Latest / AI Fire Daily / #605 Neil: GPT 6 Leaks Heat Up As Gemini 4 Moves Into The AI Race

Transcript

#605 Neil: GPT 6 Leaks Heat Up As Gemini 4 Moves Into The AI Race · AI Fire Daily

• 0:00
  So while we are happily playing with chatbots
• 0:02
  that cost pennies a day to use, there's an actual
• 0:06
  war happening to sex silence. The big labs, OpenAI
• 0:09
  and Google. They are secretly fighting over the
• 0:12
  literal silicon required to survive this next
• 0:14
  decade. Oh, absolutely. The race for GPT -6 and
• 0:17
  well, Gemini 4. It's getting incredibly messy
• 0:20
  right now behind closed doors. Yeah. And the
• 0:22
  thing is, it's definitely not just about who
• 0:24
  builds the smartest model anymore, which is what
• 0:26
  we are all used to. The entire physical battlefield
• 0:29
  is just shifting rapidly underneath our feet.
• 0:31
  It really is. Welcome to this deep dive. Our
• 0:35
  mission today is to unpack these massive, unconfirmed
• 0:39
  leaks and hidden strategies that are defining
• 0:40
  the next era of AI for you. Because there's a
• 0:44
  lot to unpack. There really is. We are exploring
• 0:46
  OpenAI's secret Astra and Bell models today,
• 0:49
  their custom jalapeno hardware, and Google's
• 0:52
  massive Gemini 4 counterattack, and ultimately
• 0:54
  why the real battle is moving from benchmark
• 0:57
  scores to cold hard economics. Right. Because
• 1:00
  the benchmark scores you see on Twitter, they're
• 1:02
  honestly just a distraction at this point. The
• 1:04
  actual war is happening deep, deep inside these
• 1:07
  server farms. Yeah. To understand where this
• 1:09
  industry is going, we first have to look at the
• 1:11
  massive puzzle pieces OpenAI is quietly hiding
• 1:14
  in plain sight. Let's start with Astra. OpenAI
• 1:18
  openly calls this their next major model, right?
• 1:20
  Yeah. They do. And the internal results for Astra
• 1:23
  are already leaking out everywhere. An internal
• 1:26
  version reportedly solved 10 long -standing computer
• 1:29
  science problems. Wow. Wait, long -standing?
• 1:31
  Like... Problems humans haven't solved. Well,
• 1:33
  incredibly difficult math and logic problems
• 1:36
  that previous AI models just completely failed
• 1:38
  at. So the community quickly connected Astra
• 1:40
  directly to the GPC6 architecture. It's seen
• 1:44
  as one of OpenAI's biggest pre -trains since
• 1:46
  GPT 4 .5. Hmm. Solving novel math problems implies
• 1:50
  the model is actively reasoning. It isn't just
• 1:53
  reciting patterns it memorized from human data.
• 1:55
  Exactly. It's a huge leap. But Astra is really
• 1:57
  only the first piece of this puzzle. The leaks
• 1:59
  also mention another massive project named Bell.
• 2:03
  Yeah, Bell makes this entire roadmap so much
• 2:05
  more fascinating. Bell is reportedly the direct
• 2:09
  successor to the Doug model, and the leaks suggest
• 2:12
  it has over 10 trillion total parameters. 10
• 2:16
  trillion. Beat. Let's pause and just think about
• 2:19
  the scale of Bell for a second. Let's define
• 2:21
  parameters here. Parameters are the internal
• 2:24
  connections a model uses to map patterns. Right,
• 2:27
  like the synapses in a brain. Exactly. And 10
• 2:30
  trillion is a staggering number to mentally comprehend.
• 2:33
  I mean, we need a way to visualize this leap.
• 2:35
  If a one trillion parameter model is like stacking
• 2:38
  Lego blocks of data to build a life -sized virtual
• 2:41
  city. Which is already massive, by the way. Right,
• 2:43
  with complex streets and buildings. But Bell,
• 2:46
  at 10 trillion, isn't just a sprawling megacity.
• 2:48
  It is more like an entirely new physics engine
• 2:51
  driving the universe. Yes, that's a great way
• 2:53
  to put it. And that unprecedented density, it
• 2:55
  changes everything. The model just has so many
• 2:58
  intricate internal connections available that
• 3:00
  it can start abstracting concepts we didn't even
• 3:02
  intentionally train it on. Reports say OpenAI
• 3:05
  plans to use heavy reinforcement learning on
• 3:07
  Bell to build stronger systems over time. So
• 3:10
  naturally, a 10 trillion parameter model reignites
• 3:13
  all those serious talks about AGI timelines.
• 3:16
  The research pipeline is clearly steps ahead
• 3:18
  of what we can access today. Two secs silence.
• 3:22
  But I have to ask about this specific parallel
• 3:23
  strategy. Why would OpenAI run two massive, distinct
• 3:29
  projects like Astra and Bell at the same time?
• 3:32
  Why wouldn't they just build one giant GPT -6?
• 3:35
  Well, because you need completely different environments
• 3:37
  for different strategic goals. Astra looks like
• 3:40
  the immediate product evolution, right, for consumer
• 3:42
  applications. But Bell looks like the foundational
• 3:45
  research engine pushing the absolute theoretical
• 3:47
  limits. Oh, I see. Yeah. So you use the massive,
• 3:50
  expensive scale of Bell to figure out what's
• 3:51
  actually possible. And then you distill those
• 3:53
  massive breaks. throughs into something efficient
• 3:55
  that you can actually run, which is Astra. So
• 3:57
  Astra is the engine and Bell is the massive testing
• 4:00
  ground. Precisely. Because you just cannot deploy
• 4:02
  a 10 trillion parameter model to everyday users,
• 4:05
  it would instantly crash your entire server network.
• 4:08
  Which brings us perfectly to the core physical
• 4:11
  limitation here. Building a 10 trillion parameter
• 4:14
  model like Bell creates a massive new problem.
• 4:17
  You can't just run that on standard infrastructure
• 4:19
  without going bankrupt. No, absolutely not. The
• 4:22
  fundamental bottleneck isn't the software anymore.
• 4:25
  It's the silicon. Yeah, and this naturally leads
• 4:28
  us to the hardware leaks, which are wild. Bigger
• 4:31
  models make compute costs skyrocket. If OpenAI
• 4:35
  pays too much for every user request, they lose
• 4:38
  money constantly. So GPT -6 is financially impossible
• 4:42
  to run on normal hardware. Right, and that is
• 4:44
  why we need to talk about the jalapeño chip.
• 4:46
  OpenAI is supposedly building this custom silicon
• 4:49
  to take the pressure off. It is essentially an
• 4:52
  inference chip. And to be clear, inference is
• 4:54
  the live process of computing a response for
• 4:56
  the user. Right, and custom silicon is vastly
• 4:59
  different from off -the -shelf commercial GPUs
• 5:01
  like what NVIDIA sells. Those standard GPUs are
• 5:04
  built to do many different graphical tasks reasonably
• 5:06
  well, but Jalapeno strips all that away to focus
• 5:09
  purely on rapid matrix math for AI. Just specialize
• 5:12
  for the exact model. Exactly. And if jalapeno
• 5:15
  works as expected, the practical improvements
• 5:17
  are just huge. I mean, the rumored specs say
• 5:19
  interactions could be 2 .1 to 4 .1 times faster.
• 5:23
  Wow. And end -to -end latency improving by 1
• 5:26
  .7 to 3 .6 times. That is a generational leap
• 5:29
  in speed. Two sec silence. Whoa. I mean, imagine
• 5:34
  scaling a 104 .3x leap in performance per watt
• 5:38
  across millions of users globally. that physical
• 5:42
  efficiency fundamentally changes the entire math
• 5:44
  of the company. It really does. It suggests OpenAI
• 5:46
  is preparing the physical infrastructure for
• 5:48
  this next generation. The intelligence race is
• 5:51
  moving deep into the hardware. But does this
• 5:53
  mean OpenAI is fundamentally transforming from
• 5:55
  a software lab into a heavy hardware manufacturer
• 5:57
  just to survive its own growth? Well, yeah. They
• 6:00
  almost don't have a choice at this extreme scale.
• 6:02
  You cannot rely entirely on off -the -shelf hardware
• 6:05
  when your compute needs are just unprecedented
• 6:07
  in human history. Custom silicon gives you the
• 6:09
  exact thermal efficiency your specific architecture
• 6:12
  needs. Right. To afford the intelligence, they
• 6:13
  have to own the metal beneath it. Exactly. They
• 6:16
  have to control the whole stack. But, you know,
• 6:18
  Google is taking a very different angle on this
• 6:20
  exact same problem. Right. While OpenAI builds
• 6:23
  custom chips to support its massive brain, Google
• 6:26
  is focused on expanding their model's working
• 6:28
  memory in toolkits. Google DeepMind has basically
• 6:31
  confirmed pre -training for Gemini 4 has started.
• 6:35
  Yeah, and if the leaks are right, this isn't
• 6:36
  just a number bump. Gemini 4 is a massive architectural
• 6:40
  upgrade. They are reportedly targeting a 1 .5
• 6:44
  million token context window. We should define
• 6:47
  that. The context window is the active working
• 6:49
  memory a model holds during a session. And I
• 6:52
  had to admit, I still wrestle with prompt drift
• 6:55
  myself, even on short tasks. Oh, totally. We
• 6:57
  all do. something, and 20 minutes later it completely
• 7:00
  forgets the original rules you established. Yeah,
• 7:03
  it's super frustrating. But a 1 .5 million token
• 7:05
  window fundamentally changes how you interact
• 7:07
  with the system. You aren't just feeding it a
• 7:09
  short PDF anymore. You can drop an entire company's
• 7:12
  massive code base into its active memory instantly.
• 7:15
  That allows the model to see the entire structural
• 7:18
  forest, not just the individual trees. And...
• 7:21
  The reported upgrades also include much stronger
• 7:25
  browser and terminal capabilities, right? Right.
• 7:27
  Terminal capabilities are where this gets incredibly
• 7:30
  powerful. The model doesn't just write code and
• 7:32
  politely wait for you to copy paste it. It can
• 7:35
  actually execute the code itself, read the error
• 7:38
  logs, and rewrite it. Which is wild. Yeah. And
• 7:41
  internal tests allegedly show it beating current
• 7:44
  frontier models entirely behind closed doors
• 7:46
  in these multi -step agent tasks. I do have to
• 7:49
  push back slightly on this, though. Why does
• 7:51
  a 1 .5 million window in terminal access matter
• 7:54
  more to an everyday user than just having a smarter
• 7:57
  chatbot? Because a brilliant chatbot still requires
• 8:00
  you to act as the exhausted manager. You have
• 8:02
  to feed it small tasks, check its work, and stitch
• 8:05
  the results together yourself. But terminal access
• 8:08
  in a massive window let the AI manage the entire
• 8:11
  project autonomously. Got it. It shifts from
• 8:13
  answering questions to actually doing the work
• 8:15
  for you. Precisely. It becomes a digital worker.
• 8:18
  We're going to take a brief pause here, but stick
• 8:19
  around. Yeah, we'll get into the brutal economics
• 8:22
  of this right after the break. We will be right
• 8:24
  back, sponsor. And we are back. Beat. So big
• 8:30
  context windows and custom silicon both point
• 8:32
  to the same ultimate reality. The next race won't
• 8:35
  be won on a simple benchmark chart. It is a fight
• 8:37
  over who can run powerful models at a price people
• 8:40
  can actually afford. This is literally the most
• 8:42
  crucial part of the AI landscape right now. As
• 8:45
  GPT -6 and Gemini 4 get larger, compute capacity,
• 8:48
  inference speed, reliability, and cost performance
• 8:51
  are the new battlegrounds. And we already see
• 8:54
  this economic pressure playing out perfectly
• 8:57
  with DeepSeq V4 Flash. Oh man, DeepSeq is the
• 9:00
  prime example. It proves you don't need to win
• 9:02
  every single benchmark if you are incredibly
• 9:03
  cheap. Right. Its scores are solid, though, like
• 9:06
  82 .7 on Turtle Bench 2 .1 and 54 .4 on DeepSWE.
• 9:10
  Right, they are respectable. But the real shock
• 9:12
  is their insanely low API price. It was once
• 9:15
  as low as 14 cents per 1 million input tokens.
• 9:17
  14 cents. Yeah. And 28 cents per 1 million output
• 9:21
  tokens. That is basically giving the compute
• 9:23
  away for free compared to everyone else. When
• 9:25
  compute is that cheap, you stop treating intelligence
• 9:28
  as a scarce resource. You don't just run one
• 9:31
  prompt. You can afford to set up 10 ,000 autonomous
• 9:34
  agents working simultaneously. Two secs silence.
• 9:38
  Now contrast that highly efficient strategy with
• 9:41
  Anthropic right now. Yeah, Anthropic is spending
• 9:44
  an absolute fortune on compute to stay in this
• 9:47
  race. They just signed a massive $45 billion
• 9:50
  infrastructure deal with Nescale for capacity.
• 9:53
  $45 billion? That is unbelievable. It's staggering.
• 9:57
  But if OpenAI and Google move fast, Anthropic
• 10:00
  has to keep up with both the quality and the
• 10:02
  incredibly expensive compute required to serve
• 10:04
  it. But will good enough and dirt cheap, like
• 10:07
  DeepSeek, ultimately defeat... Brilliant, but
• 10:09
  expensive? Like GPT -6 in the long run? History
• 10:12
  says yes, absolutely. The premium models will
• 10:15
  exist for complex scientific tasks, sure. But
• 10:17
  the vast majority of everyday business workflows
• 10:19
  just need reliable, cheap automation. Companies
• 10:22
  will route 90 % of their queries to the cheapest
• 10:24
  model that gets the job done. Makes sense. Mass
• 10:26
  adoption always follows the cheapest, most reliable
• 10:29
  utility. Exactly. Which brings us right back
• 10:32
  to why OpenAI desperately needs that jalapeno
• 10:35
  chip to lower their own costs. We have painted
• 10:37
  a picture of an incredible, expensive, futuristic
• 10:40
  battlefield today. Beat. But I think it is time
• 10:43
  to bring the hype back down to Earth and look
• 10:45
  at the actual confirmed facts. Yeah, we have
• 10:47
  to be careful. The leaks sound amazing, but many
• 10:49
  important details are totally unconfirmed. Is
• 10:52
  Astra really GPT -6? or just a stepping stone?
• 10:56
  Does Bill actually have 10 trillion parameters?
• 10:59
  Will Jalapeno be a full infrastructure replacement
• 11:02
  or just a tiny specialized fraction of their
• 11:05
  hardware? And on Google's side, you know, the
• 11:07
  1 .5 million context and internal benchmark dominance
• 11:10
  are still just leaks. We need public testing.
• 11:12
  The Internet is honestly steps ahead of the official
• 11:15
  announcements right now. The final reality depends
• 11:17
  purely on what is actually released. So what
• 11:20
  single piece of hard evidence should the listener
• 11:22
  be watching for to know which of these rumors
• 11:24
  is actually true? Honestly, just watch the pricing
• 11:28
  when the public API finally drops. If OpenAI
• 11:31
  releases GPT -6 and the API cost is surprisingly
• 11:35
  low, that's your hard evidence that the jalapeno
• 11:37
  chip is real and working. You can't fake price.
• 11:40
  So we wait for the public API release to see
• 11:42
  the true cost. Yeah. That's the ultimate moment
• 11:45
  of truth. Let's pull all of this together, Beat.
• 11:47
  What is your core synthesis of all these leaks?
• 11:51
  The big idea is that the next AI race is already
• 11:53
  crowded, but it's no longer about just being
• 11:56
  the smartest. It's an infrastructural war now.
• 11:59
  Whoever can make a genius -level model fast,
• 12:01
  reliable, and cheap at global scale, that's who
• 12:04
  wins the year. The battle lines are already being
• 12:06
  permanently drawn in the silicon. Two secs silence.
• 12:08
  I encourage you to explore these sources on your
• 12:11
  own and watch how these labs handle their infrastructure.
• 12:14
  And I will leave you with this final provocative
• 12:16
  thought. If models eventually reach 10 trillion
• 12:18
  parameters, unlocking profound new capabilities,
• 12:21
  but custom inference chips make running them
• 12:23
  practically free, what happens to human creativity
• 12:26
  when supreme intelligence becomes as cheap and
• 12:29
  invisible as electricity running through your
• 12:31
  walls? Beat. Something to think about. Thank
• 12:34
  you for taking this deep dive with us today.
• 12:36
  Out to your own music.