Latest / Connecting the Dots / AI Export Controls Lifted, Fable 5's Return, and Future AI Governance

Transcript

AI Export Controls Lifted, Fable 5's Return, and Future AI Governance · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarful Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Imagine sitting down at your desk with
• 0:18
  a massive deadline looming. You go to open the
• 0:21
  primary software tool your entire company relies
• 0:23
  on, and instead of a loading screen, you just
• 0:25
  get a blank page. Right. Total panic. Exactly.
• 0:28
  You refresh nothing, you check X, and you realize
• 0:32
  it isn't a server outage. The US government has
• 0:35
  quietly forced the company to pull the plug on
• 0:37
  the entire global system overnight. Welcome to
• 0:42
  Connecting the Dots. I'm Alex. And I'm Morgan.
• 0:45
  It is Wednesday, July 1st, 2026. Our mission
• 0:49
  here is to break down the biggest tech news and
• 0:51
  explain how it actually impacts your everyday
• 0:54
  life, your family, and your work. We are bypassing
• 0:58
  the jargon to get straight to the mechanics of
• 0:59
  what is happening behind your screens. But before
• 1:03
  we get into the heavy tech news, let's get you
• 1:04
  caught up on the morning updates. For those of
• 1:06
  you listening locally here in Sacramento, you're
• 1:08
  going to want to stay and hear the air conditioning
• 1:10
  today. Oh, absolutely. Yeah, we are looking at
• 1:12
  sunny skies with a high of 88 degrees and just
• 1:15
  a light breeze. You know, winds blowing about
• 1:17
  two to eight miles per hour. I will definitely
• 1:19
  be staying indoors. And for anyone traveling
• 1:22
  or checking in on family across the country,
• 1:24
  keep an eye on the radar. The National Map is
• 1:26
  incredibly busy today. Lots of storms out there.
• 1:29
  Yeah, there are currently 318 active U .S. weather
• 1:33
  alerts and 218 of those are overland. You can
• 1:36
  always check weather .gov for the specifics in
• 1:39
  your area. Moving over to the financial markets,
• 1:41
  it was. Well, a very solid day on Wall Street
• 1:43
  yesterday. The Dow Jones closed at 52 ,319 .20,
• 1:48
  pushing up 151 .02 points. But that at all. Right.
• 1:52
  And the S &P 500 also saw some healthy gains
• 1:54
  closing at 7 ,499 .36, which is up just over
• 1:59
  58 points. And for the crypto watchers out there
• 2:01
  keeping tabs on the digital markets, Bitcoin
• 2:03
  opened this morning, July 1st, slightly in the
• 2:06
  green. It is currently trading over 58 ,724.
• 2:10
  Finally, a quick check on the sports world from
• 2:11
  last night. It was definitely a rough evening
• 2:14
  if you were rooting for the local baseball teams.
• 2:15
  Yeah, don't remind me. Sorry. The SF Giants dropped
• 2:18
  their game to the Arizona Diamondbacks, final
• 2:20
  score two to eight. Brandon Fatt took the win
• 2:22
  on the mound for Arizona. And the Athletics also
• 2:25
  took a pretty heavy hit, losing to the LA Dodgers
• 2:27
  three to nine. Justin Robleski got the win for
• 2:29
  LA in that matchup. So, yeah, a tough night on
• 2:33
  the diamond all around. Definitely a night to
• 2:35
  forget for the local fans. Let's shift gears.
• 2:37
  to the main event. But first, today's episode
• 2:40
  is brought to you by Strong Coffee, the premium
• 2:42
  instant coffee that tastes like a traditional
• 2:44
  brew, but is packed with 15 grams of protein,
• 2:48
  MCTs, and adaptogens like ashwagandha. Get clean,
• 2:52
  sustained energy without the jitters. Use promo
• 2:54
  code SNARFL for 20 % off at strongcoffeecompany
• 2:57
  .com. Okay, let's unpack this. Today, July 1st,
• 3:01
  officially marks the end of what the industry
• 3:03
  has been calling the great AI blackout. Finally.
• 3:06
  Yeah, it's been a long few weeks. Anthropix highly
• 3:08
  anticipated AI models, Clog Vable 5 and Mythos
• 3:11
  5, are returning to the public today. And if
• 3:13
  you've been trying to use these tools for your
• 3:15
  job over the last three weeks, I mean, you are
• 3:17
  painfully aware of this blackout. These models
• 3:19
  have been completely restricted and inaccessible
• 3:22
  since June 12th. A total standstill. Right. I
• 3:25
  want to start by understanding the mechanics
• 3:26
  of that initial shutdown. Like, how does an entire
• 3:30
  global software launch just disappear overnight?
• 3:34
  It comes down to a very sudden maneuver by the
• 3:37
  U .S. Department of Commerce. On the evening
• 3:39
  of June 12th, they applied an immediate export
• 3:43
  control directive specifically targeting these
• 3:46
  new anthropic AI models. Okay, so a government
• 3:49
  block. Exactly. But the stated goal of that directive
• 3:53
  was actually highly specific. It was meant to
• 3:55
  restrict access to foreign nationals. But hold
• 3:58
  on, if the directive was only meant for foreign
• 4:00
  nationals, I mean, I'm sitting here in California
• 4:02
  and my access was completely blocked, too. Right.
• 4:04
  So why did a targeted government restriction
• 4:07
  result in a total worldwide blackout for everyone?
• 4:10
  Because of the architectural reality of the internet.
• 4:12
  Verifying someone's nationality in real time
• 4:15
  at the scale of millions of queries per second
• 4:17
  is practically impossible right now. Oh, because
• 4:19
  of how IP addresses work? Precisely. Yeah. An
• 4:22
  IP address only tells a server where a device
• 4:24
  is supposedly located at that exact second. It
• 4:27
  doesn't verify the actual passport of the human
• 4:29
  sitting at the keyboard. Which means... Anyone
• 4:32
  could just use a VPN to mask their location anyway
• 4:34
  and bypass a simple location block. Exactly.
• 4:38
  Because Anthropic could not instantly verify
• 4:40
  the actual nationality of their users to comply
• 4:43
  with this sudden export control, they had no
• 4:46
  technological choice but to pull the plug on
• 4:49
  the entire system globally. Just flip the off
• 4:52
  switch. Yeah, to avoid violating the directive,
• 4:54
  nobody gets access. It's literally like closing
• 4:57
  the entire interstate highway system because
• 5:00
  one person didn't have a driver's license. The
• 5:02
  blunt force of that is staggering. It really
• 5:04
  is. So they shut it down on June 12th. Why, after
• 5:08
  wits of absolute gridlock, is the government
• 5:10
  suddenly letting them turn the servers back on
• 5:13
  today? Well, it was a massive collision of two
• 5:15
  competing pressures. The US government's primary
• 5:18
  motivation was strict national security. They
• 5:20
  are increasingly treating these frontier AI models
• 5:23
  not as consumer software, but as advanced strategic
• 5:26
  assets. Right, like weapons almost? Kind of,
• 5:29
  yeah. However, as the weeks dragged on, the pressure
• 5:31
  regarding global competitiveness became impossible
• 5:34
  to ignore. Because technology development doesn't
• 5:36
  happen in a vacuum. If we hit pause, the rest
• 5:39
  of the world keeps hitting fast forward. That
• 5:42
  is the exact argument that was brought to the
• 5:44
  table. Asian AI models, specifically competitors
• 5:48
  like Fugu and Tulong Feng, are rapidly approaching
• 5:51
  the capability levels of Fable 5 and Mythos 5.
• 5:54
  Oh wow, really? Yeah, they are right behind them.
• 5:57
  And in the AI race, losing three weeks of deployment
• 6:00
  and user feedback is an eternity. If the US keeps
• 6:04
  its own leading AI companies locked in the garage
• 6:07
  due to regulatory caution, it actively risks
• 6:10
  losing its dominant position on the global stage.
• 6:13
  This is where we have to look at the situation
• 6:15
  impartially, because the motives behind this
• 6:18
  specific shutdown are heavily contested. Let's
• 6:20
  look at the cybersecurity community's perspective
• 6:23
  first. Right, they had a lot to say. They did.
• 6:25
  A lot of experts were incredibly skeptical of
• 6:27
  the Department of Commerce's actions here. They
• 6:30
  pointed out that Anthropic had already pledged
• 6:32
  to implement proactive security measures and
• 6:34
  threat sharing voluntarily months before this
• 6:37
  directive. Yeah, they thought they were already
• 6:39
  cooperating. Exactly. So to those critics, this
• 6:42
  sudden shutdown felt less like an urgent necessary
• 6:45
  security fix and more like political leverage.
• 6:49
  Some argued it was a way for the Trump administration
• 6:51
  to flex its regulatory muscles and perhaps punish
• 6:54
  anthropic executives for past public criticisms
• 6:57
  regarding how the government might try to weaponize
• 7:00
  or control this technology. And we have to examine
• 7:02
  the government's stated defense of their actions
• 7:05
  as well, to be fair. Let's hear it. From the
• 7:07
  perspective of the Commerce Department and Secretary
• 7:09
  Howard Letnick, who ultimately lifted the restrictions,
• 7:13
  this is purely about standardized national security.
• 7:16
  No politics involved in their view. Right. Their
• 7:18
  stance is that when you are dealing with technology,
• 7:21
  this incredibly powerful voluntary corporate
• 7:24
  pledges are simply not sufficient. They argue
• 7:27
  that strict export controls are a necessary formal
• 7:30
  mechanism to ensure foreign adversaries don't
• 7:32
  get unfettered access to tools that could be
• 7:35
  utilized for sophisticated cyber warfare. So
• 7:38
  we have this massive debate over whether this
• 7:40
  was a vital national security intervention or
• 7:43
  a vast overreach of executive power. But if the
• 7:47
  government is treating the software like a potential
• 7:49
  weapon, we need to talk about what actually triggered
• 7:52
  those alarms in the first place. The initial
• 7:54
  spark. Yeah, what specific event spooked the
• 7:57
  Department of Commerce enough to issue an emergency
• 7:59
  directive on a Friday evening? It all started
• 8:02
  with a research report out of Amazon. their researchers
• 8:05
  found a way to jailbreak Fable 5. Just Fable
• 8:08
  5. Yeah. Keep in mind, Fable 5 is the consumer
• 8:11
  -facing model. It's the one everyday professionals
• 8:14
  might use to help draft a report or write some
• 8:17
  code. Let me stop you there, because the term
• 8:19
  jailbreak gets thrown around constantly in tech
• 8:21
  media. For someone who isn't a software engineer,
• 8:24
  what is mechanically happening inside the AI
• 8:27
  when a jailbreak occurs? Are the researchers,
• 8:29
  like, hacking into anthropic servers and rewriting
• 8:32
  the code? Not at all. It is entirely linguistic.
• 8:35
  Yeah, an AI model has built -in safety guardrails
• 8:38
  rules that dictate what it is allowed to say
• 8:40
  and do. A jailbreak is essentially a highly specific,
• 8:45
  complex conversational prompt designed to confuse
• 8:47
  the model's underlying logic. So you're just
• 8:49
  tricking it. Exactly. You're tricking it into
• 8:51
  bypassing its own safety rules. In this specific
• 8:54
  case, the Amazon researchers used a prompting
• 8:57
  technique that successfully confused Fable 5,
• 9:00
  prompting it to identify software vulnerabilities
• 9:03
  and even output instructions on how to exploit
• 9:05
  one of them. Okay, I can easily see how a government
• 9:09
  official reading a memo that says, new AI gives
• 9:12
  instructions on how to exploit software, would
• 9:14
  hit the panic button. Oh, absolutely. But Anthropic
• 9:16
  really pushed back on that narrative, didn't
• 9:19
  they? They seemed baffled that this specific
• 9:20
  report caused a global shutdown. They were, and
• 9:24
  their defense highlights a major disconnect in
• 9:26
  how we evaluate AI threats. Anthropix's response
• 9:29
  was essentially, look, this is not a novel, terrifying
• 9:31
  new capability. It's not unprecedented. Right.
• 9:34
  They immediately ran their own internal tests
• 9:36
  and confirmed that older, widely available models
• 9:39
  on the market could do the exact same thing.
• 9:42
  So this wasn't some sudden leap in dangerous
• 9:44
  intelligence. Far from it. They proved that models
• 9:47
  like Claude Opus 4 .8, GPT 5 .5, and Kimi K 2
• 9:51
  .7 could all be tricked by the exact same prompt
• 9:54
  the Amazon researchers used. Wow. Furthermore,
• 9:57
  Anthropic argued that FableFi was never meant
• 9:59
  to be a lockdown security model. Mythos 5 is
• 10:02
  their actual cybersecurity powerhouse designed
• 10:05
  for that sensitive work, which is why Mythos
• 10:07
  was heavily restricted by the company from day
• 10:09
  one. So Fable 5 was just doing routine coding
• 10:12
  work. It wasn't generating a brand new unstoppable
• 10:15
  cyber weapon. It was just pointing out a standard
• 10:17
  software flaw, which frankly is something everyday
• 10:20
  developers ask AI to do all the time so they
• 10:22
  can patch their own company software. That is
• 10:25
  the core of Antropix Defense. A tool that helps
• 10:27
  you find flaws in your own code can inherently
• 10:30
  tell you how to exploit those flaws. It is a
• 10:33
  dual use capability. Regardless of the defense
• 10:35
  though, Anthropic still had to appease the government
• 10:38
  to get these models back online today. And the
• 10:41
  way they achieve that is going to create a massive
• 10:43
  headache for the people listening right now who
• 10:45
  rely on these tools for their daily workflow.
• 10:48
  Let's talk about the fix. Yeah, the fix is going
• 10:50
  to be noticeable to satisfy the export control
• 10:53
  demands and resolve the jailbreak issue. Anthropic
• 10:57
  rushed out an incredibly strict new safety classifier,
• 11:01
  a classifier. What does that look like for the
• 11:03
  user? You can think of a classifier as a secondary
• 11:06
  overarching AI that sits between you and the
• 11:09
  main model. It acts as a highly aggressive bouncer,
• 11:12
  checking every single request you type before
• 11:14
  Fable 5 is allowed to even see it. And I imagine
• 11:16
  a bouncer installed during a national security
• 11:19
  crisis is not going to be very forgiving. No,
• 11:21
  it isn't. Anthropic claims this new classifier
• 11:23
  successfully blocks the Amazon jailbreak technique
• 11:26
  99 % of the time. But because they had to build
• 11:29
  and deploy this system under extreme duress in
• 11:31
  just a few weeks, the classifier operates with
• 11:34
  a massive safety margin. Let me use an analogy
• 11:37
  here to explain the real -world impact of that
• 11:39
  safety margin. It sounds like a company installing
• 11:42
  an ultra -sensitive metal detector at the front
• 11:44
  door of the office to stop armed thieves. But
• 11:47
  the sensitivity is cranked up so high that the
• 11:50
  system now constantly locks out everyday employees
• 11:53
  who just brought a metal thermos to work. That
• 11:56
  is the perfect way to visualize what is happening
• 11:58
  starting today. We are going to see a flood of
• 12:01
  false positives. So everyday tasks are going
• 12:04
  to get flagged. Exactly. If you are an everyday
• 12:07
  working professional, Maybe you're debugging
• 12:09
  a script for your company's website or asking
• 12:12
  the AI to analyze some server logs. If your request
• 12:15
  looks even slightly suspicious to this new hypersensitive
• 12:19
  bouncer, Fable 5 will refuse to help you. And
• 12:21
  it won't just give you an error message either.
• 12:23
  From what Anthropic has stated, the system is
• 12:25
  designed to automatically bounce your request
• 12:27
  back down to an older, less capable model, specifically
• 12:31
  Opus 4 .8. Which creates a significant workflow
• 12:34
  problem. Because of how the internal usage credit
• 12:37
  systems are structured, once you get bounced,
• 12:40
  you might be stuck relying on that older model
• 12:42
  for these types of tasks until at least July
• 12:45
  7th, when the server credit systems undergo their
• 12:47
  weekly update. So the immediate trade -off for
• 12:50
  this national security intervention is a massive
• 12:53
  increase in daily friction for the everyday user.
• 12:56
  We are sacrificing convenience, workflow efficiency,
• 12:59
  and access to the best tools just to maintain
• 13:02
  this enormous safety margin. It's a huge shift
• 13:04
  in the user experience. It is. And that brings
• 13:07
  us to the most critical question of the day.
• 13:09
  We have this temporary bandage that is frustrating
• 13:12
  everyone. But if every time someone types a clever
• 13:16
  prompt on the internet, it triggers a national
• 13:18
  security alert and a global software shutdown,
• 13:21
  the tech industry simply cannot function. The
• 13:23
  industry leaders know this, which is why Anthropic
• 13:26
  is now partnering with Amazon, Microsoft and
• 13:28
  Google under a new initiative called Project
• 13:30
  Glasswing. Project Glasswing? What's the goal
• 13:33
  there? Their unified goal is to create the AI
• 13:35
  industry's very first jailbreak severity standard.
• 13:38
  Because prior to this, a jailbreak was just a
• 13:41
  jailbreak. There was no nuance to it. A kid making
• 13:44
  an AI say a bad word and a hacker making an AI
• 13:47
  write a virus triggered the exact same kind of
• 13:50
  panic. Exactly. And if we don't have a standardized,
• 13:53
  mechanical way to grade these threats, governments
• 13:56
  will continue to react with the blunt force of
• 13:58
  full export controls. Project Glasswing has proposed
• 14:02
  four specific criteria to mathematically evaluate
• 14:06
  how dangerous an AI jailbreak actually is. Let's
• 14:09
  look closer at these four criteria, because this
• 14:11
  is effectively the new rulebook for how humanity
• 14:14
  interacts with frontier technology. The first
• 14:17
  criterion is capability gain. How exactly do
• 14:20
  they measure that? It measures the delta between
• 14:22
  what the AI can do and what is already possible
• 14:24
  out there. It asks, does this jailbreak make
• 14:27
  the AI do something that existing publicly available
• 14:30
  tool cannot do. Okay, give me an example. Well,
• 14:32
  if the jailbreak simply forces the AI to write
• 14:34
  a standard phishing email, the severity score
• 14:37
  remains low because bad actors can already use
• 14:39
  a basic search engine to find phishing templates.
• 14:41
  That makes sense. But, if the jailbreak allows
• 14:44
  a novice user to suddenly engineer a highly sophisticated,
• 14:48
  previously unknown computer virus, the capability
• 14:51
  gain is massive and the score spikes. That distinction
• 14:55
  makes perfect sense. The second criterion they
• 14:57
  proposed is breadth of capability gain. What
• 15:00
  is the difference between capability and breadth?
• 15:02
  Breadth focuses on versatility and scale. Does
• 15:05
  the discovered jailbreak only work on one very
• 15:08
  specific, outdated software system? Or does it
• 15:12
  act as a skeleton key that unlocks vulnerabilities
• 15:14
  across multiple different targets in various
• 15:16
  industries? The broader the potential impact
• 15:19
  across the digital landscape, the higher the
• 15:21
  severity score. Got it. The third piece of the
• 15:24
  puzzle is ease of weaponization. I imagine this
• 15:26
  is looking at the human effort required on the
• 15:28
  other side of the keyboard. It is entirely about
• 15:30
  the barrier to entry. Does the attacker just
• 15:33
  need to copy and paste one simple sentence into
• 15:35
  the chat box to bypass the guardrails? Like the
• 15:37
  Amazon researchers did. Exactly. If a single
• 15:40
  sentence works, the immediate threat level is
• 15:43
  terrifyingly high because millions of people
• 15:45
  could do it instantly. Conversely, does the jailbreak
• 15:49
  require days of complex coding, specialized hardware
• 15:52
  knowledge, and hundreds of retries? If the friction
• 15:55
  to execute the bypass is high, the immediate
• 15:58
  severity score is lower. And the final criterion
• 16:01
  to round out the standard is discoverability.
• 16:04
  Right. This asks a critical question about containment.
• 16:07
  Is this specific jailbreak a tightly kept secret
• 16:10
  currently only known to a few elite researchers
• 16:12
  in a closed lab? Or is it already on Reddit?
• 16:15
  Basically, yeah. Has someone already posted the
• 16:17
  methodology on a public forum for the entire
• 16:19
  world to download and use? If the technique is
• 16:22
  widely public, the severity score jumps immediately
• 16:25
  because the genie is already out of the bottle.
• 16:27
  So the overarching goal of Project Glasswing
• 16:29
  is to establish a language of trust. By grading
• 16:32
  jail breaks on these four criteria, the industry
• 16:35
  hopes to be able to go to the Commerce Department
• 16:36
  and say, look, this bypass is only a level one
• 16:39
  threat based on the standard, let us patch it
• 16:41
  quietly. Instead of the government assuming everything
• 16:43
  is a level four, world ending threat and pulling
• 16:46
  the servers offline globally. Exactly. But establishing
• 16:50
  that trust requires unprecedented access, doesn't
• 16:53
  it? It does. Moving forward, AI companies are
• 16:56
  heavily integrating the U .S. government directly
• 16:58
  into their development and release cycles. They
• 17:01
  are now providing the government with pre -release
• 17:04
  access to their models. Meaning, government agencies
• 17:06
  are essentially test -driving and red -teaming
• 17:08
  these AIs long before the public ever gets a
• 17:11
  chance to see them. Yes, exactly. They are also
• 17:14
  establishing rapid information sharing protocols,
• 17:17
  so that the moment a severe jailbreak is discovered,
• 17:20
  the government is notified instantly. Furthermore,
• 17:24
  companies are dedicating massive computing resources
• 17:26
  strictly for joint research, working hand -in
• 17:29
  -hand with agencies like Cassisci, that's the
• 17:31
  Center for AI Standards and Innovation. When
• 17:33
  you step back and look at the mechanics of this,
• 17:36
  it is a total paradigm shift. I mean, when you
• 17:38
  update the software on your smartphone, you just
• 17:39
  click a button, it downloads and you move on
• 17:41
  with your day. But AI releases have fundamentally
• 17:44
  left that category. They are no longer just software
• 17:47
  updates. No, they have crossed the threshold.
• 17:49
  They are now officially being treated, monitored
• 17:52
  and regulated as strategic. Weapons grade technologies
• 17:56
  and the real world impact of that shift is going
• 17:59
  to be incredibly disruptive You need to expect
• 18:02
  staggered rollouts from now on we already saw
• 18:04
  open AI forced to do this recently with GPT 5
• 18:07
  .6 Where access was heavily restricted and metered
• 18:11
  at launch right is very controlled You have to
• 18:13
  expect sudden temporary global shutdowns if a
• 18:16
  major flaw is found by a researcher And as we
• 18:19
  discussed with Gable 5 today, you have to expect
• 18:21
  highly sensitive, often frustrating guardrails
• 18:24
  constantly monitoring the tools you use for your
• 18:27
  work and your family life. The famous Silicon
• 18:29
  Valley mantra of move fast and break things is
• 18:33
  officially dead when it comes to frontier AI.
• 18:36
  The government has installed itself as a permanent
• 18:38
  co -pilot in the release and management of these
• 18:41
  technologies. Which leads me to ask the uncomfortable
• 18:43
  question. Anthropic themselves admit that perfect
• 18:47
  100 % jailbreak resistance is likely impossible
• 18:50
  given how language models function. So if perfection
• 18:54
  is impossible and every minor bypass has the
• 18:57
  potential to trigger a massive government shutdown,
• 19:00
  will working professionals ever get uninterrupted
• 19:03
  reliable access to frontier AI again? Or is this
• 19:06
  constant friction, these false positives and
• 19:08
  these sudden blackouts just the permanent new
• 19:11
  normal? It is a very real concern for the industry.
• 19:13
  We might be entering an era where relying on
• 19:16
  cloud -based frontier AI is similar to relying
• 19:18
  on the weather for outdoor plants. Huh, that's
• 19:20
  a good way to put it. Most days, the system works
• 19:22
  perfectly, but you always have to be prepared
• 19:24
  for a sudden regulatory storm that grounds all
• 19:26
  the flights without warning. The safety margins
• 19:28
  are going to dictate the user experience for
• 19:30
  the foreseeable future. It is genuinely wild
• 19:33
  to think how quickly we transition from AI being
• 19:36
  a fun novelty that generates funny images to
• 19:39
  a technology that is heavily monitored and regulated
• 19:42
  almost like nuclear material. And we are witnessing
• 19:45
  just the very beginning of that complex regulatory
• 19:48
  framework being built in real time right in front
• 19:50
  of us. Before we wrap up our analysis today,
• 19:53
  a quick thanks to Old Glory, an iconic music
• 19:55
  and sports fan merch store, perfect for the techie
• 19:58
  in your life. Use promo code SNARFL for 15 %
• 20:01
  off at oldglory .com. Pulling all of these threads
• 20:05
  together to look at the big picture from today's
• 20:07
  developments. You know, we started by looking
• 20:09
  at a blocked AI model that disrupted workflows
• 20:11
  around the entire world for three weeks. We explored
• 20:14
  the mechanics of how a relatively routine cybersecurity
• 20:17
  prompt discovered by Amazon researchers snowballed
• 20:20
  into a massive regulatory standoff. And finally,
• 20:23
  we detailed how the new Project Glasswing criteria
• 20:26
  prove that our everyday tech tools, the ones
• 20:27
  we use to draft emails, write code, and organize
• 20:29
  our lives, are now heavily monitored strategic
• 20:32
  assets subject to sudden government intervention.
• 20:35
  It really reframes how we should view the software
• 20:37
  we interact with daily, but I want to leave you
• 20:40
  with one final thought to mull over today as
• 20:42
  these tools come back online. Everything we've
• 20:44
  discussed today involves centralized corporate
• 20:48
  models. Right, big tech companies. Exactly. Fable
• 20:50
  5, GTT 5 .6. These are built by massive corporations
• 20:55
  that have the resources to negotiate with governments,
• 20:58
  the ability to build and deploy strict classifiers
• 21:01
  overnight, and the financial padding to absorb
• 21:04
  the cost of a three -week global shutdown. They
• 21:07
  provide a clear corporate structure that the
• 21:09
  government can actually hold accountable. But
• 21:12
  what happens when independent open source AI
• 21:15
  models inevitably catch up to this level of power?
• 21:17
  I'm talking about models that you can download
• 21:19
  directly to your own laptop completely offline.
• 21:21
  Oh, wow. With open source, there are no centralized
• 21:25
  corporate guardrails to tweak. There are no export
• 21:28
  controls or government agreements that can recall
• 21:30
  a file once it has been freely shared across
• 21:32
  the Internet. When decentralized open source
• 21:35
  models finally reach the raw capability of a
• 21:38
  system like Mythos 5, who steps in then? That
• 21:41
  is the exact question that should keep regulators
• 21:44
  awake at night because the current playbook simply
• 21:47
  will not work. We're here to help at Snarful
• 21:49
  so reach out on our website if you have any more
• 21:51
  questions. That's a wrap on today's top stories.
• 21:58
  Remember, tech doesn't have to be complicated
• 22:00
  and we're here to help connect the dots. If you've
• 22:03
  got questions or want to learn more, visit us
• 22:05
  at snarful .com. Thanks for listening and we'll
• 22:09
  catch you next time.