Latest / Today’s AI News / Google Beats OpenAI 9 to 1 on Math, Mythos Finds 10K Security Holes, DeepSeek Slashes Prices

Transcript

Google Beats OpenAI 9 to 1 on Math, Mythos Finds 10K Security Holes, DeepSeek Slashes Prices · Today’s AI News

• 0:00
  Imagine an Olympic sprinter breaking a world record, only
• 0:03
  for a rival to quietly step onto the track the very next day and
• 0:08
  shatter that new record 9 times over.
• 0:11
  Well, that is exactly what just happened in the world of
• 0:14
  artificial intelligence. Welcome back to today's AI News,
• 0:17
  brought to you by 9X Productions.
• 0:20
  It really is just a stunning sequence of events, you know, It
• 0:25
  completely reframes where we thought the absolute boundaries
• 0:28
  of machine logic were resting. Yeah, so our mission for today's
• 0:31
  Deep Dive is to basically unpack all of this for you.
• 0:34
  We've been pouring over a massive stack of new AI
• 0:37
  developments. Right, giving you the quick hits
• 0:39
  on what's changing at the highest levels of tech.
• 0:41
  Exactly. But more importantly, we really
• 0:43
  want to look at how real people are applying this to their daily
• 0:46
  lives right now. We want to shortcut your path to
• 0:48
  being perfectly informed. So we should just jump straight
• 0:51
  into the biggest story. Let's do it.
• 0:53
  So last week we all saw the headlines open.
• 0:55
  AI solved an 80 year old math problem and urges conjecture.
• 1:00
  Which, I mean, was huge news on its own.
• 1:02
  Totally huge, but literally one day later Google DeepMind
• 1:06
  quietly drops Alpha Proof Nexus and this system autonomously
• 1:12
  solved 9 open errors problems. Nine of them, not just one.
• 1:19
  And we really need to emphasize here that these aren't just like
• 1:22
  difficult algebra equations. These are considered some of the
• 1:25
  absolute hardest unsolved questions in higher mathematics.
• 1:28
  Specifically spanning combinatorics and graph theory,
• 1:31
  right? Exactly.
• 1:32
  Which, just to make sure we're all on the same page here,
• 1:34
  Combinatorics is essentially the math of counting and arranging
• 1:37
  things in complex sets, right? Spot on.
• 1:39
  And graph theory is about networks, you know, studying how
• 1:41
  nodes connect to other nodes. They're incredibly abstract,
• 1:45
  conceptual puzzles. Yeah, and two of the problems
• 1:47
  Alpha Proof Nexus solved have been completely unsolved for 56
• 1:50
  years. Generations of brilliant math
• 1:52
  emissions have tried and failed. And on top of that, Google also
• 1:56
  used the system to prove 44 open conjectures from the Online
• 1:59
  Encyclopedia of Integer Sequences.
• 2:01
  OK wait, I have to stop you there and just push back on the
• 2:04
  logistics of this because when I hear that an AI solves 1/2
• 2:07
  century old math problem, my brain immediately goes to a
• 2:11
  massive supercomputer warehouse. Oh, like thousands of GPU's just
• 2:16
  running red hot. Exactly.
• 2:17
  Just brute forcing a combination lock and burning through a small
• 2:21
  town's worth of electricity. I mean, how much does a 56 year
• 2:25
  old math breakthrough cost nowadays?
• 2:27
  Well, that is the genuinely shocking part of the research
• 2:30
  papers we reviewed. The data shows it only cost a
• 2:33
  few $100 per problem in compute power.
• 2:36
  Wait a few $100 Are you serious? I could buy a nice blender for
• 2:39
  that. I know, right?
• 2:40
  It's wild. How does half a century of
• 2:43
  advanced mathematical struggle equal the cost of a kitchen
• 2:46
  appliance? It basically comes down to this
• 2:48
  brilliant architectural choice. You see, they didn't just ask an
• 2:52
  AI to guess the answer. OK, what did they do?
• 2:55
  They paired a large language model with a system called Lean,
• 2:59
  and Lean is a formal proofs assistant.
• 3:01
  So it's kind of like a ruthlessly strict math teacher
• 3:05
  who gives zero credit unless every single micro step of the
• 3:10
  equation is shown and logically sound.
• 3:13
  That is a erfect analogy. Actually.
• 3:15
  The LLM acts as the highly creative, highly intuitive
• 3:18
  student who throws out a creative hypothesis for how to
• 3:21
  solve the problem. Right.
• 3:23
  An Lean acts as the strict teacher.
• 3:26
  The LOM submits a proposed step and Lean checks it with absolute
• 3:30
  mathematical certainty. And if the logic fails.
• 3:32
  Lean rejects it entirely, but here's the magic part.
• 3:35
  The LLM learns why it failed, adjusts its approach, and then
• 3:39
  tries again. Oh wow, so it's a closed
• 3:41
  iterative loop of formal verification.
• 3:44
  Exactly. It just loops at machine speed
• 3:46
  until a complete proof passes. So it's not brute forcing the
• 3:48
  answer at all. It's essentially brute forcing
• 3:50
  the scientific method, generating hypothesis and
• 3:53
  testing them instantly. That's exactly what it's doing
• 3:55
  now. The researchers didn't know that
• 3:56
  they tested a Simler version of this agent, and while it
• 3:59
  eventually matched the results, it took way more compute and
• 4:01
  actually cost more to run because it went down South many
• 4:04
  inefficient paths that. Makes sense?
• 4:06
  And it's also imortant to clarify the limitations here.
• 4:09
  Problems that require entirely new mathematical constructions,
• 4:14
  meaning inventing a completely new type of math to solve the
• 4:17
  problem. Those are still out of reach.
• 4:19
  But for existing mathematical frameworks, this loop of
• 4:22
  generation and formal verification is just
• 4:25
  revolutionary. Right, because it means the
• 4:27
  bottlenecks for researchers just disappear.
• 4:29
  Yeah. And if you have a system that
• 4:31
  can Florida State verify abstract mathematical logic, you
• 4:34
  essentially have the perfect engine for auditing other
• 4:37
  complex logic based systems. You're just swapping math
• 4:40
  theorems for software architecture.
• 4:42
  Which leads a seamlessly into our next massive development.
• 4:46
  Yes, Entropic just shared the 1st results from Project Glass
• 4:49
  Wing using a model called Claude Mythos Preview and the numbers
• 4:52
  here are honestly staggering. They really are.
• 4:55
  I mean, in just one single month of beta testing, Claude Mythos
• 4:59
  and about 50 partner organizations found over 10,000
• 5:03
  high or critical severity vulnerabilities in software
• 5:06
  code. 10,000 critical vulnerabilities in 30 days.
• 5:09
  That is insane. It is.
• 5:11
  Cloudflare alone used it to find 2000 bugs, and their data show
• 5:15
  the AI's false positive rate was actually better than human
• 5:18
  testers. Better than humans.
• 5:20
  Wow and Mozilla used it to find and fix 271 vulnerabilities
• 5:25
  hidden specifically in Firefox 150.
• 5:27
  Yeah, and Anthropic also took it upon themselves to scan over
• 5:30
  1000 popular open source projects.
• 5:33
  Mythos flagged over 6200 of them as having high or critical
• 5:38
  issues. And what's crazy is that when
• 5:39
  independent human reviewers came in to triage those flags, 62% of
• 5:44
  them, which is nearly 3900 bugs, actually held up as legitimate
• 5:48
  critical threats. So it's essentially automating
• 5:51
  the absolute highest level of cybersecurity auditing.
• 5:54
  Right. And the real world applications
• 5:55
  are immediate, like one partner bank actually used Mythos to
• 5:59
  detect a highly sophisticated exploit and successfully blocked
• 6:03
  a $1.5 million fraudulent wire transfer in real time.
• 6:08
  Which is incredible, but I feel like there's a flip side to
• 6:11
  this. Yeah, I have to play devil's
• 6:12
  advocate here for a second. You are talking about an AI
• 6:15
  digging through millions of lines of code to find critical
• 6:19
  zero day vulnerabilities, right? And for anyone unfamiliar, a
• 6:23
  zero day is basically a secret trapdoor or a fatal flaw in
• 6:26
  software that the original creators don't even know exists
• 6:29
  yet. So if this model is this
• 6:32
  incredibly good at finding those unknown exploits, isn't
• 6:35
  Anthropic basically building a master hacker in a box?
• 6:39
  That is a very valid concern. Like you're giving an AI the
• 6:42
  ability to find the weakest point in any digital armor.
• 6:45
  Why are they keeping this locked up and only releasing it to
• 6:48
  select partners? Well, that raises A vital point,
• 6:51
  and it's exactly why Anthropic is keeping Mythos heavily gated
• 6:54
  right now. Yeah, they openly admit in their
• 6:56
  documentation that no company, not even Anthropic themselves
• 7:00
  currently, he has safeguards strong enough to prevent this
• 7:02
  kind of capability from being misused.
• 7:04
  So it's too dangerous to just release.
• 7:06
  Exactly. If a bad actor gets their hands
• 7:08
  on an autonomous agent that can spot 0 day vulnerabilities
• 7:11
  faster than an entire human security team, the damage to
• 7:15
  global infrastructure could be catastrophic.
• 7:17
  It's the ultimate double edged sword.
• 7:19
  A tool that can perfectly defend a system is naturally also the
• 7:23
  perfect tool to break into it. Yeah, but the broader context
• 7:26
  here is the impending global arms race.
• 7:29
  I mean, Anthropic can gate Mythos all they want, but Open
• 7:32
  AI is rapidly ramping up its own cyber focus models.
• 7:35
  And major Chinese tech players are catching up incredibly fast,
• 7:38
  right? Right.
• 7:40
  These capabilities will inevitably emerge on the broader
• 7:43
  open market, and when they do, the ultimate test won't be
• 7:47
  whether AI can find the bugs. The true test of our digital
• 7:51
  security will be whether the world can deploy AI to patch
• 7:54
  those vulnerabilities faster than adversarial AI can discover
• 7:57
  them. Man, it's a little terrifying to
• 7:59
  think about the geopolitical stakes of an automated
• 8:02
  cybersecurity arms race. But you know, what I find
• 8:04
  absolutely fascinating is the everyday side of this
• 8:06
  technology. Yeah, like the exact same
• 8:09
  underlying reasoning engines that are capable of securing A
• 8:12
  multinational bank are now sitting right on our desktops,
• 8:15
  ready to eliminate the most boring, tedious parts of our
• 8:18
  day. Oh, absolutely.
• 8:20
  The contrast is striking. High leverage delegation is no
• 8:23
  longer restricted to enterprise companies.
• 8:27
  You know the individual now has access to the exact same logic
• 8:30
  engine. Let's actually look at a couple
• 8:31
  of staff use cases from the sources that illustrate this
• 8:34
  perfectly. First we have Mayor, who is a
• 8:37
  content manager. Right the Downloads folder
• 8:39
  story. Yes, his Downloads folder was an
• 8:42
  absolute disaster. Like 100 gigabytes of random
• 8:45
  files, images, PDFs and zip files just accumulated over six
• 8:49
  months. We have all been there, just
• 8:51
  staring at that mess. Exactly.
• 8:53
  But instead of spending an entire Sunday organizing it, he
• 8:56
  just gave an AI assistant called Claude Co work access to the
• 9:00
  folder and basically said help me organize this and what
• 9:03
  happened. In minutes the AI scanned the
• 9:05
  entire 100 gigabytes, arranged everything by file type into
• 9:08
  neat logical folders and deleted all the duplicates.
• 9:12
  Hours of tedious work just gone in minutes.
• 9:15
  It really transforms the computer from a passive storage
• 9:17
  device into an active assistant. And then there's Shubham, an
• 9:21
  editor. His brother-in-law needed a
• 9:23
  Shopify website built for a new business.
• 9:26
  Now Shobum isn't a web developer.
• 9:28
  He doesn't know how to code. But he's clawed.
• 9:30
  Right. Yeah, he used Clawed to act as
• 9:32
  his own on demand Shopify expert.
• 9:34
  He just shared screenshots of what he wanted, linked to some
• 9:37
  design references and typed out the business requirements and
• 9:40
  the AI completely generated the liquid code.
• 9:44
  Which for context, Liquid is the specific proprietary templating
• 9:49
  language that Shopify uses to build its underlying page
• 9:52
  structures. It is highly specialized.
• 9:54
  Right, Shubham didn't need to learn a single word of Liquid.
• 9:58
  The AI wrote the code, troubleshooted some weird theme
• 10:01
  issues, modified the page layouts, and then told him
• 10:04
  exactly where to copy and paste the code inside his Shopify
• 10:07
  dashboard. He just bypassed the community
• 10:10
  forums, bypassed the technical documentation, and went straight
• 10:13
  from concept to a launch store. What both of those examples
• 10:17
  highlight is this massive transition from using AI merely
• 10:20
  as a search engine to using it as an active autonomous agent.
• 10:24
  Yeah. And there's actually a brilliant
• 10:26
  step by step framework in our materials for building your own
• 10:31
  AI taskmaster to manage your daily workflow.
• 10:34
  I love this part of the research.
• 10:36
  Break this down for us, because this is something anyone
• 10:38
  listening can literally set up on their computer today.
• 10:41
  It's surprisingly straightforward actually.
• 10:44
  You start by opening Claude code or a tool like Codex inside a
• 10:48
  dedicated folder on your computer.
• 10:50
  Your first instruction is to ask the AI to create a daily skill.
• 10:54
  OK, a daily skill. Right, you tell it to check your
• 10:56
  Slack messages, your Gmail inbox, and your daily calendar.
• 10:59
  Every single warning before you even sit down.
• 11:01
  So you're basically giving it eyes on all your incoming
• 11:04
  communication streams so it can see what people are asking of
• 11:06
  you. Precisely.
• 11:08
  Then you instructed to prioritize those incoming tasks
• 11:11
  from high to low importance, and you have it.
• 11:14
  Output this prioritized list into a specific markdown file.
• 11:17
  Let's call it mono note dot MD. Got it Mono note dot MD.
• 11:21
  Yeah, and you instruct the AI to format this file with today's
• 11:24
  date at the top, group the tasks logically by project and include
• 11:27
  direct source links back to the original emails or messages.
• 11:30
  Oh, so you don't have to go hunting for them later?
• 11:33
  That is brilliant. Exactly, and you add status
• 11:36
  check boxes so you can visually check things off as you work.
• 11:38
  I love the check boxes. It takes this abstract AI magic
• 11:42
  and turns it into a highly tangible, satisfying To Do List
• 11:46
  totally. And to keep the AI aligned with
• 11:49
  your specific job, you also have the agent create a second file
• 11:54
  called Task rules dot MD. This is where you store your
• 11:57
  personal rules for what constitutes a high priority
• 12:01
  versus a low priority task for your specific role.
• 12:04
  OK, that makes sense. But the real beauty of this
• 12:06
  system happens on day 2. After you run it manually the
• 12:09
  first time, you ask the agent to automate the process.
• 12:12
  From then on, every morning it will review what you checked off
• 12:15
  yesterday, roll over any bump tasks, update the rules based on
• 12:19
  your feedback, and generate a first perfectly prioritized mono
• 12:22
  note list for the day. It's a self cleaning To Do List
• 12:25
  and I actually have to add a pro tip from the sources here that
• 12:28
  just blew my mind. Let's hear it.
• 12:29
  Once you have this system running smoothly, you can create
• 12:33
  a weekly audit skill. You literally have the AI scan
• 12:37
  your own task list at the end of the week, look for repetitive
• 12:40
  things you keep doing manually, and have it suggest which of
• 12:43
  those tasks can be fully automated next.
• 12:45
  Wow. It's essentially an operating
• 12:48
  system for your career that continually optimizes itself.
• 12:51
  It really provides personal operational leverage at a scale
• 12:54
  we simply haven't. Seen before and you know the
• 12:56
  only reason building these daily autonomous workflows is becoming
• 12:59
  so viable for the average person is that the underlying compute
• 13:03
  tech is getting drastically cheaper and faster.
• 13:06
  Which brings us perfectly to today's rapid fire news.
• 13:09
  The speed and cost wars among the major tech companies are
• 13:12
  heating up in a big way. They really are.
• 13:14
  Yeah, DeepSeek V4 Pro just permanently slashed their
• 13:17
  pricing by a massive 75%. We're talking 0.435 dollars per
• 13:24
  million input tokens and 0.87 dollars per million output
• 13:28
  tokens. Which is staggeringly far below
• 13:31
  the closed source rivals. I mean, it's 75% price cut isn't
• 13:35
  just a discount, it changes the fundamental economics of what
• 13:38
  developers can actually build. Because it's practically free.
• 13:42
  Right when the cost of processing logic drops to near
• 13:45
  0, those continuous agent moves we just talked about, like the
• 13:48
  AI taskmaster constantly checking your e-mail, they
• 13:51
  become practically free to run all day long and.
• 13:54
  Google isn't sitting still either, they just dropped Gemini
• 13:57
  3.5 Flash which is 4 times faster than its previous
• 14:01
  versions at half the price. Wow.
• 14:03
  Plus, we're seeing entirely new specialized tools launching,
• 14:07
  like Poultia, which is marketing itself as an AI Co founder that
• 14:10
  plans, builds and operates digital businesses 24/7.
• 14:13
  The barrier to entry for building complex software is
• 14:16
  literally collapsing in real time.
• 14:18
  The ripple effects of this cheap compute are hitting every
• 14:21
  sector. Let's just run through
• 14:22
  everything else happening across the industry right now.
• 14:24
  Sounds good. Perplexity Just open sourced a
• 14:27
  tool called Bumblebee. It's a scanner for Mac OS and
• 14:30
  Linux that checks for risky packages, browser extensions and
• 14:34
  AI tool configurations during supply chain incidents.
• 14:37
  It's essentially an open source immune system for developers.
• 14:39
  Exactly. And on the medical front, NVIDIA
• 14:42
  released a new foundation model called NV Generate Mr. Brain.
• 14:46
  It generates highly realistic synthetic 3D brain MRI scans
• 14:51
  with perfectly mapped annotations.
• 14:54
  Now, the creation of synthetic data is an absolutely massive
• 14:57
  breakthrough for medical AI. For years, researchers have
• 15:00
  faced a severe data bottleneck. Because of privacy laws, right?
• 15:04
  Yes exactly. You want to train an AI to
• 15:06
  detect early stage brain tumors but you can't just feed it
• 15:09
  millions of real patient brain scans because of strict privacy
• 15:12
  laws like I pay, the data is completely locked down.
• 15:15
  So the synthetic data fixes that.
• 15:17
  Right, by generating perfectly accurate synthetic 3D brains,
• 15:21
  NVIDIA is allowing researchers to circumvent the privacy
• 15:23
  bottleneck entirely and train life saving models without ever
• 15:27
  touching a real patient's records.
• 15:28
  That is just incredible. And finally, at the national
• 15:31
  level, the White House just approved $9 billion to help US
• 15:35
  spy agencies acquire advanced AI chips.
• 15:39
  Which is a huge investment. Yeah.
• 15:41
  The driving concern there is basically ensuring they don't
• 15:44
  don't fall behind global adversaries and deploying these
• 15:46
  frontier models. Which ties right back to that
• 15:49
  cybersecurity arms race and the zero day vulnerabilities we were
• 15:53
  just discussing with Anthropic. It really does.
• 15:55
  OK, I need to pause here for a second and just look at this
• 15:57
  whole picture. We've got plummeting compute
• 16:01
  costs, medical research bypassing decades old data
• 16:04
  bottlenecks, the government pouring $9 billion into
• 16:07
  specialized chips. Sounds pretty perfect.
• 16:09
  Right, exactly. It sounds like we are gliding
• 16:11
  into this flawless, frictionless tech utopia, but anytime you
• 16:15
  force abstract, cutting edge software into the messy real
• 16:19
  world, things break. Surely there are massive growing
• 16:22
  pains when businesses force this tech into the real world.
• 16:26
  Oh, absolutely. That is a crucial reality check.
• 16:28
  The data absolutely reflects massive growing pains,
• 16:31
  especially when you look at how traditional businesses operate.
• 16:33
  Like, take the consulting world. McKenzie, you know, one of the
• 16:37
  largest management consulting firms on the planet, is
• 16:40
  currently rethinking its entire billing model.
• 16:42
  Oh wow. Yeah, for decades consulting his
• 16:45
  run on the billable hour. But AI is so efficient at data
• 16:49
  analysis and report generation that it's destroying the value
• 16:53
  of time. Because if an AI can analyze a
• 16:55
  massive corporate data set and draft a 90 page strategic brief
• 16:59
  in 12 seconds, the client isn't going to pay you for the three
• 17:03
  weeks it used to take a team of junior analysts to write it.
• 17:05
  Precisely. Clients know the work takes less
• 17:08
  time now, so they are demanding fees tied to actual business
• 17:12
  outcomes rather than just hours spent at a desk.
• 17:15
  It's a fundamental shift in how knowledge work is valued and
• 17:19
  monetized. That's a massive structural
• 17:21
  change for white collar work, but what about on the
• 17:23
  operational, physical side of things?
• 17:25
  Well, that's where we see the messy friction you mentioned
• 17:27
  most clearly. Starbucks, for example, just
• 17:29
  completely scrapped its AI inventory management system.
• 17:32
  Wait, they scrapped the entire system?
• 17:33
  Yes, after trying to make it work for nine months, they cited
• 17:38
  persistent inventory, miscounts and mislabeled products across
• 17:42
  their North American stores. Ouch.
• 17:44
  Yeah, it's a stark reminder that an AI might be able to solve an
• 17:47
  80 year old math problem in a perfectly controlled digital
• 17:50
  environment, but accurately tracking physical milk and
• 17:54
  coffee beans in a chaotic retail store is a completely different
• 17:58
  challenge. I imagine translating perfectly
• 18:00
  verified digital logic into the physical world is just an
• 18:03
  absolute nightmare. An AI can read a supply chain
• 18:06
  spreadsheet, Florida State, but it can't account for a barista
• 18:10
  accidentally dropping a bottle of vanilla syrup or a delivery
• 18:13
  driver leaving a box in the wrong hallway.
• 18:15
  Exactly, the physical world is chaotic and AI still struggles
• 18:18
  with that translation. It's a profound dichotomy,
• 18:21
  really. AI handles digital complexity
• 18:23
  beautifully, but physical unpredictability breaks it.
• 18:26
  It really does. But you know, while massive
• 18:28
  multibillion dollar corporations like Starbucks are pulling their
• 18:31
  hair out trying to force AI into thousands of physical stores,
• 18:35
  individual users operating from their laptops at home are
• 18:37
  finding life changing leverage. Because the individual isn't
• 18:41
  trying to manage global supply chains, they're just trying to
• 18:43
  manage their own cognitive load. Exactly, and that is perfectly
• 18:47
  illustrated by today's community AI workflow.
• 18:51
  This comes from a listener named Alicia in Fresno, CA.
• 18:54
  Oh I love this one. Yeah, her situation is
• 18:56
  incredibly relatable. She was completely overwhelmed
• 18:59
  trying to compare university programs.
• 19:01
  She's balancing a full time job, she has a specific income level
• 19:04
  to work around, and she found herself just drowning in browser
• 19:09
  tabs from a dozen different university website.
• 19:11
  Trying to compare tuition costs, figure out what's strictly in
• 19:14
  person versus online. It's a classic data synthesis
• 19:18
  problem. You have too many variables
• 19:20
  scattered across too many unstandardized website formats.
• 19:23
  Right, so she used Chachi PT to pull all that data together and
• 19:26
  actually compare the programs and costs side by side in a
• 19:29
  simple table. But here's where it gets really
• 19:31
  interesting. The brilliant step she took
• 19:33
  wasn't just asking for the data to be summarized.
• 19:36
  She then asked ChatGPT to build custom questions for her to ask
• 19:41
  the human admissions counselor during her upcoming phone calls.
• 19:44
  Oh wow, that is a profound shift in how to utilize the
• 19:48
  technology. She said that ChatGPT helped her
• 19:51
  come up with highly specific questions about her unique
• 19:54
  living and working constraints that she hadn't even thought to
• 19:56
  ask. It's surfaced her blind spots.
• 19:58
  Exactly. The AI essentially interviewed
• 20:01
  her life to prep her for the real interview.
• 20:04
  This is a perfect example of using AI not just as an advanced
• 20:07
  search engine, but as a genuine thought partner.
• 20:10
  When you're stressed and overwhelmed by a big life
• 20:13
  decision, your perspective naturally narrows.
• 20:16
  You simply don't know what you don't know, right?
• 20:19
  So by asking the AI to generate questions based on her specific
• 20:23
  constraints, she essentially used it to expand her own
• 20:26
  critical thinking. From Solving unsolvable 56 year
• 20:29
  old math problems to helping someone in Fresno figure out
• 20:33
  their college plan so they can advance their career, it's just
• 20:37
  incredible to see the spectrum of how this technology is
• 20:40
  touching absolutely everything. It really is.
• 20:43
  And if we pull all these threads together, it leaves us with a
• 20:45
  rather profound question to consider today.
• 20:47
  Let's hear it. If an AI system can formally
• 20:50
  verify abstract mathematical theorems for the cost of a
• 20:53
  household blender, and it can uncover critical zero day
• 20:56
  software vulnerabilities better than dedicated human security
• 20:59
  teams, how long until AI isn't just checking our existing work,
• 21:04
  but generating entirely new fields of science that humans
• 21:07
  don't even have the vocabulary to comprehend?
• 21:10
  Wow, that is a wild thought to leave off on.
• 21:13
  We are truly watching the definition of logic and
• 21:15
  discovery expand right in front of us.
• 21:18
  Well, that is all the time we have for this deep dive into the
• 21:20
  sources. Make sure you hit subscribe so
• 21:22
  you always stay updated on these rapidly moving developments.
• 21:25
  And if you enjoyed unpacking all of this with us today, we'd love
• 21:28
  it if you could read the show 5 stars.
• 21:30
  It really helps us out. We'll see you right back here
• 21:32
  for tomorrow's deep dive.