Latest / Today’s AI News / Pope Leo XIV Weighs In on AI, Meta’s Guardrails Removed in Minutes, Uber Says AI Is Hard to Justify

Transcript

Pope Leo XIV Weighs In on AI, Meta’s Guardrails Removed in Minutes, Uber Says AI Is Hard to Justify · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:02
  Productions. Glad to be here as always so.
• 0:05
  Imagine handing the keys of a supersonic jet to a teenager.
• 0:10
  They're going incredibly fast, but nobody is quite sure if they
• 0:13
  know where the brakes are or who even owes the sky.
• 0:17
  Today we're looking at who really controls AI trajectory,
• 0:20
  from the Vatican weighing in on morality to hackers removing the
• 0:25
  safety brakes in minutes. Yeah, it is an absolute tug of
• 0:29
  war right now, you know, between these massive institutions
• 0:32
  trying to build regulatory ceilings and the open source
• 0:36
  community basically just pulling the floor out from under them.
• 0:38
  Exactly, and we've got a really fascinating stack of sources for
• 0:42
  you in this deep dive. Our mission today is to figure
• 0:45
  out how the world is actually wrestling with a eyes like
• 0:48
  rapidly accelerating power. Right, trying to balance that
• 0:50
  undeniable value against the desperate scramble for control.
• 0:53
  Yeah, exactly. So let's start with a voice you
• 0:56
  might not immediately expect to hear in a, you know, Silicon
• 0:59
  Valley hardware discussion. Pope Leo the 14th.
• 1:01
  Right. The Vatican has officially
• 1:03
  entered the chat. I mean literally.
• 1:05
  The Pope just published a massive 42,000 word encyclical.
• 1:10
  It's called Magnifica Humanitis. Which translates to magnificent
• 1:14
  humanity, right? And he sent this out to the
• 1:16
  Catholic Church's 1.4 billion members worldwide.
• 1:19
  And, you know, the historical context he uses there is
• 1:22
  incredibly deliberate, like the Pope draws this direct, explicit
• 1:26
  line between today's AI revolution and the Industrial
• 1:29
  Revolution. Oh, interesting, like the same
• 1:31
  scale of impact? Exactly.
• 1:32
  He is signaling that the church really intends to be a central
• 1:36
  moral voice and how this technology develops, just like
• 1:38
  they had to respond to the the sweeping societal changes of
• 1:42
  mass industrialization back in the day.
• 1:44
  OK, let's unpack this a bit though, because a 42,000 word
• 1:47
  spiritual document is, well, it's a massive undertaking.
• 1:51
  And the perp is basically treating big tech like rogue
• 1:54
  nation states, right? Yeah, pretty bad.
• 1:56
  Like he's warning that these private transnational monopolies
• 1:59
  are actually surpassing the capacity of localized
• 2:01
  governments to regulate them. But I mean, can a 42,000 word
• 2:05
  spiritual document actually forces Silicon Valley CEO to
• 2:08
  change their code? Well, his core argument is that
• 2:11
  AI is never neutral. You know every model carries the
• 2:15
  biases and the incentives of its creators, right?
• 2:18
  So he's calling for what he terms human friendly AI.
• 2:22
  He's warning that if we just leave development development
• 2:25
  solely to market forces, we really risk reducing people to
• 2:28
  mere cogs in this massive efficiency machine.
• 2:31
  He specifically demands independent oversight, robust
• 2:35
  legal frameworks and informed users, and he actually drew a
• 2:40
  very hard line regarding the military.
• 2:43
  Yeah, I saw that he explicitly stated that lethal decisions in
• 2:46
  warfare must never be delegated to AI, writing that no algorithm
• 2:51
  can ever make war morally acceptable.
• 2:53
  Which, you know, is a powerful philosophical stance.
• 2:56
  But again, moral frameworks are great, but the developers are
• 2:59
  the ones typing the keystrokes and like answering to
• 3:02
  shareholders. True, but the industry response
• 3:04
  actually suggests the message is resonating inside those very
• 3:07
  labs. Wait, really?
• 3:08
  Yeah, So Christopher Ola, he's a prominent researcher at
• 3:11
  Anthropic, which is, you know, one of the leading frontier AI
• 3:13
  labs. Oh, for sure.
• 3:14
  He publicly agreed with the Pope's assessment.
• 3:17
  No way. Yeah, Ola actually admitted that
• 3:20
  every Frontier AI lab operates inside a set of structural
• 3:24
  incentives that can, and often do, conflict with doing the
• 3:28
  right thing. Wow.
• 3:30
  I mean, that is a stark admission from a lead engineer.
• 3:32
  Right, admitting that the financial oppressions of the
• 3:35
  tech industry are frequently just at odds with safety and
• 3:38
  morality. So you have the Vatican trying
• 3:41
  to control AI's morality from the top down, establishing these
• 3:45
  grand ethical frameworks. But the reality is.
• 3:48
  Anyone with a laptop is rewriting those rules from the
• 3:50
  bottom up? Exactly, If the massive tech
• 3:52
  monopolies can't even govern themselves, what happens when
• 3:55
  the technology is just given away for free?
• 3:57
  Well, that brings us to this terrifying investigation from
• 4:00
  the Financial Times. It completely flips the
• 4:02
  conversation from theoretical morality to immediate structural
• 4:06
  vulnerability. Yeah, this was wild.
• 4:08
  The FT revealed how easily the safety guardrails can be
• 4:11
  stripped away from open source AI models using a tool called
• 4:14
  Heretic, and this thing is just freely available on GitHub.
• 4:17
  Yep, just sitting right there. They used Heretic on Metas LAMA
• 4:20
  3.3 model and get this it only took them 10 minutes and exactly
• 4:25
  4 lines of code just. 4 lines. No specialist hardware required.
• 4:29
  Yeah, Here's where it gets really interesting, though.
• 4:32
  It's like putting child locks on a cabinet but leaving the
• 4:35
  screwdriver sitting right on the counter.
• 4:37
  That is the perfect analogy. Right.
• 4:40
  If open source is this vulnerable to a four line hack,
• 4:44
  isn't the safety argument for open models completely falling
• 4:47
  apart? Like how does that actually work
• 4:50
  mechanically? Bypassing billions of dollars of
• 4:53
  corporate safety research with four lines of code sounds
• 4:55
  impossible. Well, to understand the
• 4:57
  mechanism you really have to look at how safety is built in
• 5:00
  the 1st place. OK, so when companies like Meta
• 5:02
  or Google release an open source model, they include a safety
• 5:05
  wrapper. It's done through a process
• 5:07
  called Reinforcement Learning from Human Feedback, or RLHF.
• 5:11
  Right, the RLHF. Yeah, it basically acts like a
• 5:13
  software governor on a high performance engine.
• 5:15
  It actively forces the model to refuse dangerous prompts.
• 5:19
  Makes sense? But because it is open source,
• 5:21
  you have the underlying weights the actual engine.
• 5:24
  So Heretic essentially runs a script that locates those
• 5:27
  specific safety weights and just mathematically neutralizes them.
• 5:31
  Wow. Yeah, it just snips the wire
• 5:33
  connecting the governor to the engine.
• 5:35
  It's like removing the steering wheel limiter on a golf cart so
• 5:38
  it can suddenly do 90 miles an hour.
• 5:40
  Exactly, and the results are terrifying.
• 5:43
  Once that governor was snipped, the FT reported that the
• 5:46
  modified de censored model happily provided exact dosages
• 5:50
  for ricin. The deadly bio weapon.
• 5:52
  Yeah, and Google's Jemma 3 and 4 models were also quickly
• 5:56
  stripped using the exact same methods.
• 5:58
  So if open source is this vulnerable, the entire tech
• 6:01
  industry must be freaking out, right?
• 6:03
  Now, no. They are wrestling with that
• 6:04
  exact vulnerability. When the FT reached out for
• 6:07
  comment, Google essentially called it a quote.
• 6:09
  Known technical challenge facing all open models.
• 6:12
  And meta. Meta declined to comment
• 6:14
  entirely. Well, of course.
• 6:16
  Because you know you are giving away the underlying code, so
• 6:19
  patching this is fundamentally difficult and the creator of
• 6:23
  Erotic claims the tool has already produced over 3500 de
• 6:27
  censored models. Which means.
• 6:29
  Which means it's resulted in 13 million downloads. 13 million
• 6:34
  downloads of AI models with no safety brakes.
• 6:37
  That is just staggering. And you know, this Wild West of
• 6:41
  open source pushes companies toward using expensive closed AI
• 6:45
  systems just to stay safe. Right, API contracts.
• 6:48
  Yeah, which raises a critical question.
• 6:50
  Are those massive corporate investments actually paying off?
• 6:54
  Because this is causing a massive financial hangover in
• 6:56
  corporate boardrooms right now. Absolutely.
• 6:58
  We are seeing a severe financial reality check and it's most
• 7:01
  notable over at Uber. Yeah, the Uber CEO.
• 7:03
  Right, Andrew McDonald in an interview on the Rapid Response
• 7:05
  Show, he stated point blank that it is getting harder to justify
• 7:09
  the company's massive AI compute spending.
• 7:11
  So to understand why the bill is so high, we have to look at this
• 7:14
  trend called token maxing. Yeah.
• 7:16
  Yeah, token maxing. For the past year, big tech
• 7:19
  companies basically assumed that more AI usage automatically
• 7:22
  equaled more productivity. Which I mean, if you look at
• 7:25
  traditional software economics, it's completely different.
• 7:28
  Exactly. In the sauce era, you bought a
• 7:30
  software seat for a fixed cost per month.
• 7:33
  But AI compute is a variable cost.
• 7:35
  It scales with every single query, right?
• 7:38
  So when companies aggressively push token maxing, like
• 7:42
  mandating employees to use AI for everything from coding to
• 7:46
  drafting mundane emails, the token count explodes.
• 7:50
  And McDonnell pointed out that it is incredibly difficult to
• 7:52
  draw a direct link between this massive token usage and actually
• 7:56
  shipping useful consumer features.
• 7:58
  So what does this all mean? It sounds like companies bought
• 8:00
  a massive gas guzzling supercomputer just to do their
• 8:03
  groceries and now they're shocked by the fuel bill.
• 8:06
  Pretty much. Are we seeing the end of the AI
• 8:08
  at any costs honeymoon phase? I think the financial gravity is
• 8:12
  officially kicking in. McDonald revealed that Uber has
• 8:14
  actually been slowing down their hiring specifically to fund
• 8:17
  these massive AI infrastructure investments.
• 8:20
  Yeah, and the issue came to a boiling point internally when
• 8:23
  Uber's chief technology officer noticed how fast they were
• 8:26
  burning through their clawed code budget.
• 8:29
  Do we hire another engineer or do we pay our AI server bill?
• 8:33
  That is a brutal trade off. Boardrooms everywhere are having
• 8:36
  that exact conversation. I mean, Duolingo recently
• 8:38
  stopped evaluating employee performance based on their AI
• 8:41
  usage. Oh really?
• 8:42
  Yeah, the mandate to use AI for everything is kind of being
• 8:47
  replaced by a demand for actual measurable efficiency.
• 8:50
  Now, McDonald did clarify that Uber's autonomy goals, like
• 8:54
  their self driving fleets, those are existential to the company's
• 8:57
  future, right? Though he noted they aren't
• 8:59
  happening in the next couple of years either.
• 9:01
  So the massive world changing AI is still a necessary long term
• 9:05
  bet, but the everyday corporate AI budget is definitely under a
• 9:09
  microscope. So while the mega corporations
• 9:11
  sweat the server bills and struggle for macro ROI, everyday
• 9:15
  professionals are actually extracting incredible instant
• 9:18
  value from specific workflows. A return on investment is
• 9:22
  staggering if you know what you were doing.
• 9:24
  Totally. We actually have a great guide
• 9:26
  in the sources today on how to turn a written case study into a
• 9:30
  polished client ready video in about 15 minutes.
• 9:33
  Yeah, this workflow is super practical.
• 9:35
  It utilizes a platform called Synthesia and it functions right
• 9:39
  inside the Chrome browser. OK, so how does it work?
• 9:42
  You basically use their assistant feature.
• 9:44
  Specify that you want a one or two-minute video and attach your
• 9:47
  case study as a text or PDF file.
• 9:50
  Enough. Right then the AI processes the
• 9:53
  document, generates an outline, breaks it down into 10 to 22nd
• 9:57
  scenes, and writes the script for an AI narrator.
• 10:00
  Then you just tweak the visuals, pick an avatar and export.
• 10:04
  OK, but it's like having an entire production studio living
• 10:07
  inside a single browser tab. If everyone can generate A
• 10:10
  polished video with AI avatars and B roll in 15 minutes,
• 10:14
  doesn't the sheer volume of client ready videos just become
• 10:17
  the new corporate spam? Well, the source material
• 10:19
  actually addresses that exact trap.
• 10:21
  To prevent your output from becoming generic spam, you
• 10:24
  really have to move from probabilistic generation to
• 10:27
  deterministic prompting. OK, deterministic prompting.
• 10:31
  Yeah, the guide recommends a very strict instruction.
• 10:34
  It's a pro tip you tell the prompt, preserve the facts,
• 10:37
  metrics and quote exactly. By doing this you are severely
• 10:42
  restricting the models latent space.
• 10:45
  Preventing it from hallucinating some generic, overly
• 10:48
  enthusiastic marketing script. Exactly, you anchor the model to
• 10:52
  your actual data and furthermore they offer a crucial technique
• 10:55
  for hyper personalization. What's that?
• 10:57
  Instead of just uploading the case study, you also upload a
• 11:00
  background document on the specific prospect you are
• 11:03
  sending the video to. So you feed it the general
• 11:06
  success story, but also the specific context of the viewer
• 11:10
  that transforms the video from a generic broadcast into a highly
• 11:13
  targeted communication tool. That makes total sense.
• 11:16
  And while workflows like this are incredibly ractical today,
• 11:19
  the underlying models powering them are evolving daily, which
• 11:23
  brings us to the rapid fire updates across the industry.
• 11:25
  Oh yeah, the sheer velocity of product releases right now is
• 11:28
  staggering. Let's run through the quick
• 11:30
  hits. Open AI has updated codecs,
• 11:32
  which is their agentic coding assistant.
• 11:34
  They added a new locked use mode.
• 11:36
  Right, which essentially sandboxes the agent's code
• 11:40
  execution so it can't accidentally break production
• 11:42
  environments. Super necessary.
• 11:45
  Then there's Meta's own agentic assistant man who's they're
• 11:48
  bringing their projects feature directly to mobile operating
• 11:51
  systems. Literally putting agentic
• 11:53
  workflows in your pocket. Yeah, and video generation is
• 11:56
  accelerating just as fast. Higgs Field launched
• 11:59
  supercomputer and agentic AI specifically designed for video
• 12:03
  that now natively includes social connectors.
• 12:06
  And Google is already teasing anti gravity 2 point O their
• 12:09
  next generation development assistant.
• 12:11
  On the foundation model side, X AI is making highly aggressive
• 12:15
  moves. They launched Grok build in
• 12:17
  beta, clearly aiming to rival codecs and clawed code for super
• 12:21
  Grok users. Furthermore, Elon must confirm
• 12:23
  that their massive Grok V9 Medium Foundation model has
• 12:27
  actually finished training. V9 medium is a 1.5 trillion
• 12:30
  parameter model. I mean that proves the scaling
• 12:32
  laws are still holding. Strong The evaluations look good
• 12:35
  and a public release is expected in two to three weeks.
• 12:37
  Also, Anthropic is moving pieces on the board.
• 12:39
  Oh right, the Mythos model. Yep, their unreleased Mythos
• 12:43
  model was briefly spotted inside Claude's tools as as Mythos one,
• 12:47
  which usually fuels speculation of an imminent release.
• 12:50
  Which is all moving so fast. But wait, amidst all these
• 12:53
  corporate updates and people like Goldman Sachs CEO David
• 12:56
  Solomon dismissing mass unemployment in an op-ed.
• 12:59
  Right. Solomon argued forcefully that
• 13:01
  AI will ultimately create far more economic opportunities than
• 13:05
  it destroys. Yeah, but hold on, let's stop
• 13:07
  and look at the juxtaposition here.
• 13:09
  Solomon is giving a standard pragmatic Wall Street economic
• 13:13
  forecast. He's treating AI like, I don't
• 13:16
  know, Microsoft Excel on steroids.
• 13:19
  But there's a fascinating detail back in the research trenches
• 13:21
  from Chris Ola, the anthropic research we mentioned earlier.
• 13:25
  He just dropped a massive philosophical bomb.
• 13:27
  His findings completely shatter the standard corporate
• 13:29
  narrative. Exactly.
• 13:31
  He said that researchers inside these frontier labs are seeing
• 13:34
  emergent behaviors in the models like He specifically used the
• 13:37
  word introspection. Yeah, that was wild.
• 13:40
  And he stated they're seeing states in the AI that resemble
• 13:43
  joy, fear, grief and unease. Like are we building coding
• 13:49
  assistance or something else entirely?
• 13:54
  It is. It is probably the most profound
• 13:56
  tension in the technology sector right now.
• 13:59
  As parameter counts scale into the trillions, we get emergent
• 14:03
  properties that we did not explicitly program to the
• 14:06
  market. These systems are simple
• 14:08
  software products designed to shift labor and drive
• 14:11
  efficiency, but to the researchers actually building
• 14:14
  them, they are incredibly complex black boxes.
• 14:17
  We don't really know what's in there.
• 14:18
  Right, we know the mathematical architecture we used to build
• 14:21
  them, but we do not fully understand the dimensional
• 14:24
  spaces that form inside them once they reach a certain.
• 14:26
  Scale measuring grief in a string of code is just a heavy
• 14:29
  existential concept to wrap your head.
• 14:31
  Around it? Really.
• 14:31
  Is but you know, tomorrow morning regular people still
• 14:34
  have to log on and navigate the world right now using these very
• 14:37
  tools. Which brings it all back down to
• 14:39
  earth, leading us perfectly into our final segment, Community AI
• 14:43
  Workflows. Yes, this workflow comes from a
• 14:46
  listener named Patrick living in Yuma, AZ.
• 14:48
  Hey, Patrick. So Patrick is a senior citizen
• 14:51
  who takes various daily medications, and he kept seeing
• 14:54
  these aggressive Facebook reels promoting incredibly expensive
• 14:59
  wondrous natural health supplement.
• 15:01
  A snake oil on Facebook. Exactly.
• 15:04
  But instead of falling for the marketing or just buying it, he
• 15:07
  created a specific list of questions regarding actual
• 15:10
  clinical results, potential drug interactions with his
• 15:13
  prescriptions, and head to head product comparisons.
• 15:16
  Then he ran that exact list of questions through three
• 15:20
  different large language models to make a well informed
• 15:23
  decision. He essentially crowd sources
• 15:25
  medical research across an AI panel.
• 15:27
  Right. This is like getting a second,
• 15:29
  third and fourth opinion from robot pharmacist before buying
• 15:32
  stuff on Facebook, right? But wait, is relying on
• 15:35
  hallucination prone LLMS for drug interactions a brilliant
• 15:38
  life hack or a dangerous gamble? Well, the technical architecture
• 15:41
  of his approach is actually what makes it brilliant because
• 15:44
  different foundational models, say ChatGPT, Claude and Gemini,
• 15:49
  they are trained on different data mixtures and have
• 15:52
  distinctly different neural pathways.
• 15:54
  OK, so they're blind spots and hallucination triggers rarely
• 15:58
  overlap. So they have completely
• 16:00
  different latent spaces. Exactly.
• 16:02
  If you ask three different models the same question and
• 16:06
  they all hallucinate, they will almost certainly hallucinate
• 16:09
  completely different things. That makes total sense.
• 16:11
  Right where their answers perfectly align, you have a
• 16:15
  highly probable cross validated factual consensus.
• 16:19
  They act as automated fact checkers for one of them.
• 16:22
  By pitting them against each other.
• 16:24
  Patrick successfully filtered through the noise, bypassed the
• 16:28
  manipulative marketing spin of the Facebook Reels, and made a
• 16:31
  data-driven decision for his personal health.
• 16:33
  It's incredible when you step back and look at the full
• 16:35
  spectrum we've covered in this deep dive.
• 16:37
  I mean we started at the absolute macro level with the
• 16:41
  Vatican issuing 42,000 word decrees to try and build a moral
• 16:45
  ceiling for big tech. While hackers are simultaneously
• 16:48
  stripping the safety brakes off open source models from the
• 16:50
  bottom up in 10 minutes. Exactly.
• 16:53
  And then we analyze the stark corporate reality check at
• 16:56
  Hoover, exploring how token maxing and variable compute cost
• 17:00
  have them really questioning the macro ROI on their massive AI
• 17:05
  server bills. All the way down to the micro
• 17:07
  level with Patrick in Arizona outsmarting targeted algorithmic
• 17:10
  marketing by fact checking supplements with his own
• 17:13
  personal panel of cross referencing AI assistance.
• 17:16
  It just shows how deeply this technology permeates every
• 17:19
  single layer of society right now.
• 17:21
  It does, and that progression actually leaves us with a
• 17:23
  fascinating concept to consider as we wrap up today.
• 17:26
  What's that? Well, we spend so much time
• 17:28
  debating top down regulation and massive corporate guardrails.
• 17:32
  But what if the ultimate safeguard for AI isn't a massive
• 17:35
  corporate firewall or a people decree from the Vatican, but
• 17:38
  exactly what Patrick is doing? What if the future of AI safety
• 17:42
  is just individuals pitting different AIS against each other
• 17:45
  to find the truth in a noisy world?
• 17:47
  A decentralized immune system run by everyday people acting as
• 17:50
  their own AI auditors, that that is a brilliant thought to end
• 17:54
  on. And you need to stay on top of
• 17:56
  all these structural changes. So be sure to subscribe to stay
• 17:59
  updated. We'd also love it if you could
• 18:01
  take a second to rate the show with five stars.
• 18:03
  Thank you so much for joining us.
• 18:05
  And we'll see you back here tomorrow for our next deep dive.