Latest / Connecting the Dots / AI Decentralization Calls, Getty's Licensing Shift, and Frontier Model Bans

Transcript

AI Decentralization Calls, Getty's Licensing Shift, and Frontier Model Bans · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfl 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 your company handing over its
• 0:18
  entire artificial intelligence budget for the
• 0:21
  year 2026, just like... massive multi -billion
• 0:25
  dollar war chest designed to last 12 full months.
• 0:28
  Right. Now imagine watching every single cent
• 0:31
  of that just evaporate in four months. I mean
• 0:33
  today we're unpacking why the enterprise AI revolution
• 0:36
  is quietly turning into this catastrophic financial
• 0:40
  trap and why the biggest tech giants are suddenly
• 0:43
  warning businesses to pull back. Welcome to Connecting
• 0:46
  the Dots. I'm Alex. And I'm Morgan. We're coming
• 0:48
  to you live from Snarfl Solutions Group in Sacramento,
• 0:51
  California. That's right. Today is Monday, June
• 0:53
  22, 2026. And this is your daily show breaking
• 0:56
  down the latest tech news and really figuring
• 0:59
  out how it actually impacts your everyday life
• 1:01
  and your work. Exactly. But before we get into
• 1:03
  the heavy stuff, should we kick things off with
• 1:05
  the Morning Star Report? Yeah, let's do it. So
• 1:07
  looking out the window here in Sacramento, we
• 1:09
  are in for a hot one. It's going to be sunny,
• 1:11
  hitting a high of 91 degrees Fahrenheit with
• 1:14
  a light breeze, winds somewhere between 3 to
• 1:16
  12 miles per hour. Nationally, however, the weather
• 1:20
  map is incredibly active right now. Oh, yeah,
• 1:22
  I saw that. We are looking at 277 active U .S.
• 1:26
  weather alerts, and 206 of those are over land.
• 1:29
  Wow, that's a lot. Well, on the financial side,
• 1:32
  the markets are a bit of a mixed bag to start
• 1:34
  the week. The Dow closed slightly down on Thursday
• 1:36
  at 51 ,564 .78. But the broader market showed
• 1:41
  resilience with the S &P 500 climbing to close
• 1:44
  up at 7 ,500 .58. OK, not bad. And for the crypto
• 1:48
  watchers out there, Bitcoin actually had a remarkably
• 1:50
  strong weekend. It opened today at 63 ,242 .26
• 1:54
  and has already jumped up to 64 ,985 .38. Nice.
• 1:58
  A nice little weekend bump for the crypto portfolios.
• 2:01
  What about sports? Any good news for the local
• 2:03
  California teams? Unfortunately, no. Brace yourself
• 2:06
  for a collective sigh here. The S of Giants dropped
• 2:08
  a close one, losing 1 -2 against the Miami Marlins,
• 2:10
  and the Athletics also took a hit, losing 7 -9
• 2:12
  in a shootout against the Angels. Ouch. Well,
• 2:16
  you win some, you lose some. Let's pivot to what
• 2:18
  we are really here to explore today, because
• 2:20
  we have a fascinating stack of sources to get
• 2:23
  through. We really do. I mean, a major Wall Street
• 2:26
  Journal interview and a viral XSA from Microsoft's
• 2:29
  CEO? fresh financial disclosures from Getty Images,
• 2:33
  a sweeping new Commerce Department directive
• 2:35
  from the White House, and even some historical
• 2:38
  accounts from the 1990s crypto wars. It's quite
• 2:40
  the list. It is. But if there's one word that
• 2:43
  binds all of the source material together, it's
• 2:46
  control. We are looking at a massive three -front
• 2:49
  power struggle over the tools you use every day.
• 2:52
  who controls the underlying architecture of enterprise
• 2:54
  AI, who controls the massive data pipelines feeding
• 2:57
  those systems, and what happens when the U .S.
• 3:00
  government decides it needs to control exactly
• 3:01
  who gets access. Let's look closer. Today's episode
• 3:03
  is brought to you by Strong Coffee, the premium
• 3:06
  instant coffee that tastes like a traditional
• 3:08
  brew but is packed with 15 grams of protein,
• 3:10
  MCTs, and adaptogens like Ashwagandha. Get clean,
• 3:14
  sustained energy without the jitters. Use promo
• 3:16
  code SNARFL for 20 % off at strongcoffeecompany
• 3:20
  .com. So if Microsoft is arguably the biggest
• 3:23
  beneficiary of the generative AI boom, then their
• 3:26
  CEO, Satya Nadella, just issued a warning that
• 3:29
  should make, frankly, every corporate IT department
• 3:33
  freeze in its tracks. Oh, absolutely. In this
• 3:35
  new Wall Street Journal interview and that incredibly
• 3:38
  detailed essay posted to X, Nadella essentially
• 3:41
  pointed a finger directly at the AI frontier
• 3:44
  labs, specifically OpenAI and Anthropic. Which
• 3:47
  is wild, right? Because the context here is dizzying.
• 3:50
  Microsoft has invested billions into these exact
• 3:52
  companies. Right. But Nadella is out here issuing
• 3:55
  a sharp rebuke, warning that an industry dominated
• 3:58
  by a few massive generalized AI models is not
• 4:02
  just politically dangerous, it's economically
• 4:04
  unsustainable for the businesses relying on them.
• 4:06
  Yeah. And the phrase Nadella anchors his essay
• 4:08
  on is social license to operate, which is, you
• 4:11
  know, it's a concept usually reserved for heavy
• 4:13
  industries like mining or oil extraction. Oh,
• 4:16
  interesting. Yeah. It's the idea that an industry
• 4:18
  needs ongoing implicit approval from the public
• 4:21
  to function. Like if you extract resources without
• 4:24
  enriching the community, you lose that trust,
• 4:27
  which inevitably leads to severe regulation.
• 4:29
  Nadella's argument is that the major AI labs
• 4:32
  are currently burning through their social license
• 4:34
  at an alarming rate without doing anything to
• 4:37
  replenish it. Which completely tracks when you
• 4:40
  look at the contradictory messaging coming out
• 4:42
  of those labs. You have Anthropic's CEO casually
• 4:45
  predicting that AI could eliminate half of all
• 4:48
  entry -level white -collar jobs in just a few
• 4:51
  years. Right, pushing unemployment to 20%. Exactly.
• 4:54
  But in the very next breath, These exact same
• 4:56
  companies are demanding massive new data centers,
• 4:59
  nuclear power allocations, and total regulatory
• 5:02
  freedom so they can keep scaling. Nadella is
• 5:04
  essentially saying you cannot promise a white
• 5:07
  -collar economic apocalypse and then turn around
• 5:10
  and ask society for a blank check to build the
• 5:12
  machine that causes it. Exactly. And beyond the
• 5:15
  societal implications, there's a very tangible
• 5:17
  structural crisis happening inside the companies
• 5:21
  actually using these tools. This brings us to
• 5:23
  what the tech sector is now calling the token
• 5:26
  trap. The token trap. Right. For anyone integrating
• 5:29
  AI into their daily corporate workflows, you
• 5:32
  understand that an AI model processes information
• 5:34
  in tokens. But we aren't just talking about a
• 5:37
  user typing a prompt and getting a paragraph
• 5:39
  back anymore. We've moved firmly into the era
• 5:42
  of agentic AI workflows. Right. And the architectural
• 5:45
  consequences of those agentic workflows are exactly
• 5:48
  what is causing these budgets to implode. When
• 5:51
  you assign an AI agent to a complex software
• 5:53
  engineering task, it doesn't just read the prompt
• 5:56
  once. No, not at all. The agent writes a piece
• 5:58
  of code, runs it in a sandbox, reads the massive
• 6:00
  error log when it inevitably fails, analyzes
• 6:03
  the failure, writes a new block of code, and
• 6:06
  loops this process hundreds or thousands of times.
• 6:09
  It's having this internal compounding conversation
• 6:11
  with itself. And because you are renting that
• 6:14
  intelligence via an API, the meter is running
• 6:16
  on every single token generated during that massive
• 6:19
  background loop. Yeah, the sheer volume of computational
• 6:22
  cost is staggering, and it's hitting the biggest
• 6:25
  companies on Earth right now. Let's look at the
• 6:28
  numbers. Uber deployed Anthropix Claude code
• 6:31
  to roughly 5 ,000 of their engineers. OK. In
• 6:34
  just four months, they burned through their entire
• 6:37
  AI budget for the year 2026. Wait, really? Their
• 6:40
  whole yearly budget? The whole thing. A $3 .4
• 6:44
  billion budget evaporated in a third of a year.
• 6:48
  entirely due to the compounding token costs of
• 6:51
  agentic workflows. That is insane. And Microsoft
• 6:54
  wasn't immune either. According to the internal
• 6:56
  reports, Microsoft actually had to cancel internal
• 6:59
  API licenses for their own engineers working
• 7:01
  on Windows and Teams because the compounding
• 7:03
  costs were hitting upwards of $2 ,000 a month
• 7:06
  per single engineer. Wow. I want to bring in
• 7:08
  an analogy here because the structural risk of
• 7:10
  this feels incredibly familiar. Think back to
• 7:13
  the 1990s and the massive wave of corporate globalization.
• 7:16
  Okay. Yeah. Companies looked at their domestic
• 7:18
  manufacturing lines and realized they could outsource
• 7:21
  the labor overseas to save cash. On paper, the
• 7:25
  quarterly margins looked fantastic. But over
• 7:27
  a decade, what actually happened? Entire sectors
• 7:30
  lost their deeply ingrained institutional knowledge.
• 7:34
  The engineering expertise, the hands -on know
• 7:36
  -how just vanished. Right, the brain drain. Exactly.
• 7:39
  So wait. If we just ran out our brains to these
• 7:41
  models, aren't we just outsourcing our company's
• 7:43
  entire expertise? We pay the AI to learn how
• 7:46
  to do our jobs, and they keep the knowledge?
• 7:48
  Yes, that is exactly the parallel Nadella draws
• 7:51
  in his essay. When you rely entirely on an external
• 7:54
  API, which is really just the bridge connecting
• 7:56
  your company's servers to OpenAI or Anthropic,
• 7:59
  you're paying a toll for a service, but you aren't
• 8:01
  building any internal equity. Every time your
• 8:04
  company uses that frontier model to solve a complex
• 8:07
  supply chain issue, the model refines its understanding
• 8:10
  of that domain. That generalized intelligence
• 8:13
  lives on their servers, not yours. They can then
• 8:16
  take that newly refined capability and sell it
• 8:19
  right back to your biggest competitor at the
• 8:21
  exact same commodity price. You are basically
• 8:24
  subsidizing the destruction of your own competitive
• 8:26
  moat. So if I'm an IT director or even just an
• 8:29
  everyday professional watching my company bleed
• 8:31
  cash to the token trap while giving away our
• 8:33
  expertise, what's the actual mechanical solution?
• 8:37
  Like how does a company stop renting its brain?
• 8:40
  Well, Nadelli is advocating for a localized architecture
• 8:42
  known as learning loops. Mechanically, this means
• 8:45
  businesses need to stop sending every single
• 8:47
  task to the massive, expensive Frontier models.
• 8:50
  Instead, you deploy smaller, open -weights models
• 8:53
  internally. Models that your company physically
• 8:55
  controls on your own servers. OK, so you keep
• 8:57
  it in -house. Exactly. You use the big Frontier
• 9:00
  model sparingly, maybe just to generate high
• 9:02
  -quality synthetic training data or to tackle
• 9:05
  the absolute hardest edge cases. Then, you use
• 9:09
  those successful outputs to continually fine
• 9:11
  -tune your smaller internal model. Oh, I see.
• 9:14
  It's a distillation process. You're using the
• 9:16
  expensive genius to teach your in -house intern.
• 9:19
  And over time, the intern handles 90 % of the
• 9:23
  daily workload at a fraction of the cost. The
• 9:26
  knowledge gets encoded directly into your private
• 9:28
  system, which protects the expertise of your
• 9:30
  workers while bypassing the token trap completely.
• 9:33
  Precisely. It allows you to swap out the underlying
• 9:35
  Frontier Engine whenever you want. You can use
• 9:37
  OpenAI one month and Google the next because
• 9:39
  your core corporate knowledge isn't locked behind
• 9:41
  their proprietary API. It's a structural defense
• 9:45
  mechanism. Moving on to our next story, this
• 9:47
  actually connects perfectly. If Microsoft is
• 9:49
  urging companies to control their internal architecture,
• 9:52
  the next logical battleground is the raw material
• 9:55
  actually feeding that architecture, which is
• 9:57
  the data. And that brings us to Getty Images.
• 10:00
  Right, the new deal. Yes. We are looking at a
• 10:03
  massive new display agreement between Getty Images
• 10:05
  and OpenAI. Getty's licensed content will now
• 10:08
  natively appear across chat GPT's search and
• 10:11
  discovery features. and the market reaction was
• 10:14
  instantaneous. Getty shares ticker symbol G -E
• 10:17
  -T -Y skyrocketed over 150 % in pre -market trading,
• 10:21
  eventually settling around $1 .35, up 123%. The
• 10:26
  financial pop is undeniable, but the historical
• 10:29
  whiplash is what makes this so compelling. I
• 10:31
  mean, if we rewind the tape to 2022, Getty Images
• 10:34
  was arguably the most aggressive corporate opponent
• 10:36
  of generative AI on the planet. Oh yeah, they
• 10:38
  drew a hard line. They outright banned all AI
• 10:42
  -generated art from their massive library. Then
• 10:44
  they went a step further, launching a highly
• 10:46
  publicized lawsuit against Stability AI for rampant
• 10:49
  copyright infringement. They really positioned
• 10:51
  themselves as the ultimate defenders of human
• 10:53
  creators. It's the classic, you know, if you
• 10:55
  can't beat them, join them trope. Like a movie
• 10:57
  villain teaming up with the hero when the money
• 10:59
  is right. It totally is. The resistance held
• 11:02
  out as long as it could, but the gravitational
• 11:04
  pull of ChatGPT's massive user base was just
• 11:08
  too strong to ignore. But we really need to unpack
• 11:11
  the mechanics of this deal because the wording
• 11:12
  of the press release is incredibly specific.
• 11:15
  It carefully calls this a display agreement.
• 11:18
  This means OpenAI is likely using a mechanism
• 11:20
  called Retrieval Augmented Generation, or RRG.
• 11:25
  When you ask ChatGPT for an image of a specific
• 11:27
  news event, It doesn't hallucinate a new image
• 11:30
  from scratch. It searches Getty's database, retrieves
• 11:33
  the official licensed image, and displays it
• 11:36
  with a proper citation. And that's the core of
• 11:38
  the RGA architecture. It makes the AI's output
• 11:40
  significantly more trustworthy because it's anchoring
• 11:42
  its response in verified real -world data. So
• 11:46
  for the everyday professional, using chat GPT
• 11:49
  to build a presentation, your visual responses
• 11:51
  are about to get significantly richer and critically
• 11:54
  legally safer. Exactly. You won't have to wonder
• 11:56
  if you're going to trigger a copyright lawsuit.
• 11:58
  But Morgan, here is the critical question for
• 12:01
  the listener. This is a display agreement, but
• 12:04
  displaying an image is very different from training
• 12:06
  on an image. We don't know if OpenAI has secured
• 12:09
  the rights to ingest this pristine library of
• 12:13
  data into the actual parameter weights of their
• 12:15
  future models. Does this actually help the 600
• 12:18
  ,000 photographers who took these photos? Or
• 12:21
  is it just padding Getty stock price while ChatGPT
• 12:24
  gets to look more trustworthy? Yeah, that's the
• 12:27
  multi -million dollar pushback. And the silence
• 12:29
  on the training aspect is deafening. To give
• 12:32
  you some context, when Getty signed a similar
• 12:34
  partnership with Perplexity AI last year, they
• 12:37
  explicitly and loudly stated that their images
• 12:39
  could not be used for foundational model training.
• 12:42
  Right, they made a point of it. Exactly. The
• 12:45
  fact that this language is completely missing
• 12:47
  from the OpenAI announcement suggests a major
• 12:49
  concession. However, Getty has also been developing
• 12:52
  its own proprietary generative AI tool, promising
• 12:55
  royalty -free licenses and full legal indemnification
• 12:58
  for enterprise users. They're trying to monetize
• 13:01
  via the AI boom while trying to assure their
• 13:03
  contributors their underlying copyrights still
• 13:05
  have value. It's a tough needle to thread. Let's
• 13:07
  explore this because we have companies fighting
• 13:10
  for control of the architecture and media empires
• 13:13
  fighting for control of the data. But what happens
• 13:15
  when the actual U .S. government steps in to
• 13:18
  control who was even allowed to touch the technology?
• 13:21
  Yeah, this is huge. This brings us to a monumental
• 13:24
  shift in regulatory policy. The White House just
• 13:28
  hit anthropic with a massive export ban Immediately
• 13:31
  pulling its top -tier AI models fable and mythos
• 13:35
  from everyone outside the United States as well
• 13:37
  as restricting access for foreign nationals living
• 13:40
  inside the US We've officially crossed a threshold
• 13:42
  where the government is no longer just talking
• 13:44
  about regulating algorithms They are treating
• 13:46
  these frontier AI models with the exact same
• 13:49
  export controls used for weapons grade munitions
• 13:52
  Wow, and the triggers for this were what? Well,
• 13:55
  reportedly, two main catalysts. First, Anthropic
• 13:59
  granted limited access to a South Korean telecom
• 14:02
  company, SK Telecom. U .S. intelligence officials
• 14:05
  apparently got spooked, suspecting ties to Chinese
• 14:08
  tech sectors, which SK Telecom strongly denies.
• 14:12
  Second, researchers at Amazon allegedly found
• 14:14
  a way to reliably bypass the foundational safety
• 14:17
  guardrails on Anthropic's Fable 5 model. A jailbreak
• 14:20
  that fundamentally compromised the model's safety
• 14:22
  constraints. Yeah. Anthropic disputes how severe
• 14:25
  that jailbreak was, claiming they patched it
• 14:27
  immediately, but the damage in the eyes of the
• 14:29
  regulators was done. The Commerce Department
• 14:31
  issued a directive, and Anthropic reportedly
• 14:33
  had just 90 minutes to scramble, pull the plug
• 14:36
  on their international servers, and restrict
• 14:37
  global access. The funny, or really the most
• 14:40
  tragic, part of this story comes from an analyst
• 14:42
  on X. Oh, I saw this. Yeah, they ran a textual
• 14:45
  breakdown of how these companies communicate.
• 14:47
  They found that Anthropic used words related
• 14:49
  to risk, regulation, or restrictions. five times
• 14:53
  out of every 1 ,000 words. OpenAI used those
• 14:56
  exact same words just 0 .6 times per 1 ,000 words.
• 15:00
  Anthropic actively marketed their Mythos model
• 15:03
  as an incredibly dangerous doomsday cyber machine
• 15:06
  to prove how seriously they took safety. They
• 15:09
  wanted to be seen as the responsible adults in
• 15:11
  the room. But the US government essentially looked
• 15:13
  at their own terrifying marketing materials and
• 15:15
  said, you know what? We agree. This technology
• 15:18
  is far too dangerous to be out in the wild. They
• 15:20
  literally marketed themselves into a federal
• 15:22
  export ban. Which is just wow. But Morgan, if
• 15:26
  mythos really can't be used to supercharge hackers,
• 15:29
  automating the discovery of zero -day vulnerabilities,
• 15:31
  shouldn't the government block it? Or are we
• 15:33
  just going to force all the bad actors to build
• 15:35
  their own unregulated AI overseas? Well, to understand
• 15:38
  why this ban is likely to fail, we have to look
• 15:40
  back at the 1990s and the crypto wars. The US
• 15:44
  tried to ban PGP encryption by treating it as
• 15:47
  a weapon. They placed it on the restricted list
• 15:49
  and opened a federal criminal investigation against
• 15:52
  its creator, Phil Zimmerman. They were treating
• 15:54
  math like physical weapons. Exactly. And Zimmerman
• 15:58
  used a brilliant loophole. He printed the source
• 16:00
  code into a physical book and published it. You
• 16:02
  can ban the digital export of software, but a
• 16:05
  book is protected under the First Amendment.
• 16:07
  Once the book shipped globally, anyone could
• 16:09
  scan the pages and compile the code. The government's
• 16:12
  case completely fell apart. But there's a crucial
• 16:14
  difference between 1990s encryption and 2026
• 16:17
  AI. You could print PGP in a single paperback.
• 16:22
  You cannot print a 100 billion parameter neural
• 16:24
  network in a book. It requires terabytes of storage
• 16:27
  and massive data centers. Doesn't that make an
• 16:30
  AI export ban much more enforceable? It makes
• 16:32
  the deployment harder to hide, yes, but it doesn't
• 16:35
  stop the proliferation. It just changes the geography.
• 16:38
  We saw this in the 2010s with the Wassenaar arrangement
• 16:41
  regarding spyware. When the US and Europe tightened
• 16:44
  the screws, malicious companies didn't stop building
• 16:46
  spyware. They just moved their headquarters to
• 16:48
  countries outside the treaty's jurisdiction.
• 16:51
  If US labs are forced to navigate massive bureaucratic
• 16:54
  hurdles for every foreign customer, the global
• 16:57
  demand doesn't vanish. The capital and talent
• 17:00
  will just flow offshore. which directly shapes
• 17:03
  the tools you, the listener, will have access
• 17:05
  to. If American companies are bogged down in
• 17:07
  compliance, their research slows down, and labs
• 17:10
  in other countries race ahead without any of
• 17:12
  those safety guardrails Anthropic was trying
• 17:14
  to pioneer. Before we wrap up, a quick thanks
• 17:17
  to Old Glory, an iconic music and sports fan
• 17:20
  merch store. Perfect for the techie in your life.
• 17:22
  Use promo code SNARFL for 15 % off at oldglory
• 17:25
  .com. So bringing all these threads together,
• 17:28
  today is really all about the power struggle
• 17:30
  for the future of work. Whether it's Microsoft
• 17:32
  warning companies not to go broke, renting AI
• 17:35
  brains by the token, Getty licensing their visual
• 17:38
  history to Chad GPT, or Anthropic accidentally
• 17:41
  marketing themselves into a government export
• 17:43
  ban. The undeniable theme here is that whoever
• 17:46
  controls the computational architecture, the
• 17:48
  massive data streams it trains on, and the legal
• 17:51
  right to access it, ultimately controls the entire
• 17:54
  future economy. It really leaves us with a final
• 17:57
  thought to mull over. If the government starts
• 17:59
  treating advanced AI like weapons -grade material,
• 18:03
  and companies like Uber are already burning billions
• 18:05
  just to use it, at what point does the cost of
• 18:08
  artificial intelligence outweigh the human intelligence
• 18:10
  we already have? Are we building a brilliant
• 18:13
  tool to elevate our workforce? Or are we constructing
• 18:16
  a massive looping machine that we simply can
• 18:19
  no longer afford to turn on? We're here to help
• 18:21
  it snarfle, so reach out on our website if you
• 18:23
  have any more questions. That's a wrap on today's
• 18:29
  top stories. Remember, tech doesn't have to be
• 18:32
  complicated, and we're here to help connect the
• 18:34
  dots. If you've got questions or want to learn
• 18:37
  more, visit us at snarfle .com. Thanks for listening,
• 18:41
  and we'll catch you next time.