Latest / AI Fire Daily / 🎙️ EP 389: Karpathy’s 2-Hour YouTube Alternative & Doximity’s Lethal AI Dosing Warning!

Transcript

🎙️ EP 389: Karpathy’s 2-Hour YouTube Alternative & Doximity’s Lethal AI Dosing Warning! · AI Fire Daily

• 0:00
  You know, I was just thinking about this. Imagine
• 0:02
  a machine building a flawless video platform
• 0:05
  from scratch. Yeah, just entirely on its own.
• 0:07
  Right. And it does this easily on a quiet Saturday
• 0:10
  afternoon. But then that same machine constantly
• 0:14
  hallucinates a fatal medical dosage. Moving a
• 0:17
  decimal point on a prescription will actually
• 0:18
  kill you. It is an incredibly bizarre reality
• 0:21
  we are living in right now. It really is. Welcome
• 0:24
  to today's deep dive. We are exploring the shifting
• 0:27
  tectonic plates of modern AI. There is just so
• 0:30
  much happening in the space today. There is.
• 0:33
  First, we're dissecting Andrej Karpathy's fascinating
• 0:35
  weekend project, Hackertude. Then we are going
• 0:38
  to look at the hidden infrastructure war. This
• 0:41
  includes some massive updates from Anthropic
• 0:42
  and OpenAI. And then we have to confront that
• 0:45
  new Doximity study on medical AI. Exactly. It
• 0:48
  reveals exactly why you just can't trust AI with
• 0:52
  your health yet. So let's start with the sheer
• 0:54
  speed of digital creation. Yeah, the HackerTube
• 0:56
  story is wild. Karpathy got really frustrated
• 1:00
  with YouTube's central recommendation algorithm.
• 1:03
  The corporate algorithm always pushes, you know,
• 1:05
  mindless engagement over actual quality. So he
• 1:08
  just spent a single weekend building a functional
• 1:10
  alternative. Right. He built HackerTube in exactly
• 1:14
  two hours of actual work. Utilized Claude Code
• 1:17
  and the Opus 5 .5 model to do it. And his manual
• 1:21
  input was practically non -existent, right? Basically,
• 1:23
  yeah. He just grabbed a YouTube API key. He set
• 1:27
  up a server and bought a domain name. The AI
• 1:29
  models totally took over from there. He just
• 1:31
  wrote the complex code and structured the database
• 1:33
  autonomously. It's totally wild to see that level
• 1:36
  of execution. The scale of the data processing
• 1:39
  is what really gets me. It analyzed over 43 million
• 1:43
  different hacker news comments. Wow. Yeah. These
• 1:46
  archives span nearly 20 years of internet history.
• 1:49
  It identified about 39 ,000 comments recommending
• 1:52
  highly specific videos. And that curated over
• 1:55
  6 ,000 individual YouTube videos. Right. Across
• 1:58
  150 different technical and scientific topics.
• 2:01
  We're talking about like deep programming concepts
• 2:04
  and physics. Whoa. I mean, imagine crunching
• 2:08
  43 million comments to build a perfect video
• 2:11
  feed. It is genuinely hard to comprehend that
• 2:14
  sheer volume. It really is. It read two decades
• 2:17
  of human opinions. Almost instantly. Right. Because
• 2:20
  it isn't reading them like a human reads a book.
• 2:22
  It converts the text into mathematical vectors
• 2:25
  to find the semantic meaning. I think it consumed
• 2:28
  roughly 504 million input tokens in total. Tokens
• 2:32
  are just the basic building blocks of text AI
• 2:34
  reads. Exactly. It is how the machine breaks
• 2:37
  down human language. And it generated about 7
• 2:39
  .6 million output tokens, too, mostly using Anthropic's
• 2:43
  Sonnet model. It feels like a massive... paradigm
• 2:47
  shift is quietly happening here. Relying on AI
• 2:50
  completely changes the standard curation dynamic
• 2:52
  online. It absolutely does. It is like having
• 2:56
  a million internet sleuths handpicking your daily
• 2:58
  feed. You aren't relying on a corporate algorithm
• 3:01
  optimizing for ad revenue anymore. Plus, your
• 3:05
  user preferences are stored locally within your
• 3:07
  own browser. Which protects your digital privacy
• 3:09
  in a very structural way. You own the algorithm
• 3:12
  instead of the algorithm owning you. Does this
• 3:17
  mean the era of the centralized recommendation
• 3:19
  algorithm is ending? I strongly believe it points
• 3:21
  directly toward that future. Centralized algorithms
• 3:25
  optimize heavily for mass engagement. They just
• 3:27
  want to keep you endlessly scrolling. But personalized
• 3:30
  AI curation optimizes for your deep, specific
• 3:33
  interests, removing from passive consumption
• 3:36
  to active, highly personalized curation. So personalized
• 3:41
  AI curation is replacing mass market algorithms.
• 3:43
  100%. But... Owning your own algorithm comes
• 3:47
  with a massive hidden cost. The compute power.
• 3:50
  Exactly. Building tools like HackerTube requires
• 3:53
  absolutely incomprehensible computing power.
• 3:56
  You need massive server farms running constantly.
• 4:00
  Which brings us to the actual physical war happening
• 4:03
  right now. Right, the infrastructure updates.
• 4:06
  Anthropic just updated their entire model lineup.
• 4:09
  Yeah, they officially renamed Claude Max 20X
• 4:11
  to Claude Ultra. And they finally removed that
• 4:14
  five -hour usage limit for users. That is a huge
• 4:18
  relief. And Max 5X is now simply Claude Max.
• 4:21
  Right. And internet sleuths found something else
• 4:23
  hiding on their site, too. They spotted Claude
• 4:26
  Mythos 5 .5 lurking in the background. Geared
• 4:28
  specifically toward cyber defense, right? Yeah,
• 4:31
  heavy security tasks. And we might even see Fable
• 4:34
  5 .5 tag along shortly. People expect a public
• 4:37
  launch as early as next week. The pace of these
• 4:40
  releases is accelerating so rapidly. I still
• 4:43
  wrestle with prompt drift myself on my personal
• 4:46
  projects. It is definitely hard to keep up. It
• 4:48
  really is. So the idea of these massive compute
• 4:52
  clusters feels incredibly daunting. OpenAI is
• 4:55
  completely rethinking their entire physical footprint
• 4:57
  right now. Yeah, they want to rebrand data centers
• 5:00
  as AI compute campuses. Which... clearly aims
• 5:04
  to win over the skeptical American public. They
• 5:07
  are preparing for a massive fight over local
• 5:10
  power grids. Yes, and their financial reality
• 5:13
  is shifting quite dramatically, too. Their mega
• 5:16
  -funding round target just dropped from $30 billion...
• 5:19
  Down to $20 billion, right? Yeah, and they've
• 5:21
  only managed to secure roughly $11 billion so
• 5:23
  far. Big investors are definitely cooling on
• 5:26
  this massive unchecked AI spending. The economic
• 5:29
  reality is finally catching up to the initial
• 5:31
  hype. Exactly. Massive returns aren't materializing
• 5:35
  instantly. Meanwhile, Apple is taking a totally
• 5:39
  different philosophical approach. John Ternes
• 5:41
  made their firm design stance clear. They absolutely
• 5:44
  won't ship AI for AI's sake in their products.
• 5:46
  The technology must make the physical product
• 5:48
  genuinely much better, even if users barely notice
• 5:52
  the AI working in the background. They want seamless
• 5:54
  integration, not flashy gimmicks. Speaking of
• 5:57
  the industry, though, there is some incredibly
• 5:59
  petty drama unfolding. Oh, yeah. Met as Alexander
• 6:02
  Wang, but a very specific internet domain name.
• 6:06
  It previously belonged to an OpenAI employee
• 6:08
  named Thibaut Saru. And Wang set up a troll redirect
• 6:12
  just to mock his rival. It is petty tech industry
• 6:16
  drama at its absolute finest. It just shows the
• 6:19
  intense rivalry happening behind the scenes.
• 6:22
  It really does. But on a much more serious hardware
• 6:26
  note, Sebastian Thrun recently made a move. Yeah,
• 6:29
  the founder behind Udacity and Waymo's self -driving
• 6:31
  program. He just raised $130 million for Dulo.
• 6:35
  That's a massive investment for a new hardware
• 6:37
  startup. What exactly is Dulo trying to accomplish
• 6:40
  with that money? They build highly specialized
• 6:42
  AI models that design complex physical hardware.
• 6:46
  They use generative design to map out entirely
• 6:48
  new physical components. So the AI can test thousands
• 6:52
  of structural variations in a simulation. Right.
• 6:54
  It aims to speed up global manufacturing quite
• 6:57
  significantly. It's fascinating how different
• 6:59
  companies frame their physical infrastructure.
• 7:01
  Wait, why is OpenAI trying to rebrand data centers
• 7:06
  as campuses? It's a very deliberate, calculated
• 7:09
  public relations strategy. Data centers sound
• 7:11
  industrial, resource -heavy, and corporate to
• 7:14
  locals. Right, like massive cooling towers draining
• 7:17
  municipal water. Exactly. But campuses sound
• 7:21
  academic, innovative, and beneficial to society.
• 7:23
  They desperately need public support to tap into
• 7:26
  massive public power grids. They want facilities
• 7:29
  to sound like universities, not industrial power
• 7:31
  drains. That is exactly the playbook. We are
• 7:34
  going to take a very brief pause here for our
• 7:36
  sponsor. We will be right back with the Doximity
• 7:39
  study. And we're back. So Apple's philosophy
• 7:43
  of avoiding AI for AI's sake is critical right
• 7:46
  now. It really is, especially when you look at
• 7:48
  high -stakes fields. Right, because when we blindly
• 7:51
  trust this massive compute in the real world,
• 7:54
  the results can be terrifying. Let's pivot to
• 7:59
  the highly regulated medical field. This is where
• 8:01
  the reality check hits hard. Doximity recently
• 8:05
  tested nine leading AI models, quite extensively.
• 8:08
  They asked over 8 ,000 specific medication questions.
• 8:11
  And those covered over 2 ,204 unique prescription
• 8:16
  drugs, right? And the empirical results they
• 8:19
  found were incredibly concerning. Over 2 .3 %
• 8:23
  of AI generated doses were factually incorrect.
• 8:26
  That was based on roughly 261 ,000 evaluations.
• 8:31
  That might sound like a small percentage at first
• 8:33
  glance. But at a global medical scale, it is
• 8:36
  completely unacceptable. Almost 43 .8 % of those
• 8:40
  specific errors were considered serious. Even
• 8:44
  worse, 3 .4 % were classified as potentially
• 8:46
  lethal. One in five severe errors involved simple
• 8:51
  basic unit mix -ups. Yeah, the models confused
• 8:54
  vital decimal point placements. They confused
• 8:58
  milligrams with micrograms, which is an instantly
• 9:01
  deadly mistake. Relying on AI for medication
• 9:04
  is like stacking Lego blocks of data, but you
• 9:08
  are completely missing the single structural
• 9:10
  piece that prevents a collapse. That is the perfect
• 9:13
  way to visualize the core underlying problem.
• 9:16
  An LLM doesn't actually understand the biological
• 9:20
  reality of human kidneys. It just looks at the
• 9:23
  statistical probability of the next word. Right.
• 9:26
  The AI frequently failed to adjust doses for
• 9:29
  patient age. It completely ignored body weight,
• 9:32
  kidney function, and critical drug interactions.
• 9:35
  It hallucinates a dosage without understanding
• 9:38
  frail human biology. Exactly. It confidently
• 9:42
  invents a highly plausible but totally false
• 9:45
  clinical fact. And it missed critical context,
• 9:48
  even when patient information was explicitly
• 9:50
  provided. The text was right there, but the model
• 9:53
  couldn't synthesize it. It just doesn't reason
• 9:56
  like a doctor. What's even more concerning is
• 9:59
  how the models handle newer breakthroughs. Yeah,
• 10:02
  newer medications actually had a 5 .1 times higher
• 10:06
  error rate. These are complex new drugs officially
• 10:09
  approved since the year 2024. AI models frequently
• 10:12
  overlook these critical new patient dosing details.
• 10:15
  And the confident tone of the AI completely masks
• 10:19
  its clinical incompetence. It sounds incredibly
• 10:22
  authoritative while giving you a lethal prescription
• 10:24
  dosage. It is terrifying. So Doximity launched
• 10:28
  a safety feature called Dose Check very recently.
• 10:31
  To automatically check AI -generated doses against
• 10:34
  verified medical databases, right? Yes, it acts
• 10:37
  as a necessary deterministic guardrail for the
• 10:40
  generative model. It compares the output to official
• 10:43
  FDA labeling and clinical guidelines. They claim
• 10:46
  it achieves over 99 % dosing accuracy now, but
• 10:49
  it requires that secondary non -AI layer to be
• 10:52
  safe. Why exactly do newer medications have a
• 10:55
  five times higher error rate? It comes down entirely
• 10:58
  to the model's training data cutoffs. These AI
• 11:01
  models are essentially frozen in time after their
• 11:04
  training finishes. So they don't continuously
• 11:06
  learn new facts from the live internet automatically.
• 11:09
  Right. If a life -saving drug was invented later,
• 11:12
  the model just guesses. It tries to extrapolate
• 11:15
  dosages from older, somewhat similar sounding
• 11:18
  drugs. The AI simply hasn't read enough about
• 11:20
  drugs invented after its training ended. And
• 11:23
  hallucinating a dosage for a brand new drug is
• 11:26
  medically catastrophic. We clearly can't trust
• 11:30
  AI with lethal medical prescriptions yet. The
• 11:33
  fundamental grounding in physical reality simply
• 11:35
  isn't there today. No, not at all. But we can
• 11:38
  trust it to automate daily software workflows
• 11:40
  safely. Right. The stakes are drastically lower
• 11:43
  in basic digital workflow management. Several
• 11:46
  new empowered digital tools are making massive
• 11:48
  waves today. Yeah, the digital workflow space
• 11:51
  is where AI truly shines right now. Let's look
• 11:54
  at what actually works reliably for us. Screen
• 11:57
  Studio makes your basic screen recordings look
• 11:59
  incredibly professional. And it achieves this
• 12:02
  without requiring a dedicated, expensive video
• 12:06
  editor. Screen Studio features automatic zooms
• 12:09
  and transcript -based video editing capabilities.
• 12:12
  It tracks your mouse movements completely seamlessly
• 12:14
  across the screen. It knows exactly where the
• 12:17
  viewer's attention should naturally focus. It
• 12:19
  even handles dynamic background removal for your
• 12:22
  screen recordings automatically. You literally
• 12:24
  don't need a green screen to look like a professional.
• 12:26
  Then you have Klaus Go Science for the dedicated
• 12:29
  academic researchers. That one is amazing. It
• 12:31
  literally turns simple research questions into
• 12:34
  actual cited scientific discoveries. It automates
• 12:38
  over 500 complex lab tools completely seamlessly.
• 12:41
  And it processes over 90 massive data sets entirely
• 12:45
  automatically for researchers. It accelerates
• 12:47
  the boring parts of tedious scientific discovery
• 12:51
  immensely. It removes the mechanical friction
• 12:53
  from the core scientific process. It lets the
• 12:56
  human scientist focus entirely on the actual
• 12:59
  hypothesis. Then there's Riverside Agents, which
• 13:02
  is a digital team for video. It handles recording
• 13:05
  prep, complex video editing, and viral social
• 13:08
  clips. It acts exactly like a digital producer
• 13:11
  living right in your browser. Yeah, it even manages
• 13:14
  your captions and automatic video scheduling
• 13:15
  across platforms. Then there's a powerful tool
• 13:19
  called Embage, available today. It lets you build
• 13:22
  AI chat and voice agents very easily. These digital
• 13:25
  agents can automate workflows and rapidly qualify
• 13:28
  your leads. They connect hundreds of different
• 13:30
  enterprise software tools together seamlessly.
• 13:33
  They basically act as the connective tissue between
• 13:36
  your isolated applications. Of all these workflow
• 13:40
  tools, which actually saves the most human time?
• 13:43
  Honestly, video editing is notoriously slow and
• 13:46
  highly tedious manual work. You spend hours constantly
• 13:50
  scrubbing through timelines for tiny audio mistakes.
• 13:53
  So the agents that handle that save the most
• 13:55
  time? Exactly. Agents that cut clips and handle
• 13:58
  zoom tracking save countless hours. You eliminate
• 14:01
  the raw mechanical clicking from the creative
• 14:03
  process completely. So anything eliminating tedious
• 14:06
  video editing yields the highest immediate return.
• 14:09
  By far. It frees the human creator completely
• 14:12
  to focus on the narrative. It's fascinating to
• 14:17
  step back and look at this bigger picture today.
• 14:19
  We are looking at a technology with an incredibly
• 14:22
  profound duality. It really is a double -edged
• 14:26
  sword right now. It possesses blistering speed
• 14:29
  and vast pattern recognition capabilities today.
• 14:32
  It curates 20 years of hacker news data in two
• 14:35
  hours. It understands what programmers want to
• 14:38
  watch perfectly well online. Right. And it builds
• 14:41
  a flawless video platform almost instantly. The
• 14:44
  raw generative capability in the digital realm
• 14:47
  is completely staggering. Yet it entirely lacks
• 14:50
  the basic common sense required for daily survival.
• 14:53
  Exactly. It forgets to check a crucial decimal
• 14:56
  point on a prescription. It will confidently
• 14:59
  give you a lethal medical dosage without hesitation.
• 15:02
  It simply doesn't understand the physical world
• 15:04
  at all yet. It perfectly simulates human taste
• 15:07
  in our digital media landscape, but it hallucinates
• 15:11
  deadly math in our physical hospitals and pharmacies.
• 15:15
  It's a scary thought. Are we rushing to use it
• 15:18
  for the wrong real -world tasks? Maybe AI should
• 15:21
  remain our... creative digital muse for now.
• 15:24
  I think that's exactly where it belongs today.
• 15:26
  It can brainstorm, curate feeds, and inspire
• 15:29
  our daily creative work. And traditional deterministic
• 15:33
  software should remain our highly reliable calculator.
• 15:37
  Keep that profound duality in mind this coming
• 15:39
  week. Always verify the output, folks. Definitely.
• 15:43
  Double -check your AI's work, no matter how convincing
• 15:45
  it sounds. Thank you for joining this deep dive
• 15:48
  with us today. Keep exploring and always keep
• 15:51
  questioning the final automated outputs.