Latest / AI Fire Daily / 🎙️ EP 334: Google Releases Gemini 3.7 Flash & Michael Burry Issues Warning on Nvidia

Transcript

🎙️ EP 334: Google Releases Gemini 3.7 Flash & Michael Burry Issues Warning on Nvidia · AI Fire Daily

• 0:00
  We have AI models currently processing text 14
• 0:04
  times faster than standard speeds. Yeah, it is
• 0:07
  wild. They are actively diagnosing complex canine
• 0:10
  cancer. Right. Yet at the same time, major tech
• 0:13
  companies are quietly killing off their flagship
• 0:15
  AI features. It is a weird contradiction. It
• 0:18
  really is. Welcome to this custom -tailored deep
• 0:21
  dive. Glad you are here with us. We are exploring
• 0:24
  the intersection of a raw compute power and brutal
• 0:27
  business realities today. We really do have a
• 0:29
  fascinating journey ahead for you. We absolutely
• 0:31
  do. We are looking at how sheer speed is fundamentally
• 0:35
  enabling autonomous AI agents. And that velocity
• 0:39
  is triggering a massive financial ripple effect
• 0:41
  across the whole industry. We will also look
• 0:44
  at some stark warnings from prominent investors.
• 0:46
  Yeah, some heavy stuff there. And finally, we
• 0:48
  will see how one man leveraged all of this raw
• 0:50
  tech. To try and save his dog's life. Exactly.
• 0:53
  To truly understand what artificial intelligence
• 0:57
  is capable of today, you first have to grasp
• 1:00
  the underlying physics of it. You really do.
• 1:02
  You have to understand the velocity. You also
• 1:04
  need to look at the plunging costs of the engines
• 1:07
  powering it all. Yeah, and the clearest example
• 1:09
  of this just dropped. Google released Gemini
• 1:11
  3 .7 Flash. Which... Well, what is wild is that
• 1:15
  this came out exactly three weeks after version
• 1:19
  3 .6. Just three weeks. Three weeks. That is
• 1:22
  a blisteringly fast, almost aggressive release
• 1:25
  cycle. It really is. They are clearly focusing
• 1:29
  heavily on complex coding and deep web development
• 1:32
  here. Let us look at the actual specs for a moment.
• 1:34
  Okay. It has a one million token context window.
• 1:38
  Meaning the amount of text the AI can remember
• 1:41
  at once. Exactly. And it can generate up to 64
• 1:44
  ,000 output tokens. In a single prompt. Yeah,
• 1:47
  which is just massive. We are talking about capabilities
• 1:49
  that are expanding horizontally now. It can generate
• 1:53
  complete functional websites from scratch. Right.
• 1:55
  It can look at a static screenshot and perfectly
• 1:58
  rebuild the entire user interface. It turns flat
• 2:01
  PDFs into interactive web pages. Yeah. And it
• 2:04
  coordinates multiple AI agents in the background
• 2:06
  to pull all this off. Beat. It is fascinating
• 2:09
  when you look at the benchmarks, too. Oh, for
• 2:11
  sure. It scored 43 .6 % on Frontier Code 1 .1.
• 2:16
  Which places it ahead of CloudSonic 5. Right.
• 2:19
  It even beats GPT 5 .6 Terra in that specific
• 2:23
  test. Yeah. Though it is worth noting, it still
• 2:26
  trails Terra slightly on long horizon coding.
• 2:29
  You know, the really complex multi -step software
• 2:32
  architecture tasks. Right. Exactly. But honestly.
• 2:36
  The real story here is not just the benchmark
• 2:38
  score. No, it is the raw speed. And it is the
• 2:41
  cost. The economics are shifting dramatically.
• 2:44
  They really are. Google is engaging in a blatant
• 2:46
  price war right now. Oh, absolutely. Till the
• 2:48
  end of 2026, their API is aggressively cheap.
• 2:52
  Like, shockingly cheap. It is 75 cents per million
• 2:55
  input tokens. Right. And output tokens are $3
• 2:59
  .75. Yeah, and they plan to double those prices
• 3:02
  in 2027. It is a classic move. Totally. They
• 3:05
  want to capture the entire... developer ecosystem
• 3:06
  right now while it is cheap they are forcing
• 3:09
  everyone else to lower their prices to compete
• 3:11
  two sec silence but the speed is what i find
• 3:14
  most profound yeah gemini 3 .7 hits around 340
• 3:18
  tokens per second which is fast that is like
• 3:21
  watching a machine read an entire novel in a
• 3:23
  single heartbeat it is nuts yet open ai just
• 3:26
  countered oh they definitely did they launched
• 3:28
  an ultra fast mode for gpt 5 .6 sol it is completely
• 3:32
  absurd how fast is it Salon. pushes output up
• 3:37
  to 750 tokens per second. Wow. That is 14 times
• 3:41
  faster than what we consider standard speed today.
• 3:44
  14 times. Yeah. It is not even reading anymore.
• 3:47
  It is just downloading a matrix of text instantly.
• 3:50
  Does this 14x speed increase actually change
• 3:53
  how we use AI or is it just a vanity metric?
• 3:56
  Oh, it completely changes the paradigm. Speed
• 3:58
  is everything. Right. When a model responds instantly,
• 4:01
  you stop waiting for a search engine to give
• 4:02
  you answers. You start having a real -time...
• 4:05
  zero -latency conversational thinking process
• 4:07
  with the machine. So zero -latency transforms
• 4:10
  AI from a search engine into a real -time collaborator.
• 4:14
  Exactly. It changes from a tool you query into
• 4:17
  a partner you brainstorm with. Beat. But because
• 4:20
  these models are now so incredibly fast and so
• 4:23
  cheap, their fundamental nature is shifting.
• 4:26
  They do not just generate passive blocks of text
• 4:29
  anymore. Right. They're becoming independent
• 4:30
  agents. Yeah. They can actively navigate the
• 4:33
  web for you now. This is where that massive speed
• 4:35
  actually translates into physical action. Exactly.
• 4:39
  Think about Claude Cowork. It just got a massive
• 4:42
  update. It can now take over repetitive tasks
• 4:45
  directly inside your Chrome browser. It is a
• 4:48
  surreal concept. It really is. It is like handing
• 4:51
  a physical mouse over to a ghost. That is a great
• 4:55
  way to put it. It literally takes over the browser.
• 4:57
  It clicks dropdowns. It types in forms. It scans
• 4:59
  webpages. All in real time. Yeah. Gemini 3 .7
• 5:03
  is doing the exact same thing with rebuilding
• 5:05
  interfaces. Right. And because of this, we are
• 5:07
  seeing a sudden flood of new, highly empowered
• 5:11
  tools. King CLI, for example. Yeah, that one
• 5:15
  is interesting. It allows for natural language
• 5:16
  browser testing. You just tell it what to test
• 5:19
  in plain English and it runs it. It actually
• 5:22
  returns path or fail results based on real browser
• 5:25
  interaction. Then you have tools like Edo. Edo
• 5:28
  is fascinating. It catches software bugs using
• 5:32
  actual runtime evidence. Right. And that is a
• 5:34
  huge shift. Yeah. Traditional static analysis
• 5:36
  tools and even older AI reviewers, they just
• 5:40
  look at the code. They guess where a bug might
• 5:42
  be. Exactly. But Edo actually executes the code.
• 5:45
  It runs the application before reviewing it.
• 5:47
  Right. It finds the bug because it literally
• 5:50
  watched the program crash. And then there is
• 5:53
  Nufos. It provides a shared workspace for these
• 5:56
  AI agents to collaborate. They investigate production
• 5:59
  issues and operate systems together. And I have
• 6:02
  to admit, I still wrestle with prompt drift myself.
• 6:05
  Oh, sure. You know, where the AI slowly wanders
• 6:08
  off topic over time. Yeah, like a bad game of
• 6:10
  telephone. Exactly. So the idea of letting it
• 6:13
  click around my browser autonomously makes me
• 6:16
  slightly anxious. You are definitely not the
• 6:18
  only one feeling that. There is actually a massive
• 6:21
  undercurrent of anxiety about this. People are
• 6:24
  starting to actively fight back against this
• 6:26
  level of AI control and tracking. We are seeing
• 6:29
  a notable backlash. A new open source tool was
• 6:32
  just released specifically to strip inviolable
• 6:35
  watermarks. Right. It removes these hidden digital
• 6:38
  markers from the outputs of Claude, Gemini, and
• 6:40
  OpenAI. Users want to maintain ownership of their
• 6:43
  content. Yeah, they do not want invisible tracking
• 6:46
  mechanisms embedded in the code or text these
• 6:49
  agents generate. They want the utility of the
• 6:51
  ghost, but they do not want it leaving fingerprints.
• 6:53
  Exactly. Two sec silence. With agents independently
• 6:57
  testing code and browsing, how do we trust their
• 7:00
  runtime evidence over human oversight? Well,
• 7:04
  tools like Edo provide verifiable proof of those
• 7:07
  bugs. Right. It shifts our trust away from the
• 7:09
  AI's subjective opinion of the code and places
• 7:12
  it squarely on the actual hard execution logs.
• 7:15
  Trust the code execution logs, not the AI's subjective
• 7:18
  analysis. Make it prove it. Precisely. The environment
• 7:22
  itself verifies the work. Sponsor. This deep
• 7:26
  dive is brought to you by Workflow Pro. That
• 7:56
  is Workflow Pro. building the infrastructure
• 8:03
  for millions of these autonomous ghosts to move
• 8:05
  our cursors is not cheap. No, it is not. It requires
• 8:07
  astronomical amounts of money. It requires massive
• 8:10
  physical infrastructure to support the compute
• 8:11
  load. And that brings us to a massive financial
• 8:13
  crossroads in the industry today. Are we looking
• 8:16
  at a permanent technological boom or are we inflating
• 8:19
  a massive bubble? Because the scale of the money
• 8:21
  involved right now is truly staggering. Let us
• 8:24
  look at Databricks as the prime example of this
• 8:26
  infrastructure surge. Okay. They are a massive
• 8:29
  player in the data layer required for AI. Right.
• 8:33
  They originally planned to raise $1 billion in
• 8:36
  this current round. But investor demand was so
• 8:38
  utterly rabid, it pushed that number up to $15
• 8:41
  billion. Which is insane. They eventually compromised
• 8:43
  and settled on raising $5 billion. That places
• 8:47
  their total valuation at $190 billion. Right,
• 8:51
  yeah. Their revenue has reached a $7 billion
• 8:54
  annual run rate. Whoa. Imagine scaling to a billion
• 8:59
  queries. It is hard to even picture. The physical
• 9:02
  servers, the cooling systems, the sheer electricity
• 9:05
  required to support that kind of infrastructure
• 9:08
  is just mind bending. It is a staggering flow
• 9:10
  of capital. It really is. But it is important
• 9:13
  to note that not everyone is celebrating these
• 9:15
  massive valuations. No, definitely not. Michael
• 9:18
  Burry, the investor famous for predicting the
• 9:20
  2008 housing crisis, is raising some significant
• 9:24
  red flags. He is looking very closely at NVIDIA's
• 9:27
  recent $500 billion funding deal. Right. He's
• 9:31
  pointing out how interconnected all this funding
• 9:33
  is. Burry stated that this specific situation
• 9:36
  has, in his words, echoes of Enron. That is a
• 9:40
  heavy comparison. It is. Now, we are simply reporting
• 9:43
  his stance here. Right. We are not endorsing
• 9:45
  his financial prediction, nor are we taking a
• 9:48
  side on the market's future. It is just a very
• 9:50
  stark warning from a prominent voice. Exactly.
• 9:54
  regardless of the warnings these companies are
• 9:56
  pushing incredibly hard to find new sustainable
• 9:59
  revenue streams openai for instance is currently
• 10:02
  testing a 100 per month Premium tier. This is
• 10:06
  aimed at their GPT business platform. Yeah. It
• 10:09
  promises five times more usage for their power
• 10:12
  users. Meanwhile, Apple is taking a very different
• 10:14
  financial angle on the AI space. Apple reportedly
• 10:17
  has a nine -figure budget on the table right
• 10:19
  now. Wow. They plan to actually pair publishers
• 10:22
  when Siri uses their content to answer user queries.
• 10:25
  This is a fascinating model. It really is. It
• 10:27
  could become a major new revenue stream for traditional
• 10:29
  media companies. But amidst all of this aggressive
• 10:31
  fundraising, spending, and rapid expansion, There
• 10:34
  is a very quiet reality check happening. Yeah.
• 10:37
  Microsoft is abruptly cleaning house. They are
• 10:40
  actively killing off several of their flagship
• 10:43
  co -pilot AI features starting August 18th. It
• 10:47
  is a surprisingly sudden pruning of their product
• 10:50
  line. It really is. They are shutting down co
• 10:53
  -pilot group chats, the podcast features, labs,
• 10:56
  deep research, and Miko. They are just pulling
• 10:59
  the plug. Yeah. Why is Microsoft killing copilot
• 11:02
  features right when the underlying models are
• 11:05
  getting radically cheaper and faster? Because
• 11:07
  raw capability does not automatically equal product
• 11:11
  market fit. Users want highly specific workflows
• 11:14
  that solve real problems. They do not just want
• 11:16
  generalized AI gimmicks slapped onto their software.
• 11:19
  Raw intelligence doesn't guarantee a good product.
• 11:22
  The market demands highly specific, useful tools.
• 11:25
  Exactly. Utility always wins over novelty in
• 11:27
  the long run. We have talked about the physics
• 11:29
  of token speeds. Yeah. We explored agents taking
• 11:31
  over browsers, and we discussed billion -dollar
• 11:34
  valuations. We covered a lot of ground. We did.
• 11:36
  But what is the actual tangible payoff of all
• 11:39
  this compute power in the physical world? For
• 11:42
  that, we have to look past the screens. Right.
• 11:44
  We have to look at biological applications. Let
• 11:47
  us talk about Paul Cunningham. Okay. He has a
• 11:49
  dog named Rosie. Right. Rosie was diagnosed with
• 11:53
  a highly aggressive form of cancer. Yeah, it
• 11:55
  is a tough story. She went through all the standard
• 11:57
  protocols, she had surgery, she endured chemotherapy,
• 12:02
  and she tried immunotherapy. And unfortunately,
• 12:04
  all of those traditional treatments failed. The
• 12:07
  veterinarians gave Rosie only a few months to
• 12:09
  live. It is a devastating position for any pet
• 12:12
  owner to be in. Beat. This is where the story
• 12:15
  profoundly intersects with everything we have
• 12:18
  discussed today. Cunningham did not accept that
• 12:20
  timeline. No, he did not. He leveraged the raw
• 12:23
  speed of these new AI tools. He used ChatGPT,
• 12:27
  he used AlphaFold, and he used advanced genomic
• 12:30
  analysis. He used those models to analyze Rosie's
• 12:34
  specific, unique biology. AlphaFold is incredible
• 12:37
  for this. It predicts protein structures. So
• 12:40
  they used it to help identify the exact, unique
• 12:43
  mutations that were driving Rosie's specific
• 12:46
  tumor. He then partnered with researchers and
• 12:48
  scientists at the University of New South Wales.
• 12:50
  Together, using the AI's analysis, they helped
• 12:53
  manufacture a personalized mRNA vaccine just
• 12:57
  for Rosie. Which is incredible. It is. To be
• 12:59
  clear, an mRNA vaccine is a genetic messenger
• 13:02
  that teaches the body to fight disease. Right.
• 13:05
  So they essentially taught Rosie's immune system
• 13:08
  exactly what her specific cancer looked like.
• 13:12
  And after administering this custom vaccine,
• 13:14
  the results were stunning. Several of Rosie's
• 13:16
  tumors reportedly shrank significantly. It is
• 13:19
  a deeply moving application of customized biotechnology.
• 13:22
  It really is. And it was so successful in extending
• 13:25
  Rosie's life that it actually sparked an entirely
• 13:27
  new company. It is now a Y Combinator -backed
• 13:30
  startup. It is called Gamgee. Gamgee wants to
• 13:32
  scale this exact process. They want to do this
• 13:35
  for other dogs facing terminal diagnoses. They
• 13:37
  plan to sequence each individual canine tumor.
• 13:40
  Then they use these high -speed AI models to
• 13:43
  identify the specific mutations. Once identified,
• 13:46
  they design a custom mRNA vaccine tailored precisely
• 13:49
  to that dog. They then manage the treatment directly
• 13:52
  in partnership with local veterinarians. It is
• 13:55
  moving fast. Yeah. The company is already running
• 13:58
  clinical trials in Australia. They're accepting
• 14:00
  complex canine cases globally. There are reportedly
• 14:03
  more than 300 dogs on their wait list right now.
• 14:06
  Two sex islands. We must include a very crucial
• 14:09
  caveat here. Yes, definitely. The scientific
• 14:12
  reality is that the Rosie case is still an early,
• 14:15
  isolated example. Researchers have strongly cautioned
• 14:19
  about drawing definitive, sweeping conclusions
• 14:21
  just yet. That is a vital point. Rosie received
• 14:24
  multiple other aggressive treatments before she
• 14:27
  got the custom vaccine. Scientifically, it is
• 14:30
  very difficult to completely isolate. how much
• 14:33
  of her improvement came specifically from the
• 14:35
  mRNA vaccine versus the lingering effects of
• 14:39
  the prior therapies. Given the caveat about Rosie's
• 14:42
  multiple treatments, is the real breakthrough
• 14:44
  the vaccine itself or the speed at which AI analyzed
• 14:48
  her genomics? The breakthrough is the compression
• 14:50
  of time. AlphaFold and these massive context
• 14:54
  window models took what used to be years of biological
• 14:57
  puzzle solving and they shortened it. They shortened
• 15:00
  it into a time frame fast enough to actually
• 15:02
  treat a living breathing patient. Exactly. The
• 15:05
  true breakthrough is time. AI compressed years
• 15:09
  of biological research into mere weeks. Precisely.
• 15:13
  That unprecedented speed is the only thing that
• 15:16
  makes hyper -personalized medicine viable. Let
• 15:18
  us synthesize everything we have unpacked in
• 15:20
  this deep dive today. Okay. Our narrative arc
• 15:22
  started with the raw underlying physics of compute.
• 15:25
  We looked at models downloading text at 750 tokens
• 15:30
  per second. We saw AI agents autonomously moving
• 15:33
  cursors and clicking around Chrome browsers.
• 15:35
  We examined the infrastructure boom. Right. We
• 15:38
  saw companies like Databricks raising $5 billion
• 15:40
  at staggering mind -bending valuations. But we
• 15:44
  also saw the market's reality check. with Microsoft
• 15:47
  quietly pruning features that ultimately lacked
• 15:49
  clear utility. Yeah, that was a stark contrast.
• 15:52
  And it all culminated in a custom, genetically
• 15:55
  tailored cancer vaccine for a dog. It is a powerful
• 15:58
  reminder. It really is. Artificial intelligence
• 16:01
  is scaling vertically in the cloud at breakneck
• 16:04
  speeds. But its most profound, lasting impacts
• 16:07
  are intimately physical. The bridge between digital
• 16:10
  processing speed and tangible biological outcomes.
• 16:13
  is finally being built it is amazing to watch
• 16:16
  i want to leave you with a final thought to mull
• 16:17
  over today sure a single determined owner used
• 16:21
  off -the -shelf ai tools today yeah he used them
• 16:25
  to help engineer a custom cancer vaccine for
• 16:28
  his pet right What happens when those underlying
• 16:32
  AI models become 14 times faster again? It is
• 16:35
  wild to think about. What happens when they act
• 16:38
  as fully autonomous researchers silently navigating
• 16:40
  our medical systems tomorrow? Yeah. Will hyper
• 16:43
  -personalized medicine eventually become as cheap
• 16:46
  and as instant as a simple web search? It is
• 16:50
  a profound question. The future is arriving faster
• 16:52
  than we think. It really is. Thank you for going
• 16:55
  on this deep dive with us today. We appreciate
• 16:57
  you tuning in. Look closely at your own daily
• 16:59
  workflows this week. Yeah, look for the friction.
• 17:01
  See where the raw speed of these new models might
• 17:04
  unlock something that was previously impossible
• 17:06
  for you. OTRO Music.