Latest / Today’s AI News / AI Outperforms Doctors in the ER, Anthropic Left Off Pentagon's List Again, AI Bans Spread

Transcript

AI Outperforms Doctors in the ER, Anthropic Left Off Pentagon's List Again, AI Bans Spread · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X Roductions.
• 0:03
  Yeah, it is a it's great to be back.
• 0:06
  Imagine a brilliant medical resident who never sleeps, has
• 0:10
  read every medical journal in existence, and is subtly beating
• 0:14
  seasoned attending doctors at their own game.
• 0:17
  That's exactly what an older AI model just pulled off in a busy
• 0:20
  emergency room, and it's our lead story today.
• 0:23
  It really is a watershed moment, not just for medical technology,
• 0:27
  but for how we actually understand expertise itself.
• 0:30
  Right. And that is the exact mission
• 0:32
  for today's Deep Dive. We are exploring how AI is
• 0:35
  fundamentally redefining authority across the board.
• 0:38
  We're looking at everything from life or death decisions in the
• 0:41
  ER to the highest level of the Pentagon's classified networks,
• 0:44
  all the way down to how you, the listener, can take back
• 0:46
  authority at your local grocery store.
• 0:48
  It's a massive landscape to cover, and you know, the
• 0:50
  connective tissue through all of these stories is trust.
• 0:53
  Exactly. We are watching a real time
• 0:55
  shift in where society places its trust, moving away from
• 0:58
  traditional human institutions and well, toward artificial
• 1:02
  intelligence. OK, let's unpack this lead
• 1:04
  story, because mechanics of it are wild.
• 1:06
  A new Harvard study was just published in the journal
• 1:08
  Science. They took Open AI's 01 preview
• 1:12
  model. Right.
• 1:13
  Now, for context, that was released back in 2024, so we
• 1:16
  weren't even talking about the bleeding edge of current compute
• 1:19
  here. They tested this older model
• 1:22
  across 76 real historical emergency room cases and we.
• 1:26
  Have to look at how they leveled the playing field because they
• 1:29
  actually handicapped the AI in a very specific way.
• 1:32
  The model didn't have access to X-rays, MRI scans, or like any
• 1:36
  visual data at all. It was fed purely on raw
• 1:39
  electronic health record text. Just the written triage notes
• 1:42
  and vital signs. Right.
• 1:43
  And the numbers they got back are staggering.
• 1:46
  At the initial ER triage stage, which is arguably when the room
• 1:49
  is the most chaotic and time sensitive, the AI diagnosed the
• 1:52
  patients correctly 67.1% of the time.
• 1:56
  Yeah, and they compared that against two human attending
• 1:58
  physicians working the same cases.
• 2:00
  The human scored 55.3% and 50.0. Percent.
• 2:04
  That's a huge gap. It is.
• 2:06
  And just to completely eliminate any bias, the researchers had
• 2:10
  two blind physician reviewers score the results without
• 2:14
  knowing if the diagnosis came from a human or a machine.
• 2:17
  And they? Couldn't tell the difference
• 2:18
  right? They literally could not
• 2:19
  distinguish the AIS work from the human doctors.
• 2:22
  But here is the craziest part. In a standout case, the AI
• 2:27
  flagged a rare flesh eating infection in a transplant
• 2:30
  patient 12 to 24 hours before the treating doctor caught it.
• 2:34
  What's fascinating here isn't just the overall score, but the
• 2:37
  broader implications of it all. It's not just about beating
• 2:39
  humans. It's the fact that a model that
• 2:41
  is already, you know, generations behind current tech
• 2:45
  is doing this. Like 1,000,000 use AI for health
• 2:49
  advice already, but this proves its utility as a formal tool for
• 2:53
  the doctors themselves. See.
• 2:54
  OK, that is where I had to play devil's advocate for a second.
• 2:56
  Go for it. An emergency room is pure
• 2:59
  sensory overload. It's the smell of the room, the
• 3:02
  sound, the specific look of exhaustion in a patient's eyes.
• 3:06
  Doctors rely so heavily on that human gut instinct when someone
• 3:10
  walks through the door. How does an AI do this with just
• 3:14
  raw text without physically seeing the patient or relying on
• 3:18
  a doctor's gut instinct? Well, that sterile nature is
• 3:21
  exactly the feature, not the bug.
• 3:23
  I mean, the human gut instinct is incredibly valuable, yes, but
• 3:26
  it's also highly vulnerable to cognitive bias and physical
• 3:30
  fatigue. OK, Fairpoint.
• 3:32
  Think about the mechanics of human attention.
• 3:34
  Human doctors often use a mental spotlight.
• 3:36
  They see a prominent symptom like chest pain and shine all
• 3:40
  their diagnostic focus on the most statistically probable
• 3:43
  cause. And they miss the other stuff.
• 3:44
  Precisely. The spotlight leaves the rest of
• 3:47
  the patient's history in the shadows.
• 3:48
  We call that anchoring bias. So they get tunnel vision on the
• 3:51
  obvious problem. Exactly, especially a doctor at
• 3:55
  hour 11 of a 12 hour shift. The AI on the other hand uses a
• 3:59
  floodlight processing just the raw text.
• 4:02
  It doesn't get tired. It takes all 5000 data points in
• 4:07
  that patient's record and illuminates them simultaneously.
• 4:10
  So it's essentially acting as this exhaustive, tireless safety
• 4:13
  net for the raw data itself. Yeah, exactly.
• 4:17
  And if a handicap 2024 model can beat an attending ER physician
• 4:21
  by over 10 percentage points in triage accuracy, just imagine
• 4:25
  what a current frontier model is going to do for patient
• 4:28
  outcomes. It's an incredible shift in
• 4:29
  authority. And you know, if we follow that
• 4:31
  through line of trust, it perfectly sets up the
• 4:34
  geopolitical landscape because if we are starting to trust AI
• 4:37
  with our physical health in the ER, the next logical step is
• 4:41
  trusting it with our national security.
• 4:42
  Oh, absolutely. The stakes escalate immediately
• 4:45
  from individual survival to National Defense.
• 4:47
  Here's where it gets really interesting.
• 4:49
  The Pentagon just made a definitive move in that
• 4:52
  direction. They added eight AI companies to
• 4:54
  their highly classified networks to make the US military an AI
• 4:58
  first fighting force. Which is a huge shift.
• 5:01
  It's massive. The list includes SpaceX, Open
• 5:04
  AI, Google, NVIDIA, Microsoft, AWS, Oracle, and a smaller
• 5:09
  player called Reflection. But the real story is who isn't
• 5:13
  on the list? Entropic was explicitly
• 5:16
  excluded. They were blacklisted.
• 5:18
  Which creates A fascinating paradox, because, according to
• 5:21
  reporting from the Washington Post, the new contracts for
• 5:24
  those eight approved companies contain the exact same
• 5:27
  autonomous weapons and surveillance limits that
• 5:29
  supposedly caused anthropics blacklisting in the first place.
• 5:32
  Right. The rules didn't change, just
• 5:33
  the guest list. And it gets a little
• 5:35
  contradictory. DOGCTO, Emil Michael cited
• 5:38
  ancropics supply chain risk, basically a fear that foreign
• 5:41
  adversaries might have compromised their models.
• 5:43
  Yeah, but in the exact same breath, he called Anthropics new
• 5:47
  unreleased de Mythos model model a separate national security
• 5:50
  moment. And compounding that, the White
• 5:52
  House recently advised against a broader roll out of that Mythos
• 5:54
  model over severe compute concerns, yet the government
• 5:57
  seemingly still wants priority access to it.
• 6:00
  Right. So the So what analysis here is
• 6:02
  that the government is trying to balance intense national
• 6:05
  security risks with the absolute terror of falling behind on
• 6:09
  frontier AI compute power? Yeah, it's a tightrope.
• 6:12
  It really is. They're basically looking at a
• 6:14
  frontier model like Mythos and saying, well, we don't fully
• 6:17
  trust who built this engine, but if we don't put it in our jet,
• 6:21
  we lose the war. It's incredible and this rush to
• 6:25
  secure the best engines is reshaping where the money comes
• 6:28
  from. For example, Reflection, that
• 6:30
  smaller company that did make the cut recently raised $2
• 6:34
  billion from 1789 capital, which is a Donald Trump junior backed
• 6:38
  fund. And I want to be super clear
• 6:40
  here, we are maintaining strict political impartiality.
• 6:43
  We aren't taking sides on White House policies or specific
• 6:46
  investment funds. We are just impartially
• 6:48
  reporting the facts in the source material about how these
• 6:50
  military AI contracts are playing out.
• 6:53
  Absolutely, and the fact is the barrier to entry for building
• 6:56
  these models is so astronomically high that
• 6:59
  companies have to source capital from deeply interconnected,
• 7:03
  politically adjacent funds. Everyone realizes that military
• 7:07
  AI contracts are the defining gold rush of this decade for
• 7:10
  sure. But you know, while the
• 7:12
  government figures out national security, ordinary people are
• 7:15
  wielding AI to take back control of their daily lives.
• 7:19
  Let's shift the focus to everyday AI mastery.
• 7:22
  Yeah, I love these community examples.
• 7:24
  This represents a huge shift from AI as just a corporate or
• 7:28
  military tool to like a personal advocate.
• 7:30
  Totally. Yeah, our first round table use
• 7:32
  case is from a user named Jennifer.
• 7:35
  Her daughter was diagnosed with a rare autoimmune disease, so
• 7:38
  she used Gemini and Chachi PT to research it.
• 7:40
  That's. Amazing, yeah.
• 7:41
  She sort of dense medical literature, explored global
• 7:44
  treatment options, and actually successfully identified leading
• 7:47
  US specialists for her daughter's exact condition.
• 7:49
  Whether you're decoding a rare disease or you know anything
• 7:52
  else complex, AI is primary superpower right now is
• 7:55
  demystifying jargon for the consumer.
• 7:57
  That is so true, and it applies to completely mundane things
• 8:01
  too. Our second use case is from
• 8:03
  Shovem. He uses ChatGPT as a label
• 8:06
  reading filter for packaged snacks.
• 8:08
  Oh, like at the grocery store? Yes, he uploads photos of the
• 8:11
  ingredient list to decode industrial jargon.
• 8:14
  It flags INS numbers, stabilizers, emulsifiers and
• 8:18
  hidden sugars or palm oils. It basically bypasses those
• 8:21
  misleading sugar free claims. That is so.
• 8:23
  Smart. It's like having a bilingual
• 8:25
  nutritionist in your pocket. It forces transparency.
• 8:28
  Corporations have relied on that asymmetry for decades, and the
• 8:32
  AI just removes the friction of doing that research.
• 8:34
  Exactly. And it's not just about
• 8:36
  consuming better information, it's about building things too.
• 8:40
  Anthropic just dropped a new guide for Claude Design, which
• 8:43
  is their tool for building high converting landing pages.
• 8:45
  Right for people with 0 coding experience.
• 8:47
  Exactly. The process is wild.
• 8:50
  You go to Claude dot A slash design, select wireframe and
• 8:53
  describe your audience and product.
• 8:55
  But then you upload screenshots of a niche page you like, Andy a
• 8:58
  high transaction page like Amazon or eBay.
• 9:00
  You literally just give it pictures.
• 9:02
  Just flat screenshots and you ask for four variations.
• 9:05
  You can refine it via comments and the ultimate pro tip is to
• 9:08
  use the Handoff to Claude code web feature.
• 9:11
  It autonomously build and deploys the final site for you.
• 9:14
  That completely collapses the traditional production pipeline.
• 9:17
  What used to take a copywriter, a designer and a developer is
• 9:22
  now just one person directing an AI.
• 9:24
  Yeah, it's democratizing production.
• 9:26
  And because this democratization is happening so incredibly fast,
• 9:30
  you know the rules, tools and legal guardrails are dropping
• 9:33
  daily. Let's wrap but fire through the
• 9:35
  rest of the landscape and some quick hits.
• 9:37
  Let's do it. First up on the tool side,
• 9:39
  Custom Voices just launched. It allows incredibly accurate
• 9:42
  voice cloning via short clips, specifically for Grok apps.
• 9:46
  Nice. Then open AI shipped Codex Pets.
• 9:49
  These are animated desktop companions that let you track
• 9:52
  your AI agents progress without having to switch apps.
• 9:55
  Which is actually great for keeping you in a flow state,
• 9:57
  right? 11 Music also launched a
• 9:59
  platform for AI song generation that actually includes creator
• 10:02
  payouts, and overseas, Xiaomi released a powerful open source
• 10:06
  model called MIMO V 2.5 Pro. Which is a big deal because it
• 10:10
  proves you don't always need massive centralized servers to
• 10:13
  run top tier AI. Exactly Over on the industry and
• 10:16
  legal side, Open AICEO Sam Altman announced Open Claw users
• 10:20
  can now use ChatGPT subscriptions inside the tool,
• 10:24
  which counters Anthropics previous restrictions.
• 10:26
  It's a total battle for ecosystem dominant.
• 10:29
  Oh, totally. Now this next one is a massive
• 10:31
  win for consumer protection. Maryland just signed EU s s
• 10:35
  first ban on AI driven grocery pricing.
• 10:38
  Wow. Yeah, they're finding stores up
• 10:40
  to $25,000 for using personalized data to market
• 10:43
  prices. Is this the start of a broader
• 10:46
  legislative trend? I really think it is dynamic.
• 10:48
  Pricing is insidious, and lawmakers are finally catching.
• 10:51
  On Thank Goodness also sag after I secured new AI guardrails in a
• 10:55
  four year studio deal and a Chinese court just ruled that
• 10:59
  replacing a worker with AI does not legally justify firing them.
• 11:02
  They actually ordered wrongful termination damages.
• 11:04
  That is a staggering legal precedent.
• 11:07
  It forces companies to legally consider the human cost of
• 11:10
  automation. Seriously.
• 11:12
  OK, before we wrap up, let's talk about community AI
• 11:15
  workflows. Let's look at how a listener is
• 11:17
  putting this toward today. Always the best part.
• 11:19
  This one is from reader Ollie T out in Finland.
• 11:22
  He's a new real estate investor, and he was struggling with a
• 11:25
  lack of historical real estate data and market transparency in
• 11:28
  his local county. So what did he do?
• 11:30
  He used Gemini Pro to create a concept tool that analyzes
• 11:33
  market info, renovation history, and future needs.
• 11:37
  It even generates investment calculations for him.
• 11:40
  That's. Brilliant, but the best part is
• 11:42
  how he built it. He refined the tool using
• 11:46
  multiple free AIS like Claude and Codex by doing just one
• 11:49
  prompt per day. Oh, I love that.
• 11:51
  If we connect this to the bigger picture, that one prompt per day
• 11:54
  method is a brilliant, sustainable way to build complex
• 11:57
  workflows, right? Yeah, without feeling
• 11:59
  overwhelmed. You don't have to be a prompt
• 12:01
  engineer in a weekend, you just need daily curiosity.
• 12:04
  Consistency beats intensity. Well, we have covered an immense
• 12:08
  amount of ground today. We really have, and it leaves us
• 12:11
  with a final lingering question. If a 2024 AI model can already
• 12:16
  out diagnose seasoned ER doctors, and international
• 12:20
  courts are now ruling that AI replacements don't legally
• 12:23
  justify firing humans, how will society redefine the value of
• 12:27
  human expertise in the next five years?
• 12:29
  That is such a heavy thought to leave on, but it is so
• 12:32
  necessary. Well, thank you for diving deep
• 12:34
  with us today. Make sure you hit that subscribe
• 12:36
  button to stay updated on all these shifts.
• 12:38
  We'd love it if you can rate the show 5 stars and we'll see you
• 12:40
  back for tomorrow's episode.