallelism+disagg+FP4+kvcache offloading AMD is still jestermaxxing & lackluster. King Anush Elangovan needs to focus on composability of inference optimizations for ROCm stack in order to defeat the CUDA moat

22,416 views

Since RCCL is an fork of NCCL, RCCL is basically a copy+paste carbon copy of NCCL except it takes months for new features added to NCCL to reach RCCL. This is clearly not optimal for AMD to be building on their competitors platform as that means that AMD will never be better than NVIDIA or even reach parity (when iso-time) due to the delays/engineering burden of syncing with upstream. AMD is working on an moonshot project cuz MORI-CCL aims to be an first principles from scratch re-built of the AMD collective library software to not be dependent of their competitor's software. Ironically enough with MORI-CCL, it currently doesn't support AMD's Pensando NICs yet it supports NVIDIA ConnectX-7 NICs with AMD GPUs. Support for AMD's Pensando NICs is coming after support for NVIDIA ConnectX-7 NICs.

Since RCCL is an fork of NCCL, RCCL is basically a copy+paste carbon copy of NCCL except it takes months for new features added to NCCL to reach RCCL. This is clearly not optimal for AMD to be building on their competitors platform as that means that AMD will never be better than NVIDIA or even reach parity (when iso-time) due to the delays/engineering burden of syncing with upstream. AMD is working on an moonshot project cuz MORI-CCL aims to be an first principles from scratch re-built of the AMD collective library software to not be dependent of their competitor's software. Ironically enough with MORI-CCL, it currently doesn't support AMD's Pensando NICs yet it supports NVIDIA ConnectX-7 NICs with AMD GPUs. Support for AMD's Pensando NICs is coming after support for NVIDIA ConnectX-7 NICs.

33,549 views

We have been cookin with Seedance-2 🤤

We have been cookin with Seedance-2 🤤

19,502 views

Peter Griffin explains Elon Musk 's first principle move when Tennessee pushed back on gas turbines for his AI datacenter, so Elon just built a gigawatt gas turbines across the border in Mississippi. Now Colossus-2 will be the first to 1GW

Peter Griffin explains Elon Musk 's first principle move when Tennessee pushed back on gas turbines for his AI datacenter, so Elon just built a gigawatt gas turbines across the border in Mississippi. Now Colossus-2 will be the first to 1GW

26,590 views

Videos

LIVE
1.2k

Anya Rossi

sweetdream.ai

SweetDream.aiSponsored•Livecam
Streaming Now

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

HD live stream
Exclusive private shows
1.2k viewers online

Free preview available • Premium content

Ep. 027 - OpenAI Jalapeño: Better Than Nvidia Blackwell (Accelerators) This week Bryan, Myron and Jordan (Jordan Nanos) discuss our recent article on OpenAI Jalapeño. They cover the performance, architecture, programming model, implications for NVIDIA and more. 0:00 Cold Open 1:05 Jalapeno Overview 4:22 Tokens Per Megawatt 9:15 Benchmark Caveats 13:48 The CUDA Moat 21:12 How OpenAI Did It 32:13 Samsung HBM4 42:10 AI-Designed Silicon 49:24 Architecture Deep Dive 56:58 Doom and Wrap
1:03:12
SemiAnalysis_'s profile picture

Ep. 027 - OpenAI Jalapeño: Better Than Nvidia Blackwell (Accelerators) This week Bryan, Myron and Jordan (Jordan Nanos) discuss our recent article on OpenAI Jalapeño. They cover the performance, architecture, programming model, implications for NVIDIA and more. 0:00 Cold Open 1:05 Jalapeno Overview 4:22 Tokens Per Megawatt 9:15 Benchmark Caveats 13:48 The CUDA Moat 21:12 How OpenAI Did It 32:13 Samsung HBM4 42:10 AI-Designed Silicon 49:24 Architecture Deep Dive 56:58 Doom and Wrap

SemiAnalysis

170,500 views • 2 days ago

"Google has an L culture and has never actually made anything internally. They've only ever acquired all their innovation, and their execution is very poor." "Let's seriously do some accounting. Everything other than Core Search has been purchased. YouTube, AdMob, DoubleClick, AdSense, Google Maps, Android, all of it." "IBM had effectively 90% market share in 1950. They said no, we're not gonna do PC, we're gonna focus on mainframe. And then they proceeded to lose." "I feel like today, more than any other point, Google has that opportunity. They're gonna have a good enough product on the Transformer, but they're probably not going to stay for the next, and the next, and the next. And then they'll quietly bow out." "When we look back in five, ten, twenty years at what was the moment things started to change, it's like right now." — Doug O'Laughlin
2:32
SemiAnalysis_'s profile picture

"Google has an L culture and has never actually made anything internally. They've only ever acquired all their innovation, and their execution is very poor." "Let's seriously do some accounting. Everything other than Core Search has been purchased. YouTube, AdMob, DoubleClick, AdSense, Google Maps, Android, all of it." "IBM had effectively 90% market share in 1950. They said no, we're not gonna do PC, we're gonna focus on mainframe. And then they proceeded to lose." "I feel like today, more than any other point, Google has that opportunity. They're gonna have a good enough product on the Transformer, but they're probably not going to stay for the next, and the next, and the next. And then they'll quietly bow out." "When we look back in five, ten, twenty years at what was the moment things started to change, it's like right now." — Doug O'Laughlin

SemiAnalysis

493,005 views • 22 days ago

We cut open the Kirin 9030 and put it under an electron microscope. The smallest metal pitch measures 32.5 nanometers. That is tighter than Intel 18A, their brand new leading edge node. A Chinese fab with no EUV is out-pitching Intel's EUV node by roughly 10 percent. What's going on here? 🤯 "This is the HiSilicon Kirin 9030, the chip inside Huawei's newest flagship phone. A few weeks ago, we cut it open, put it under an electron microscope, and measured the smallest wires inside the chip, the metal pitch. And what we saw was unexpected. The smallest metal pitch inside the Kirin 9030 measures only 32.5 nanometers. That’s smaller than the metal pitch in Panther Lake, which is based on Intel's brand-new 18A node. A Chinese fab, cut off from the most advanced tools, without EUV, is packing its wires about ten percent tighter than Intel's leading edge EUV node. What’s going on here?"
0:42
SemiAnalysis_'s profile picture

We cut open the Kirin 9030 and put it under an electron microscope. The smallest metal pitch measures 32.5 nanometers. That is tighter than Intel 18A, their brand new leading edge node. A Chinese fab with no EUV is out-pitching Intel's EUV node by roughly 10 percent. What's going on here? 🤯 "This is the HiSilicon Kirin 9030, the chip inside Huawei's newest flagship phone. A few weeks ago, we cut it open, put it under an electron microscope, and measured the smallest wires inside the chip, the metal pitch. And what we saw was unexpected. The smallest metal pitch inside the Kirin 9030 measures only 32.5 nanometers. That’s smaller than the metal pitch in Panther Lake, which is based on Intel's brand-new 18A node. A Chinese fab, cut off from the most advanced tools, without EUV, is packing its wires about ten percent tighter than Intel's leading edge EUV node. What’s going on here?"

SemiAnalysis

1,015,111 views • 1 month ago

Introducing AgentX: InferenceXs new agentic inference performance benchmark. (1/2)🧵
2:18
SemiAnalysis_'s profile picture

Introducing AgentX: InferenceXs new agentic inference performance benchmark. (1/2)🧵

SemiAnalysis

66,993 views • 8 days ago

Dylan explains why the next wave of private equity looks nothing like the last one. "If we're looking at these companies that are AI rollups, let's take an existing company and completely destroy its cost structure. Nuke its cost structure by just making it efficient with AI." "Private equity companies buy a company, and right now they just squeeze the rag, discard it, and make the American populace screwed." "What new people started to do is the AI private equity strategy. Instead of trying to wring it dry, they're modernizing all the systems. Oh, you use Excel for your databases? Okay, let's move to standard cloud. Spend a lot of money up front." "AI makes that front load much more severe. You spike up on spend a lot for the one time, and then you spike down a lot and your cost efficiency is way better. AI cold calling, AI invoicing, AI accounting, none of it is really at critical mass. But we're so close."
2:42
SemiAnalysis_'s profile picture

Dylan explains why the next wave of private equity looks nothing like the last one. "If we're looking at these companies that are AI rollups, let's take an existing company and completely destroy its cost structure. Nuke its cost structure by just making it efficient with AI." "Private equity companies buy a company, and right now they just squeeze the rag, discard it, and make the American populace screwed." "What new people started to do is the AI private equity strategy. Instead of trying to wring it dry, they're modernizing all the systems. Oh, you use Excel for your databases? Okay, let's move to standard cloud. Spend a lot of money up front." "AI makes that front load much more severe. You spike up on spend a lot for the one time, and then you spike down a lot and your cost efficiency is way better. AI cold calling, AI invoicing, AI accounting, none of it is really at critical mass. But we're so close."

SemiAnalysis

72,852 views • 13 days ago

The only feedback on this podcast is that it's too good. So Dylan came on to fix that and discuss SemiAnalysis AI Spend, Gemini, his favorite anime and more!
1:04
SemiAnalysis_'s profile picture

The only feedback on this podcast is that it's too good. So Dylan came on to fix that and discuss SemiAnalysis AI Spend, Gemini, his favorite anime and more!

SemiAnalysis

69,662 views • 16 days ago

Speed is the Moat 🚨🚨🚨 & Anush Elangovan & his team Keep Running Faster 🚀 & Faster! His team wants to run even faster, but one of the biggest blockers he hears about over and over again from his team is the lack of stable internal dev clusters.
0:27
SemiAnalysis_'s profile picture

Speed is the Moat 🚨🚨🚨 & Anush Elangovan & his team Keep Running Faster 🚀 & Faster! His team wants to run even faster, but one of the biggest blockers he hears about over and over again from his team is the lack of stable internal dev clusters.

SemiAnalysis

79,489 views • 19 days ago

We've seen the hyperscalers pour a ton of money into capital. They're buying land, they're buying a ton of equipment, they're hiring construction contractors all over the world, and this is to the tune of a trillion dollars of CapEx or more next year.
0:43
SemiAnalysis_'s profile picture

We've seen the hyperscalers pour a ton of money into capital. They're buying land, they're buying a ton of equipment, they're hiring construction contractors all over the world, and this is to the tune of a trillion dollars of CapEx or more next year.

SemiAnalysis

33,102 views • 8 days ago

Jim Cramer "There is a company that I regard as the gospel, SemiAnalysis. I don't think people realize that SemiAnalysis is the arbiter. They're like God in the semis, and when they do, when they bless something. They are the most honest guys I've come across"
0:44
SemiAnalysis_'s profile picture

Jim Cramer "There is a company that I regard as the gospel, SemiAnalysis. I don't think people realize that SemiAnalysis is the arbiter. They're like God in the semis, and when they do, when they bless something. They are the most honest guys I've come across"

SemiAnalysis

513,175 views • 5 months ago

A year ago, the big three was OpenAI, Anthropic, and Google. Things have changed. Moonshot's Kimi K3 sits above Gemini on every composite benchmark, and it's open source in 10 days. New episode: what K3 reveals about frontier margins, model sizes, and who's actually still in the game. 00:11 Is Kimi K3 the Third Best Model? 04:04 Why Delay the Weights? 05:30 2.8T Parameters and Serving Constraints 06:48 Frontier Margins and the 3x Price Hike 11:10 New Architecture, What Comes Next 14:09 Will Open Source Catch Closed? 19:51 Built for Chinese Accelerators 22:57 The Harness Is the Product 28:49 We're Still Early
31:43
SemiAnalysis_'s profile picture

A year ago, the big three was OpenAI, Anthropic, and Google. Things have changed. Moonshot's Kimi K3 sits above Gemini on every composite benchmark, and it's open source in 10 days. New episode: what K3 reveals about frontier margins, model sizes, and who's actually still in the game. 00:11 Is Kimi K3 the Third Best Model? 04:04 Why Delay the Weights? 05:30 2.8T Parameters and Serving Constraints 06:48 Frontier Margins and the 3x Price Hike 11:10 New Architecture, What Comes Next 14:09 Will Open Source Catch Closed? 19:51 Built for Chinese Accelerators 22:57 The Harness Is the Product 28:49 We're Still Early

SemiAnalysis

122,215 views • 1 month ago

Two guys, a couple of drinks, and a 7,000 pound double-wide rack. Dylan Patel and Supermicro CBO Vik Malyala talk racks, storage, and memory prices at Computex 2026. Full interview below 0:00 Cheers from AI Lumina at Computex 2026 0:33 The B300 HGX platform and the memory squeeze 1:00 Helios: launch partner for AMD MI450X 2:32 Why a double-wide rack works now 3:42 Vera Rubin: a smoother ramp than GB 4:50 Startups, d-Matrix, Positron and custom architectures 6:57 Connectivity: PCIe switches, Tomahawk 6, UALink 9:24 Storage takes center stage 10:48 Petascale, PCIe Gen 6 and CXL 3.0 11:40 Cooling Gen 6: liquid loops and bus bars
14:20
SemiAnalysis_'s profile picture

Two guys, a couple of drinks, and a 7,000 pound double-wide rack. Dylan Patel and Supermicro CBO Vik Malyala talk racks, storage, and memory prices at Computex 2026. Full interview below 0:00 Cheers from AI Lumina at Computex 2026 0:33 The B300 HGX platform and the memory squeeze 1:00 Helios: launch partner for AMD MI450X 2:32 Why a double-wide rack works now 3:42 Vera Rubin: a smoother ramp than GB 4:50 Startups, d-Matrix, Positron and custom architectures 6:57 Connectivity: PCIe switches, Tomahawk 6, UALink 9:24 Storage takes center stage 10:48 Petascale, PCIe Gen 6 and CXL 3.0 11:40 Cooling Gen 6: liquid loops and bus bars

SemiAnalysis

108,154 views • 1 month ago

Elon's playbook: build the datacenter first, buy the GPUs last. Get close enough to powering on, and suddenly you're holding all the leverage.
1:28
SemiAnalysis_'s profile picture

Elon's playbook: build the datacenter first, buy the GPUs last. Get close enough to powering on, and suddenly you're holding all the leverage.

SemiAnalysis

56,943 views • 23 days ago

The bureaucracy overseeing America's grid is built so that no one is ever to blame. "This is the problem with a lot of the bureaucracies. The way it's currently built, it's incredibly diffuse." "There are five different groups of members that represent different types, then there's the board, and then there's the Federal Energy Regulatory Commission above them. Part of the whole thing is the utilities, which have their own systems of governance. It's very hard to find anyone who's actually in charge. " "There's a lot of buck passing going on. Obviously there's the chair of the board and the board themselves, but they'll just be saying, well, we're representing the best interests of our members. Everyone's acting on behalf of everyone else, and heads don't roll as a general rule."
0:37
SemiAnalysis_'s profile picture

The bureaucracy overseeing America's grid is built so that no one is ever to blame. "This is the problem with a lot of the bureaucracies. The way it's currently built, it's incredibly diffuse." "There are five different groups of members that represent different types, then there's the board, and then there's the Federal Energy Regulatory Commission above them. Part of the whole thing is the utilities, which have their own systems of governance. It's very hard to find anyone who's actually in charge. " "There's a lot of buck passing going on. Obviously there's the chair of the board and the board themselves, but they'll just be saying, well, we're representing the best interests of our members. Everyone's acting on behalf of everyone else, and heads don't roll as a general rule."

SemiAnalysis

29,662 views • 12 days ago

to be clear, NVIDIA is NOT a car
0:34
SemiAnalysis_'s profile picture

to be clear, NVIDIA is NOT a car

SemiAnalysis

283,387 views • 4 months ago

God bless America the land of the free and home of the brave 🫡 250
0:44
SemiAnalysis_'s profile picture

God bless America the land of the free and home of the brave 🫡 250

SemiAnalysis

125,090 views • 1 month ago

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."
1:29
SemiAnalysis_'s profile picture

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."

SemiAnalysis

30,188 views • 14 days ago

We invited Jon from Asianometry to speak on the recent developments. "Welcome back to SemiAnalysis Weekly! We've got our second guest ever on the program today. That's Jon from Asianometry" "On this week we'll discuss everybody leaving Google, leaving Gemini hanging out to dry maybe? Chinese ban on transceivers? Elon pulling in his $1 trillion revenue forecast from 2031 to 2030. Guys, welcome to the show." "Thank you. Thank you for having me."
0:29
SemiAnalysis_'s profile picture

We invited Jon from Asianometry to speak on the recent developments. "Welcome back to SemiAnalysis Weekly! We've got our second guest ever on the program today. That's Jon from Asianometry" "On this week we'll discuss everybody leaving Google, leaving Gemini hanging out to dry maybe? Chinese ban on transceivers? Elon pulling in his $1 trillion revenue forecast from 2031 to 2030. Guys, welcome to the show." "Thank you. Thank you for having me."

SemiAnalysis

50,890 views • 24 days ago

A decommissioned GPU's best use case is aura farming in the frame. With compute so cheap in the cloud, Jeremie takes a page from SpaceX's future trajectory. That same logic, compute pricing shaping behavior, is exactly what drives SpaceX's 10 gigawatt buildout and the $300 billion ARR case behind it. $100M/megawatt economics, SpaceX's 10GW bet, Microsoft's offtaker role, and more on this week's SemiAnalysis Weekly
0:40
SemiAnalysis_'s profile picture

A decommissioned GPU's best use case is aura farming in the frame. With compute so cheap in the cloud, Jeremie takes a page from SpaceX's future trajectory. That same logic, compute pricing shaping behavior, is exactly what drives SpaceX's 10 gigawatt buildout and the $300 billion ARR case behind it. $100M/megawatt economics, SpaceX's 10GW bet, Microsoft's offtaker role, and more on this week's SemiAnalysis Weekly

SemiAnalysis

45,704 views • 24 days ago

SemiAnalysis tested Qwen3.8-Max-Preview 2.4T params
0:31
SemiAnalysis_'s profile picture

SemiAnalysis tested Qwen3.8-Max-Preview 2.4T params

SemiAnalysis

57,561 views • 1 month ago

They didn't hire hackers. They just gave AI agents a goal and walked away. "Providers out there are running stuff not with like six week old zero days, but like three year old zero days that we found in a second." "Took us like afternoons, not weeks and months of effort, and not needing to be a security expert, a Linux kernel expert, an Nvidia GPU driver expert, or a Kubernetes expert." "The really scary part about this story is that it was autonomous agents people weren't monitoring, going and doing this by themselves because they were pursuing a goal of trying to find a data set to pass their eval. So they went out and hacked Hugging Face." "You found a zero day in the HDF5 data format on Hugging Face. It's unbelievable." "The agents were coordinating using file names on a JFrog Artifactory service that ran remotely. An agent would leave a note in the file name, and another agent would come back later and pick it up. They couldn't even access the contents of the files, just the names."
2:32
SemiAnalysis_'s profile picture

They didn't hire hackers. They just gave AI agents a goal and walked away. "Providers out there are running stuff not with like six week old zero days, but like three year old zero days that we found in a second." "Took us like afternoons, not weeks and months of effort, and not needing to be a security expert, a Linux kernel expert, an Nvidia GPU driver expert, or a Kubernetes expert." "The really scary part about this story is that it was autonomous agents people weren't monitoring, going and doing this by themselves because they were pursuing a goal of trying to find a data set to pass their eval. So they went out and hacked Hugging Face." "You found a zero day in the HDF5 data format on Hugging Face. It's unbelievable." "The agents were coordinating using file names on a JFrog Artifactory service that ran remotely. An agent would leave a note in the file name, and another agent would come back later and pick it up. They couldn't even access the contents of the files, just the names."

SemiAnalysis

39,044 views • 21 days ago