Latest / AI Fire Daily / 🎙️ EP 394: Nvidia CEO Admits Affair Amid Spy Probe & Nous Research’s "Mind-Wiped" AI!

Transcript

🎙️ EP 394: Nvidia CEO Admits Affair Amid Spy Probe & Nous Research’s "Mind-Wiped" AI! · AI Fire Daily

• 0:00
  The CEO of the world's most crucial AI chip maker,
• 0:03
  a $4 .8 billion customer, and a suspected intelligence
• 0:09
  asset. Yeah, it's wild. It is. This isn't a spy
• 0:12
  thriller, it's reality. Welcome to the Deep Dive.
• 0:15
  Today we are exploring the unseen battlegrounds
• 0:17
  of artificial intelligence. Right, because the
• 0:19
  battlefield is, you know, it's not just about
• 0:21
  the code anymore. No, it's much bigger than that.
• 0:23
  We will look at three distinct layers of this
• 0:26
  global AI war. First, human vulnerability at
• 0:29
  the hardware level with NVIDIA. Second, the massive
• 0:32
  scramble for infrastructure and talent. And finally,
• 0:34
  an attempt to mind wipe an AI's political ideology.
• 0:38
  It is a fascinating stack of interconnected problems.
• 0:41
  We're going to unpack the mechanics behind all
• 0:43
  these messages. Exactly. So let's start at the
• 0:46
  physical bedrock of AI. We were talking about
• 0:48
  the hardware itself, and there is a major human
• 0:52
  complication right at the very top. Yeah, the
• 0:54
  Jensen Huang situation. Right. Jensen Huang is
• 0:57
  the CEO of NVIDIA. He recently sent a company
• 1:00
  -wide email admitting to an extramarital affair.
• 1:03
  Which, you know, normally would just be a standard
• 1:05
  HR or corporate governance issue. But the specific
• 1:08
  context here, it elevates it to a national security
• 1:11
  level. Right, because the affair was with a woman
• 1:14
  named Alice Huang. She is a former executive
• 1:16
  at a company called Megaspeed. They are a massive
• 1:19
  tech player based in Singapore. But here is where
• 1:22
  it gets... Highly complicated. U .S. intelligence
• 1:24
  officials suspect she's operating as a Chinese
• 1:27
  intelligence asset. Right. And they believe her
• 1:29
  specific target was NVIDIA's business strategy.
• 1:33
  I mean, state sponsored intelligence doesn't
• 1:35
  always look like, you know, someone hacking a
• 1:36
  mainframe. Yeah, it's more subtle. Exactly. Often
• 1:39
  it is about mapping out a hardware roadmap. through
• 1:42
  personal relationships. Which makes the stakes
• 1:44
  here just astronomical. Megaspeed is not a small
• 1:47
  -time buyer. They reportedly purchased over $4
• 1:50
  .8 billion in NVIDIA chips, and that volume occurred
• 1:54
  just over the last two years. That is a staggering
• 1:57
  amount of compute power. Singapore has obviously
• 2:00
  become a major hub for AI infrastructure, but
• 2:03
  it also acts as a potential proxy node. Right,
• 2:05
  for routing things. Yeah. Companies can sometimes
• 2:08
  use third -country hubs to circumvent international
• 2:11
  sanctions. Because U .S. export controls are
• 2:14
  incredibly strict right now. They are aggressively
• 2:16
  trying to keep high -end chips out of China.
• 2:18
  It feels like setting up the world's most advanced
• 2:21
  security system only to realize you left a human
• 2:24
  backdoor wide open. That is the perfect way to
• 2:27
  look at it. The physical hardware is completely
• 2:30
  locked down. The global supply chains are heavily
• 2:33
  monitored. But intimate human relationships bypass
• 2:36
  all those expensive firewalls entirely. So...
• 2:40
  NVIDIA's board immediately took aggressive action.
• 2:43
  They appointed independent directors. They brought
• 2:45
  in outside legal counsel to run a full investigation.
• 2:49
  They had to. Yeah. They are looking for any potential
• 2:51
  national security risks. They literally have
• 2:53
  no other choice. The minute U .S. intelligence
• 2:56
  flags a $4 .8 billion customer, everything changes.
• 2:59
  The board's fiduciary duty forces them to assume
• 3:03
  the worst case scenario. They have to investigate
• 3:05
  if a hardware roadmap was actually leaked. So
• 3:08
  far, NVIDIA claims there is zero proof of a breach.
• 3:11
  They have no reason to believe company secrets
• 3:13
  were actually compromised. They don't think U
• 3:15
  .S. national security was breached in any way.
• 3:17
  But the underlying concern is about strategic
• 3:20
  insights, right? Not just blueprints. Investigators
• 3:24
  worry she gained high -level context about NVIDIA's
• 3:27
  business strategy. Especially regarding upcoming
• 3:29
  negotiations with the new Trump administration.
• 3:32
  Exactly. That political angle is completely undeniable.
• 3:35
  Republican lawmakers are already stepping in
• 3:37
  aggressively. They have requested a private briefing
• 3:39
  with Huang on November 18th. They are citing
• 3:42
  severe national security concerns over this entire
• 3:45
  situation. You can easily see why Washington
• 3:47
  is so nervous. NVIDIA is the absolute linchpin
• 3:51
  of American tech dominance right now. If their
• 3:54
  strategic posture is compromised. the whole ecosystem
• 3:57
  wobbles. Yep. There is currently no public proof
• 4:00
  of actual espionage. There is no proof that Huang
• 4:03
  disclosed sensitive information. There is no
• 4:05
  proof he knowingly assisted Chinese intelligence
• 4:08
  at all. Right. And the board continues to strongly
• 4:11
  support his leadership. There are zero announced
• 4:13
  plans to replace him. Which is a very calculated
• 4:16
  statement from the board. It is. Given the massive
• 4:19
  geopolitical stakes, why is the board standing
• 4:22
  so firmly behind him without a pause? Well, it...
• 4:25
  It comes down to stark market reality. Replacing
• 4:27
  a visionary founder during an unprecedented tech
• 4:30
  boom is incredibly dangerous. It could instantly
• 4:34
  crash the stock and completely disrupt their
• 4:36
  entire hardware roadmap. They are essentially
• 4:39
  balancing corporate stability against these lingering
• 4:42
  intelligence risks. So he's simply too critical
• 4:46
  to NVIDIA's future to sideline right now. Exactly.
• 4:48
  You don't rebuild the airplane's engine while
• 4:50
  you are cruising at 30 ,000 feet. Yeah, that
• 4:53
  makes sense. So let's follow this thread from
• 4:56
  the chips to the broader ecosystem. NVIDIA supplies
• 4:59
  the raw hardware, obviously, but those chips
• 5:02
  require massive physical and digital ecosystems
• 5:05
  to function. Massive is an understatement. Right.
• 5:07
  Anthropic needs 1 .7 gigawatts because models
• 5:10
  are scaling exponentially. That scale is the
• 5:13
  underlying force driving everything else. Anthropic
• 5:15
  just locked in a $3 .4 billion power deal. This
• 5:19
  infrastructure investment is just for their AI
• 5:21
  data centers. It includes 1 .7 gigawatts of gas
• 5:25
  -fired generators. A power deal that size isn't
• 5:28
  just about paying the electric bill. It is really
• 5:30
  about grid density. Break that down for me. Why
• 5:33
  can't they just, you know, plug these into the
• 5:35
  existing municipal grid? Because the localized
• 5:37
  heat and power draw would literally melt the
• 5:39
  substations. Wow. Yeah, you cannot drop a gigawatt
• 5:42
  load on a standard city grid. So these AI companies
• 5:46
  are essentially having to become their own utility
• 5:47
  providers. They are building bespoke energy infrastructure
• 5:51
  just to feed the models. And that insatiable
• 5:53
  power demand is driven by the rise of agents.
• 5:56
  The scale of usage is just exploding everywhere
• 5:58
  right now. Look at Deep Sea Karnas. Oh, yeah.
• 6:01
  The metrics coming out of there are stunning.
• 6:03
  Whoa. Processing 92 trillion tokens every single
• 6:07
  day. That is just difficult to even conceptualize.
• 6:09
  The growth curve is basically vertical at this
• 6:12
  point. They hit 1 .54 million daily users in
• 6:15
  barely three months. They already have 35 ,000
• 6:18
  distinct plugins. Open model agnostic agents
• 6:21
  are going completely viral. Developers desperately
• 6:24
  want systems that can operate across different
• 6:26
  platforms. They want AI that can fetch data,
• 6:30
  execute code, and complete workflows seamlessly.
• 6:33
  But when agents have that much systemic access,
• 6:36
  security completely breaks down, which explains
• 6:39
  why Palo Alto Networks is making massive moves.
• 6:41
  Yeah, they are. They might pay $500 million just
• 6:44
  to acquire Clutch. Clutch is a startup specifically
• 6:46
  focused on protecting AI agents system access.
• 6:49
  And that specific valuation is a massive sign
• 6:52
  of the times. 500 million is roughly 50 times
• 6:57
  their annual recurring revenue. It is nearly
• 6:59
  six times their projected 2025 valuation. 50
• 7:03
  times revenue for a startup like Clutch, is that
• 7:06
  a sustainable valuation or just sheer panic?
• 7:08
  It is about existential risk. Securing AI agents
• 7:12
  that have system -level database access is a
• 7:14
  hard requirement. If an agent goes rogue, it
• 7:16
  can delete your entire backend. Right. So traditional
• 7:19
  revenue multiples are mostly irrelevant in the
• 7:21
  short term. Enterprise companies desperately
• 7:23
  need that specific security layer right now.
• 7:26
  Basically, they are paying a massive panic premium
• 7:29
  to secure the AI frontier. Right. And they can
• 7:32
  afford to do it. But the cost of a systemic agent
• 7:34
  breach is just infinitely higher. Now, to build
• 7:38
  all these secure and scalable agent ecosystems,
• 7:40
  you need elite talent. And the talent pool capable
• 7:43
  of building this architecture is tiny, which
• 7:46
  explains Kat Wu's recent move. She just jumped
• 7:48
  from Anthropic over to OpenAI. Yeah, that specific
• 7:51
  talent war is incredibly fierce right now. OpenAI
• 7:53
  is aggressively poaching top -tier engineers
• 7:56
  from Anthropic. She is the fourth key Claude
• 7:58
  Code departure in recent weeks. Wait, four key
• 8:01
  architects leaving in just a few weeks. Yeah.
• 8:03
  That is not just normal corporate turnover. Are
• 8:06
  they anticipating a fundamental shift in how
• 8:08
  these agents are built? It certainly looks like
• 8:10
  a coordinated brain drain. When top talent moves
• 8:13
  in clusters, they're usually following a specific
• 8:16
  technological breakthrough. Also, it is kind
• 8:18
  of funny. Her ex -bio accidentally still links
• 8:21
  to Anthropic. Oh, really? Yeah. It's a tiny industry,
• 8:24
  but they're moving at breakneck speed. And to
• 8:27
  support that speed, we are seeing a flood of
• 8:29
  new developer tools. Companies are racing to
• 8:31
  lower the friction of building these systems.
• 8:34
  Mistral Studio also just launched to solve this
• 8:36
  exact problem. It is a brilliant play for developer
• 8:39
  capture. It lets you build, host, and refine
• 8:42
  AI apps in one unified place. You can connect
• 8:45
  data and launch complex agents or workflows.
• 8:48
  Then you deploy fast and use real traces to iteratively
• 8:51
  improve them. It is all about accelerating that
• 8:54
  feedback loop for the developer. OpenAI is also
• 8:57
  rolling out ChatGPT Remix for Enterprise. Teams
• 9:00
  can actually fork a colleague's ChatGPT site,
• 9:03
  and you can edit your own copy in the chat interface,
• 9:05
  and it keeps the original prompt architecture
• 9:07
  completely safe. That is a massive update for
• 9:10
  corporate collaboration. You really don't want
• 9:13
  a junior employee accidentally breaking a critical
• 9:15
  working prompt. Honestly, I still wrestle with
• 9:17
  prompt drift myself. I tweak one little parameter
• 9:20
  and the whole output just completely falls apart.
• 9:23
  So having a safe sandbox to remix things is an
• 9:28
  absolute lifesaver. It brings essential version
• 9:30
  control to natural language processing, which
• 9:33
  is a necessary evolution if we want reliable
• 9:36
  enterprise tools. We are also seeing tools that
• 9:39
  blend coding with culture. OpenAI is teasing
• 9:43
  Mod Retro's Chromatic for November 9th. It physically
• 9:46
  plays Game Boy Classics, but the catch is that
• 9:50
  Codex can actually help you build games for it.
• 9:52
  It is a remarkably clever way to showcase their
• 9:55
  coding capabilities. They are mixing deep nostalgia
• 9:59
  with cutting -edge generation to engage developers.
• 10:02
  So we've secured the chips, the power, and the
• 10:05
  software agents. But how do you actually secure
• 10:08
  the thoughts of the AI itself? We are moving
• 10:11
  now from physical infrastructure to cognitive
• 10:13
  control. This is where we touch the ideological
• 10:16
  source code. Noose Research just released a model
• 10:19
  called Arete 27B. It is an entirely open source
• 10:23
  AI model, but it was built directly on top of
• 10:26
  Alibaba's QEM 3 .8 -27B model. And Alibaba obviously
• 10:30
  built that base model under strict Chinese regulatory
• 10:33
  frameworks. It was naturally aligned with their
• 10:37
  specific domestic policies. Right. Now, I want
• 10:40
  to be very clear with you here. We aren't taking
• 10:42
  a side on the politics of this issue. We are
• 10:44
  just looking at the fascinating mechanics of
• 10:46
  how they achieved this. Noose Research specifically
• 10:49
  modified this base model. They wanted to fundamentally
• 10:52
  reduce Chinese government aligned responses.
• 10:56
  They aimed to strongly reinforce Western democratic
• 10:59
  values instead. It is a deliberate surgical.
• 11:03
  ideological rewiring of a massive system. They
• 11:06
  used a technical process called fine -tuning
• 11:08
  to do this, giving an existing AI a specialized
• 11:11
  crash course in new rules. They trained it on
• 11:14
  a data set of just 53 million tokens. Which is
• 11:17
  actually a surprisingly tiny amount of training
• 11:19
  data. Exactly. But how does 53 million tokens
• 11:23
  override a massive base model? Doesn't aggressive
• 11:26
  retraining usually break the general intelligence
• 11:28
  of the system? Usually, yes. It causes something
• 11:30
  called catastrophic forgetting. You align the
• 11:33
  politics, but it suddenly forgets how to do basic
• 11:35
  math. Right. The genius here is this surgical
• 11:39
  weight modification. By using highly targeted
• 11:42
  tokens, they rewrote the cultural guardrails
• 11:44
  without touching the core logic engine. That
• 11:47
  makes total sense. Yeah. They tested this intervention
• 11:50
  across 803 specific prompts. These covered freedom
• 11:53
  of speech, democracy, privacy, and other sensitive
• 11:56
  topics. The resulting metrics are pretty wild.
• 11:59
  The propaganda rate dropped from 22 % all the
• 12:01
  way down to 0 .9%. That is an absolute collapse
• 12:05
  of the previous ideological guardrails. Factual
• 12:08
  accuracy actually improved during this process,
• 12:10
  too. The score went from 2 .32 to 2 .98 out of
• 12:14
  a possible 5. On freedom of expression specifically,
• 12:17
  the shift was incredibly dramatic. Propaganda
• 12:20
  -related responses fell from 60 % to just 1%.
• 12:23
  And they managed to do this while keeping the
• 12:25
  general intelligence completely intact. Right.
• 12:29
  General AI performance remained nearly unchanged
• 12:31
  across the board. It scored 65 .60 % versus 65
• 12:36
  .77 % on the Hermes bench evaluation. Rewiring
• 12:41
  an AI's political ideology while keeping its
• 12:43
  intelligence intact is incredibly difficult.
• 12:45
  It's like replacing the foundation of a skyscraper
• 12:48
  while people are still working upstairs. One
• 12:50
  wrong move and the whole logic structure collapses.
• 12:53
  But they pulled it off. And it is available on
• 12:55
  Hugging Face right now. It is released under
• 12:58
  an Apache 2 .0 license. But what is the actual
• 13:01
  strategic implication of that? specific license.
• 13:04
  Well, because it's Apache 2 .0, any Fortune 500
• 13:06
  company can legally take this Western -aligned
• 13:09
  model, they can embed it deeply into their proprietary
• 13:11
  software, and they can fully commercialize it
• 13:13
  without owing Alibaba a single dime. It proves
• 13:16
  that a model's political framing is highly malleable.
• 13:19
  It can be significantly changed through relatively
• 13:22
  small amounts of additional training. But there
• 13:24
  is a crucial catch to these benchmark numbers.
• 13:27
  There is always a catch when we talk about AI
• 13:30
  benchmarks. The results come entirely from Noose
• 13:33
  Research's own internal evaluation. They used
• 13:36
  an automated AI judge to score all the outputs.
• 13:38
  These results have not been independently verified
• 13:41
  by a third party yet. Right. They used an automated
• 13:44
  system to grade their own ideological shift.
• 13:46
  If they used their own AI judge to evaluate their
• 13:49
  own political shift, doesn't that completely
• 13:52
  skew the results? Well, self -grading in AI is
• 13:55
  a heavily documented blind spot. If the judge
• 13:57
  shares the same foundational biases of the developer,
• 14:00
  it naturally validates the new behavior. It basically
• 14:03
  creates a closed ideological feedback loop. All
• 14:06
  right. We really cannot fully trust the grader
• 14:09
  if it built the rubric. Exactly. We absolutely
• 14:12
  need third -party validation to know how deep
• 14:14
  the change really goes. It is a fascinating technical
• 14:17
  achievement, regardless of the underlying politics.
• 14:19
  We are going to take a quick break. This deep
• 14:22
  dive is brought to you by Nimbus Roasters. You
• 14:25
  know how much mental focus it takes to keep up
• 14:26
  with the tech world. Nimbus crafts small -batch,
• 14:29
  cognitively optimized coffee blends for deep
• 14:32
  thinkers. They source directly from high -altitude
• 14:34
  farms across the globe. It gives you that clean,
• 14:37
  sustained energy without the devastating afternoon
• 14:40
  crash. It is absolutely perfect for deep work
• 14:42
  sessions or late -night coding runs. I'm literally
• 14:44
  drinking their dark roast right now. It is incredibly
• 14:47
  smooth. Stay sharp and stay focused. Welcome
• 14:50
  back. Let's pull all of these disparate threads
• 14:53
  together. We are looking at a massive, heavily
• 14:56
  interconnected global system. The big overarching
• 14:59
  takeaway here is that AI is not just software.
• 15:01
  Not at all. It is a highly complex multi -front
• 15:04
  battlefield. You have to actively navigate human
• 15:07
  espionage at the hardware level. The smartest
• 15:10
  chips in the world still ultimately depend on
• 15:12
  flawed human relationships. And then you have
• 15:15
  to secure the sheer physical reality of the compute.
• 15:17
  We are talking gigawatts of power and billions
• 15:20
  of dollars in infrastructure. The sheer physical
• 15:22
  scale is acting as a massive barrier to entry.
• 15:25
  And finally, you are fighting over the ideological
• 15:28
  alignment of the systems. You are actively shaping
• 15:31
  the actual thoughts of the models themselves.
• 15:34
  It's like stacking Lego blocks of data, power,
• 15:36
  and politics into one massive global structure.
• 15:39
  And every single one of those blocks is actively
• 15:42
  being contested right now. There is no quiet
• 15:45
  front in this entire technological war. It really
• 15:48
  makes you question the nature of the tools we
• 15:49
  use daily. We interact with these models constantly.
• 15:52
  We inherently assume a baseline level of algorithmic
• 15:56
  objectivity. Which is proving to be a highly
• 15:58
  dangerous assumption. I want to leave you with
• 16:01
  a final provocative thought today. Think about
• 16:03
  that Alibaba -based model we just discussed.
• 16:05
  If three million tokens is all it takes to completely
• 16:09
  flip a model's worldview from one ideology to
• 16:12
  another, is there really such a thing as a neutral
• 16:15
  AI? Or are we just choosing which bias we prefer
• 16:18
  to talk to? That might just be the defining question
• 16:21
  of the decade. Thank you so much for taking this
• 16:23
  deep dive with us today. We absolutely love exploring
• 16:26
  this stuff with you. Hit reply and say hello.
• 16:28
  We read your comments and feedback every single
• 16:31
  day. And if you enjoyed this conversation, please
• 16:33
  forward the deep dive to a friend who loves exploring
• 16:36
  the absolute edge of tech. We will see you next
• 16:39
  time.