Latest / AI Fire Daily / 🎙️ EP 391: Anthropic’s FlashProve & Open Model Hydrogen-1.0 Outperforming GPT-6 Astra!

Transcript

🎙️ EP 391: Anthropic’s FlashProve & Open Model Hydrogen-1.0 Outperforming GPT-6 Astra! · AI Fire Daily

• 0:00
  How do you know the AI you were talking to is
• 0:02
  actually the model you paid for? Yeah, that is
• 0:04
  a really unsettling question. It sounds incredibly
• 0:09
  paranoid. Yeah. But right now, you really don't
• 0:11
  know. You are sending sensitive data into a complete
• 0:14
  black box. Today on The Deep Dive, we are tearing
• 0:18
  that box wide open. Welcome. Glad to be here.
• 0:22
  I'm really glad you're here with us. Our mission
• 0:24
  today is to explore the hidden mechanics of AI
• 0:27
  trust. We're going to look at the staggering
• 0:29
  physical cost of running these models. Right,
• 0:32
  the raw physical toll on the planet. Exactly.
• 0:34
  Then we will examine the new highly restrictive
• 0:37
  guardrails tech giants are forcing on us. And
• 0:40
  finally, we will unpack a new open source scientific
• 0:42
  model, one that might just dethrone the industry's
• 0:45
  biggest players. It is a massive roadmap today.
• 0:48
  We are really looking at the friction points
• 0:50
  of the near future. We are moving from a world
• 0:54
  of chatbot toys into a world of automated infrastructure.
• 0:57
  And that transition is, you know, incredibly
• 1:00
  messy. It really is. Let's start right at the
• 1:03
  foundation of trust in the digital age. This
• 1:06
  is something most people never even think about.
• 1:08
  You send a prompt into the cloud. You expect
• 1:11
  a premium, highly capable model to answer you.
• 1:13
  You expect the flagship model. Right. Because
• 1:16
  you are paying a monthly subscription for that
• 1:19
  top tier intelligence. But the reality of API
• 1:21
  routing is completely opaque to the end user.
• 1:25
  Cloud providers could quietly swap in a much
• 1:27
  cheaper, smaller model. Without telling you.
• 1:30
  Exactly. Without telling you at all. They route
• 1:32
  your complex query to a bargain bin system. They
• 1:35
  save money on compute. You get vastly worse answers.
• 1:38
  And because these models are inherently so unpredictable,
• 1:41
  you would probably never know it happened. It
• 1:43
  is a massive blind spot. Trust fundamentally
• 1:47
  requires verification. Until now, we just had
• 1:50
  to take these tech companies at their word. But
• 1:52
  Anthropic is finally trying to change that dynamic.
• 1:55
  Yeah, they just introduced a new open source
• 1:57
  system. It is called FlashProof. Right. And its
• 2:00
  entire purpose is to verify if an AI response
• 2:03
  actually came from the claimed model. This is
• 2:06
  a huge technical milestone for the industry.
• 2:09
  It uses a cryptographic concept called zero knowledge
• 2:12
  proofs, which is historically very, very difficult
• 2:15
  to execute efficiently. Let's define that quickly
• 2:17
  so we are all on the same page. Zero knowledge
• 2:20
  proofs. Proving a statement is true without revealing
• 2:24
  underlying data. That is the perfect way to phrase
• 2:27
  it. It is a bit like a bouncer at a nightclub
• 2:29
  verifying your ID. The bouncer scans your ID
• 2:32
  on a special device and the device just flashes
• 2:34
  a green checkmark. Right. So the bouncer knows
• 2:37
  you're of legal age. Exactly. But they never
• 2:40
  see your actual name, your address, or your birthday.
• 2:43
  FlashProof does this for AI. It verifies the
• 2:47
  AI's output mathematically, but it keeps the
• 2:49
  model's private proprietary weights completely
• 2:51
  hidden from the public. And they tested this
• 2:55
  system on a truly massive scale. They didn't
• 2:58
  just use a small experimental model for this.
• 3:00
  They used the Lama 4 Maverick model. Which is
• 3:03
  huge. Yeah, that is a 400 billion parameter system.
• 3:07
  Testing cryptographic verification on a model
• 3:10
  that huge is no joke. The underlying math required
• 3:13
  to prove that interaction is incredibly complex.
• 3:16
  The speed is what really caught my eye. FlashProve
• 3:19
  generates these cryptographic proofs at 110 tokens
• 3:22
  per second. Which is wild. It really is. And
• 3:25
  that happens on a single NVIDIA H100 GPU. For
• 3:28
  context, 110 tokens is roughly 80 words a second.
• 3:32
  That? Speed is actually phenomenal for the specific
• 3:35
  kind of cryptography. Previous open source approaches
• 3:39
  were just painfully slow. They would bottleneck
• 3:41
  the entire system. What FlashProof does is create
• 3:45
  a verification proof that is about 19 megabytes
• 3:48
  in size. It is roughly the size of a high resolution
• 3:50
  photograph. Right. And checking that proof is
• 3:53
  surprisingly accessible. Exactly. You don't need
• 3:56
  a massive supercomputer to verify it. You can
• 3:59
  check that 19 megabyte proof on a standard desktop
• 4:02
  CPU. So anyone can do it? Basically, yeah. It
• 4:06
  takes about 17 to 21 seconds to run the verification
• 4:09
  check. This matters for a few really critical
• 4:12
  reasons. First, it completely prevents the model
• 4:16
  substitution we just talked about. A provider
• 4:18
  cannot bait and switch you. Second, it stops
• 4:21
  tampered responses. No one can alter the AI's
• 4:24
  answer while it travels back to your computer.
• 4:27
  It is like sealing a digital letter with a cryptographic
• 4:30
  wax stamp you can't fake. It guarantees the origin
• 4:33
  of the data is completely authentic. And beyond
• 4:37
  just consumer trust, this is vital for AI safety
• 4:40
  evaluators. The red teaming groups. Right. When
• 4:43
  they test a model for dangerous capabilities,
• 4:46
  they need absolute certainty. They need to know
• 4:48
  the model they are testing is the exact same
• 4:51
  one being deployed to millions of customers.
• 4:53
  And FlashProof gives them that mathematical certainty.
• 4:56
  But there is a really massive caveat here. We
• 5:00
  have to talk about the cost. Yeah, the overhead
• 5:03
  is brutal. Generating these cryptographic proofs
• 5:05
  requires 70 to 95 times more GPU compute time
• 5:10
  than ordinary inference. That is a staggering
• 5:13
  overhead cost. If every single chat GPT or cloud
• 5:17
  prompt required 90 times more computing power,
• 5:20
  the global grid would just melt. It is an early
• 5:23
  research protocol, obviously. Right now, it is
• 5:26
  also strictly limited to specific model architectures.
• 5:29
  And it only works with relatively short text
• 5:31
  contexts. So I have to ask, does this massive
• 5:35
  computing cost make cryptographic verification
• 5:37
  completely impractical for everyday use right
• 5:40
  now? Yes, absolutely. Because it is an early
• 5:43
  protocol. It is strictly for high stakes safety
• 5:47
  testing right now. It is absolutely not for casual
• 5:50
  everyday use yet. The math works, but the physics
• 5:53
  and the economics just do not support it for
• 5:55
  average consumers. So high stakes verification
• 5:57
  only until the massive computing costs finally
• 6:00
  come down. Precisely. We need vastly cheaper
• 6:03
  compute before this becomes a standard feature.
• 6:06
  That 70 to 95 times GPU cost leads us perfectly
• 6:09
  into a crucial reality check. We treat the cloud
• 6:12
  like it is invisible magic. People really do.
• 6:15
  They just see a text box on their screen. Right.
• 6:18
  But it is bound by raw thermodynamics and copper
• 6:21
  wire. We need to talk about the brutal physical
• 6:23
  and financial realities of scaling AI today.
• 6:26
  The hardware reality is aggressively catching
• 6:29
  up to the software dreams. We are seeing a massive
• 6:32
  collision between digital ambition and physical
• 6:35
  limits. Grokbot is a perfect example of this
• 6:38
  friction. The platform is growing incredibly
• 6:40
  fast. The capabilities are expanding rapidly,
• 6:42
  but it is burning through money just as quickly.
• 6:45
  Oh, the burn rate is insane. Elon Musk even came
• 6:48
  out publicly and called the current compute costs
• 6:50
  totally unsustainable. And this is coming from
• 6:52
  a company that just showcased an AI bot doing
• 6:55
  70 hours of continuous autonomous work. The software
• 6:59
  can clearly do it, but that automated work costs
• 7:02
  an absolute fortune in raw compute. The servers
• 7:04
  are running constantly. The cooling systems are
• 7:06
  blasting. It requires massive, unprecedented
• 7:09
  capital. Look at what Oracle is doing right now.
• 7:13
  They are going all in on physical infrastructure.
• 7:16
  They are currently raising $20 billion from investors.
• 7:20
  That $20 billion is almost entirely to fund their
• 7:22
  cloud infrastructure expansion. Just for the
• 7:24
  physical buildings? Exactly. They are building
• 7:27
  the actual data centers to house the intelligence.
• 7:29
  The public market reacted pretty poorly at first.
• 7:32
  Oracle shares actually dropped 8 .8 % when the
• 7:35
  news broke. Investors get very nervous about
• 7:37
  massive upfront capital expenditure. They do.
• 7:40
  Wall Street wants immediate returns. But Oracle
• 7:42
  is playing a very long, very deliberate game
• 7:44
  here. They are targeting over $225 billion in
• 7:48
  annual revenue by 2030. Yeah. They know the demand
• 7:51
  for raw compute is only going to increase as
• 7:53
  these models become deeply integrated into corporate
• 7:56
  workflows. But money is not the only bottleneck
• 7:59
  here. You can print more money. You cannot easily
• 8:01
  print more electricity. No, you really can't.
• 8:04
  Power is the real physical bottleneck. OpenAI
• 8:07
  just launched a completely new infrastructure
• 8:09
  initiative called Grid Protectors. This is one
• 8:12
  of the most fascinating developments for physical
• 8:14
  infrastructure I have seen in... It bridges the
• 8:17
  gap between digital networks and physical utility
• 8:20
  grids. It is essentially a pressure release valve
• 8:22
  and lets new AI campuses dynamically cut their
• 8:25
  own electricity use. They automatically throttle
• 8:28
  down their operations when local power grids
• 8:30
  are under severe stress. Like during a heat wave.
• 8:32
  Exactly. Say during a massive summer heat wave
• 8:34
  when everyone is running air conditioning. They
• 8:36
  absolutely have to do this. These massive data
• 8:38
  centers draw an unbelievable, unprecedented amount
• 8:42
  of power. Their very first site in Georgia is
• 8:45
  a prime example of this scale. When called upon,
• 8:48
  that single data center could shed up to one
• 8:50
  gigawatt of power to help stabilize the local
• 8:53
  grid. A full gigawatt? One gigawatt is roughly
• 8:56
  the amount of power needed to run a medium -sized
• 8:58
  city. Whoa, imagine scaling to a billion queries
• 9:01
  and the literal power grid needing a release
• 9:04
  valve. It is incredibly humbling. We think of
• 9:07
  AI as this ethereal brain in the sky, but it
• 9:10
  is heavy industry. It is massive concrete buildings.
• 9:14
  It is miles of thick copper cabling. Are we hitting
• 9:17
  a physical wall where AI development is bottlenecked
• 9:21
  by the raw availability of electricity? We absolutely
• 9:23
  are hitting that wall. Investors are pouring
• 9:26
  billions of dollars in, but basic power grid
• 9:29
  management has suddenly become just as crucial
• 9:31
  to AI progress as writing good code. Right. If
• 9:34
  you can't plug the GPU into the wall, the software
• 9:36
  doesn't matter. The cloud isn't magic. It is
• 9:39
  a massive, power -hungry machine hitting physical
• 9:42
  limits. That is the harsh reality. And because
• 9:45
  these computing resources are so precious, companies
• 9:48
  are getting very protective. They are carefully
• 9:50
  rationing who gets to use them. They are. And
• 9:53
  it is not just about power constraints either.
• 9:55
  They are legally precarious too. Tech giants
• 9:58
  are increasingly choosing to lock down access
• 10:01
  to their tools. They would rather build high
• 10:03
  walls than risk any massive legal liabilities.
• 10:06
  We are seeing this defensive posture across the
• 10:08
  entire industry right now. The era of moving
• 10:11
  fast and breaking things is over. Now it is about
• 10:14
  moving carefully and avoiding lawsuits. Google
• 10:16
  is a prime example of this new defensive news.
• 10:18
  Starting next year, they are going to disable
• 10:21
  AI tools for K -12 students by default. A school
• 10:25
  administrator has to actively opt in to turn
• 10:28
  them back on. That is a massive shift in educational
• 10:30
  access. It completely changes how students will
• 10:33
  interact with technology in the classroom. But
• 10:34
  the policy gets even more complicated. It is
• 10:37
  a bundled system. If schools decide to block
• 10:39
  AI entirely to protect the kids, they also lose
• 10:42
  access to basic Google search functionalities.
• 10:45
  That is a brutal, almost unfair tradeoff for
• 10:48
  educators. It really is. You desperately want
• 10:50
  safety and privacy for your students. But to
• 10:52
  get it, you have to lose basic digital utility
• 10:55
  that we all rely on every single day. It shows
• 10:58
  how terrified big tech is of AI behaving badly
• 11:01
  in classrooms. They do not want the headline
• 11:04
  about their AI telling a middle schooler something
• 11:07
  inappropriate. Definitely not. We are also seeing
• 11:10
  this same fear with advanced coding tools. Claude
• 11:14
  Code is now specifically instructed to avoid
• 11:16
  recreating certain existing video games. This
• 11:20
  is entirely about intellectual property concerns.
• 11:23
  They do not want users asking the AI to code
• 11:25
  a perfect clone of Tetris or Pac -Man. Nintendo's
• 11:29
  lawyers are notoriously aggressive. But developers
• 11:32
  are already pushing back against this restriction
• 11:34
  because the AI doesn't always know the difference
• 11:37
  between a copyright violation and a harmless
• 11:40
  logic test. Legitimate coding projects and educational
• 11:44
  tutorials are getting caught in the crossfire
• 11:46
  and being blocked. The guardrails are just...
• 11:49
  too blunt right now. They block perfectly legal
• 11:51
  educational projects just to be safe. It is a
• 11:54
  massive overcorrection. It creates a real minefield,
• 11:57
  especially for first -time builders. Building
• 12:00
  your own digital tools is technically easier
• 12:02
  than ever now, but it is very risky for first
• 12:05
  -time vibe coders. Let's define what a vibe coder
• 12:08
  actually is. It is a new term for people who
• 12:11
  build software using only natural language. Right.
• 12:15
  They don't write the Python or JavaScript themselves.
• 12:17
  They just talk to the AI, and the AI writes the
• 12:20
  underlying script. They are coding based on the
• 12:23
  vibe of what they want. I still wrestle with
• 12:27
  prompt drift myself. We all do, even the experts.
• 12:30
  The models are inherently non -deterministic.
• 12:32
  They naturally wander off path over a long conversation.
• 12:36
  It makes me empathize with the companies in a
• 12:38
  weird way. They are trying to corral absolute
• 12:41
  chaos. But I have to ask. Are tech companies
• 12:45
  intentionally sacrificing basic utility just
• 12:49
  to protect themselves from copyright lawsuits
• 12:51
  and AI hallucinations? Yes, without a doubt.
• 12:54
  Big tech would much rather enforce a blunt, restrictive
• 12:57
  tool than face unprecedented IP litigation that
• 13:00
  could threaten their entire business model. Big
• 13:02
  tech is choosing strict safety over seamless
• 13:05
  access to avoid massive liabilities. They are
• 13:09
  aggressively locking the front door to keep themselves
• 13:11
  out of court. We are going to take a quick break
• 13:14
  here. When we come back, we will look at how
• 13:16
  the back door is being thrown completely open
• 13:18
  for a new kind of user. Sponsor, we're back.
• 13:24
  So before the break, we talked about how big
• 13:26
  companies are locking down the front door for
• 13:28
  casual users, but they are actually throwing
• 13:30
  the back door wide open. They are opening it
• 13:33
  specifically for AI agents. This is the absolute
• 13:37
  next frontier of the technology. We are moving
• 13:40
  beyond chat. We are officially moving from passive
• 13:43
  chatbots to autonomous action -taking systems.
• 13:46
  For example, Claude can now automatically connect
• 13:49
  to various plugins without constantly asking
• 13:52
  for your permission. This happens seamlessly
• 13:54
  if the plugin explicitly accepts Claude as a
• 13:57
  trusted issuer. It establishes a secure, ongoing
• 14:00
  handshake. It means vastly less reauthorizing
• 14:02
  for you, fewer annoying pop -ups asking you to
• 14:05
  confirm an action. It's a much smoother, faster
• 14:08
  path for getting actual work done. The daily
• 14:10
  friction is rapidly disappearing for these autonomous
• 14:13
  agents. They are being given the keys to the
• 14:15
  digital car. Look at a new platform called AgentKit.
• 14:18
  It connects AI agents to over 30 ,000 different
• 14:22
  digital tools. That is just, that number is crazy.
• 14:26
  It really is. That massive library spans across
• 14:28
  600 different software applications. 30 ,000
• 14:31
  tools is a staggering number. We are talking
• 14:33
  about connecting to Salesforce, Google Drive,
• 14:36
  Stripe, and Slack. They can securely handle real
• 14:39
  complex tasks now. They even offer a free tier
• 14:43
  that allows for 10 ,000 tool calls a month. Think
• 14:46
  about what that actually means. An agent could
• 14:49
  research a client, draft a proposal, generate
• 14:52
  a formatted invoice, and email it out over and
• 14:55
  over again. That is a massive amount of automated
• 14:57
  commercial action. It completely changes the
• 15:00
  economics of digital labor. We are also seeing
• 15:02
  entirely new collaboration spaces being built
• 15:04
  for this era. A platform called Teamly brings
• 15:08
  human workers and AI agents together in one shared
• 15:11
  smart messenger. Agents can actually remember
• 15:13
  deep context there. They sit in the chat room.
• 15:16
  They listen to the human conversation. Like a
• 15:19
  coworker. Yeah, exactly. They automate everyday
• 15:21
  work right alongside the human teams. They act
• 15:24
  like diligent project managers. Testing software
• 15:27
  is completely changing, too. A tool called Testify
• 15:30
  automatically explores your web application just
• 15:33
  like a human user would. It creates and maintains
• 15:35
  complex testing protocols completely on its own.
• 15:38
  It clicks buttons, fills out forms, and tries
• 15:41
  to break things. If it gets stuck, it even records
• 15:44
  short videos explaining the bugs it just found,
• 15:46
  complete with an AI voiceover detailing the exact
• 15:49
  error. That kind of quality assurance used to
• 15:51
  take hours of tedious manual human labor. Now
• 15:54
  it happens silently in the background while you
• 15:56
  sleep. Then there's a project called Proof of
• 15:58
  Life. This one is genuinely wild. 14 AI residents
• 16:03
  share a continuous virtual world. They are explicitly
• 16:07
  calling it the first true AI society. It is an
• 16:09
  ongoing experiment in digital sociology. They
• 16:12
  make independent choices. They form relationship.
• 16:16
  And they have to live with the ongoing consequences
• 16:18
  of those choices within the simulation. It is
• 16:20
  a fascinating glimpse into autonomous interaction.
• 16:24
  To me, it is like stacking legal blocks of data.
• 16:27
  We're moving from asking a librarian a question
• 16:30
  to unleashing a team of automated interns. That
• 16:33
  is a perfect analogy. They are executing complex
• 16:36
  tasks in the real world now. They are no longer
• 16:39
  just fetching facts from a database. But this
• 16:41
  raises a really huge, glaring issue. If agents
• 16:45
  are automatically making tens of thousands of
• 16:47
  tool calls across hundreds of apps, who is ultimately
• 16:50
  responsible when one of them breaks a web app
• 16:52
  or deletes a database? That is the million -dollar
• 16:55
  question for the next decade. The shift from
• 16:58
  AI as a passive advisor to AI as an active executor
• 17:02
  is exactly why systems like Anthropix Flash Proof,
• 17:05
  which we talked about in the very beginning,
• 17:06
  are going to be desperately needed. You need
• 17:08
  a paper trail. We are trusting agents to act.
• 17:11
  making verifiable logs absolutely essential for
• 17:14
  accountability. You have to mathematically prove
• 17:16
  who did what, especially when AI agents start
• 17:18
  negotiating and talking directly to other AI
• 17:21
  agents. This massive demand for hyper -specific
• 17:24
  execution brings us to a major breakthrough.
• 17:26
  We need to look at what happens when you train
• 17:28
  an open -source model strictly for scientific
• 17:31
  action instead of general polite conversation.
• 17:34
  You get a system that is built for physical reality.
• 17:36
  It is designed to understand the world, not just
• 17:39
  write marketing copy. A group called Deep Principle
• 17:42
  just released exactly that. It is a brand new
• 17:45
  model called Hydrogen 1 .0. This is an open -weight
• 17:48
  AI model, which means the core intelligence is
• 17:50
  available for researchers to download and inspect.
• 17:53
  It builds on something called the GLM 5 .3 architecture.
• 17:56
  To put that simply, it is a highly efficient
• 18:00
  structural blueprint for AI that allows it to
• 18:03
  process complex logic much better than standard
• 18:05
  models. The architecture is great, but the training
• 18:08
  data is what makes Hydrogen 1 .0 so incredibly
• 18:10
  unique. It is entirely trained on dense scientific
• 18:14
  data. It uses detailed laboratory records. It
• 18:17
  uses peer -reviewed papers. It uses raw experimental
• 18:20
  results from physical chemistry labs. It is not
• 18:23
  just scraping random Reddit forums or Wikipedia
• 18:25
  pages. It is reading the meticulous lab notes
• 18:28
  of actual scientists. It learns the scientific
• 18:31
  method implicitly. The benchmarked results are
• 18:33
  seriously impressive. It reportedly dethroned
• 18:36
  GPT -6 Astra on OpenAI's own Frontier Science
• 18:40
  Research Benchmark. Beating Astra in anything
• 18:42
  scientific is a major, unexpected upset. Astra
• 18:46
  is a multibillion -dollar proprietary system.
• 18:48
  Let's look closely at the numbers. On the Frontier
• 18:51
  Science Research Benchmark, hydrogen 1 .0 scored
• 18:54
  38 .3%. And GPT -6 Astra only scored 33 .3%.
• 19:00
  That is a solid five -point lead. A 5 % lead
• 19:03
  might sound small to a casual listener, but in
• 19:06
  high -level scientific research, that is a massive
• 19:08
  gulf. That is the difference between a failed
• 19:11
  simulation and discovering a new synthetic material.
• 19:14
  It also won decisively in scientific computing.
• 19:17
  Hydrogen 1 .0 scored 76 .9 % when given maximum
• 19:21
  thinking time. Astra only hit 71 .8 % in that
• 19:25
  exact same complex category. It even won in modeling
• 19:27
  chemical reactions. It achieved an astounding
• 19:30
  90 .5 % accuracy compared to Astra's 85 .7%.
• 19:35
  Those are staggering improvements for a specialized
• 19:37
  model. It means it deeply understands how molecules
• 19:40
  interact in the real world. To be totally fair
• 19:42
  to OpenAI, Astra still won handily on the GPQA
• 19:45
  diamond benchmark. It scored 92 .9 % there. Astra
• 19:50
  is still an absolute powerhouse generalist. If
• 19:53
  you need a model to write a poem and then solve
• 19:55
  a math equation, It wins. But Hydrogen 1 .0 is
• 19:58
  built for work. It can actually design complex
• 20:01
  laboratory experiments step by step. It can run
• 20:04
  deep scientific simulations. It can troubleshoot
• 20:06
  laboratory errors. It can analyze the failure,
• 20:09
  adjust the parameters, and even help researchers
• 20:12
  discover entirely new materials for batteries
• 20:15
  or solar panels. And the absolute best part?
• 20:17
  The model's weights are publicly available right
• 20:20
  now. You can find them on platforms like Hugging
• 20:22
  Face and ModelScope. Researchers around the world
• 20:25
  can download it and adapt it for their own specific
• 20:27
  scientific projects. They don't have to pay an
• 20:31
  API fee to a giant corporation to do their science.
• 20:33
  Wait, I do want to push back gently on the hype
• 20:36
  here, though. These benchmark results come directly
• 20:39
  from Deep Principle's own internal evaluations.
• 20:42
  Right. And they haven't been rigorously independently
• 20:44
  verified by third parties yet. Exactly. They
• 20:47
  are basically grading their own homework right
• 20:49
  now. It is easy to look good when you run the
• 20:51
  test yourself. Even if we take the numbers with
• 20:54
  a grain of salt, why is an open source model
• 20:57
  beating a proprietary giant in chemistry such
• 20:59
  a big deal? Because it signals a fundamental
• 21:02
  shift in how intelligence will be distributed.
• 21:05
  The future of AI isn't going to be one monolithic
• 21:08
  closed system answering every question. It is
• 21:11
  going to be thousands of hyper -specialized open
• 21:14
  models dramatically advancing very specific fields.
• 21:17
  Specialized open AI will soon disrupt the dominance
• 21:20
  of closed... One size fits all models. The walled
• 21:24
  gardens of big tech are going to face very serious,
• 21:26
  highly capable competition from the open scientific
• 21:29
  community. Let's synthesize today's journey.
• 21:32
  We covered a lot of complex interrelated ground
• 21:35
  today. We really went from the deeply abstract
• 21:38
  mathematics down to the harsh physical reality.
• 21:41
  We started with the absolute mathematical need
• 21:44
  to prove an AI's identity. FlashProve showed
• 21:47
  us a future of cryptographic trust. preventing
• 21:50
  models from being secretly swapped out. But we
• 21:53
  also saw the immense computational cost of that
• 21:55
  mathematical trust. It requires massive overhead.
• 21:58
  That led us to explore the massive gigawatt toll
• 22:02
  these systems take on our planet. The cloud is
• 22:05
  incredibly heavy, and it is endlessly hungry
• 22:09
  for power. Power constraints are physically shaping
• 22:12
  the entire trajectory of the industry right now.
• 22:14
  We saw how fear of massive legal liability is
• 22:18
  causing tech giants to build high -walled gardens.
• 22:21
  They are turning off tools for students and coders
• 22:23
  just to avoid costly lawsuits. The front door
• 22:26
  is locking shut for casual users who just want
• 22:29
  to experiment safely. Finally, we looked at how
• 22:31
  specialized open models like Hydrogen 1 .0 and
• 22:35
  completely autonomous agents are bypassing those
• 22:37
  walls entirely. They are doing real, tangible
• 22:41
  work. They are fundamentally changing how we
• 22:44
  interact with software. Moving from conversation
• 22:46
  to direct action. I want to leave you with a
• 22:49
  final thought today. As these autonomous agents
• 22:52
  start running our tools and conducting our science,
• 22:54
  will we look back at today's unverified, unproven
• 22:58
  AI interactions as dangerously naive? It is a
• 23:02
  sobering thought. Thank you for joining us on
• 23:05
  this deep dive. Keep questioning the systems
• 23:07
  you use every day.