Latest / Connecting the Dots / Claude Fable 5 Unleashed, Safeguarding Frontier AI, and Stealthy Model Restrictions

Transcript

Claude Fable 5 Unleashed, Safeguarding Frontier AI, and Stealthy Model Restrictions · 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 hosts.
• 0:16
  Imagine buying like a world -class power tool,
• 0:20
  right? Like a top -of -the -line table saw. But
• 0:22
  then you find out the manufacturer actually programmed
• 0:25
  it. to quietly ruin your cuts if it suspects
• 0:28
  you're, you know, using it to build a competing
• 0:30
  furniture factory. Oh, wow. Yeah, you literally
• 0:33
  wouldn't know if you just forgot how to measure
• 0:34
  your wood or if the saw was actively working
• 0:36
  against you behind the scenes. That is wild.
• 0:39
  It really is. Anyway, welcome to Connecting the
• 0:41
  Dots. I'm Alex. And I'm Morgan. And we are broadcasting
• 0:44
  live from the SNARFL Solutions Group in sunny
• 0:46
  Sacramento, California. Today is Wednesday, June
• 0:49
  10th, 2026. Yep, 2026. Time is flying. It really
• 0:54
  is. And before we unpack the hidden sabotage
• 0:57
  built into the world's most powerful new artificial
• 0:59
  intelligence, let's just get you caught up on
• 1:01
  the day. So if you're local, definitely crank
• 1:03
  the AC. We are looking at a scorcher here in
• 1:05
  Sacramento. Oh, yeah, it is hot out there. sunny
• 1:08
  and 98 degrees today was just a slight breeze
• 1:10
  at like 5 to 12 miles per hour and Nationally
• 1:14
  the map is just lit up. There are 218 active
• 1:18
  weather alerts across the country with a hundred
• 1:20
  thirty three of those overland So definitely
• 1:23
  check your local listings if you're flying out
• 1:25
  today. Absolutely and Over on Wall Street. The
• 1:28
  morning energy is just a bit fractured I mean
• 1:30
  the Dow is actually up slightly today gaining
• 1:32
  fifty seven point six nine points But it's still
• 1:35
  sitting on a net loss for the week, right? Yeah,
• 1:37
  and the S &P 500 isn't really sharing that morning
• 1:39
  optimism either. It's down 52 .01 points right
• 1:43
  now. And for the crypto watchers out there, Bitcoin
• 1:45
  is hovering right around 61 .4K. It took us light
• 1:49
  dip today, but it's still holding onto a pretty
• 1:50
  solid gain for the week. Well, on the sports
• 1:53
  front, we do actually have a local win to celebrate.
• 1:55
  The Oakland Athletics pulled off a pretty solid
• 1:57
  seven to five victory over the Brewers yesterday.
• 2:00
  Nice. Yeah, JT Jin. taking the win on the mound
• 2:03
  it was a rougher news for the giants though they
• 2:06
  felt the nationals three to six and you know
• 2:08
  if you're king's fan you got a night off to rest
• 2:10
  your voices there was no game on the schedule
• 2:12
  yesterday which is honestly probably for the
• 2:15
  best because the tech news today is It requires
• 2:18
  your full attention. But first, let's pay the
• 2:21
  bills. Today's episode is brought to you by Strong
• 2:23
  Coffee, the premium instant coffee that tastes
• 2:26
  like a traditional brew, but is packed with 15
• 2:28
  grams of protein, NCTs, and adaptogens like ashwagandha.
• 2:32
  Get clean, sustained energy without the jitters.
• 2:35
  Use promo code SNARFL for 20 % off at strongcoffeecompany
• 2:39
  .com. So let's look closer at what is happening
• 2:42
  in the AI space today because Anthropic has officially
• 2:46
  dropped their fifth generation models and we
• 2:48
  really need to pause and kind of reestablish
• 2:50
  what capable even means now because this is a
• 2:52
  monumental leap. Right, it's totally different
• 2:54
  now. Exactly. So they're releasing this two -tiered
• 2:57
  system. The public and everyday enterprise users,
• 3:00
  they get access to CloudFable 5. But behind closed
• 3:03
  doors for highly trusted partners like Project
• 3:05
  Glasswing and the U .S. government, there is
• 3:07
  this completely unrestricted powerhouse called
• 3:09
  Quad Mythos 5. And, okay, the real world application
• 3:13
  of Fable 5 is what's blowing my mind this morning.
• 3:17
  Think about the payment processor, Stripe. They
• 3:19
  just used Fable 5 to migrate a 50 million line
• 3:23
  Ruby code base. Which is just an unfathomable
• 3:27
  amount of code. It is. For anyone who hasn't
• 3:30
  worked in software development, a code base of
• 3:32
  that size is essentially the entire architectural
• 3:36
  foundation of a massive digital city. I mean,
• 3:38
  it handles billing logic, user data, security
• 3:42
  protocols. It's literally millions of text files
• 3:44
  interlocking with one another. Right. And moving
• 3:46
  that from an old system to a new one isn't just
• 3:49
  a copy paste job. No, not at all. You have to
• 3:51
  translate the grammar of the code while the website
• 3:53
  is live and processing credit cards. And historically,
• 3:56
  you know, mapping those dependencies takes a
• 3:58
  dedicated team of senior engineers a minimum
• 4:00
  of two months. Yeah. The Fable 5 did it in a
• 4:03
  single day. A single day. Yeah. And the reason
• 4:05
  it can do that is because of its expanded memory
• 4:07
  or what they call the context window. It doesn't
• 4:11
  just look at one line of code at a time. It can
• 4:13
  hold the entire architecture of that 50 million
• 4:16
  line digital city in its working memory simultaneously.
• 4:19
  It can just spot the broken connections and rewrite
• 4:21
  them on the fly. And it's not just text -based
• 4:24
  logic anymore either. I mean, the vision capabilities
• 4:27
  have crossed a completely new threshold. The
• 4:29
  Pokemon thing. Yes. Previously, if you showed
• 4:32
  an AI an image, it was basically just labeling
• 4:35
  objects. It needed massive amounts of scaffolding
• 4:38
  to understand context. But Fable 5 just beat
• 4:42
  the video game Pokemon Fire Red using nothing
• 4:44
  but screenshots. Yeah, that is a really crucial
• 4:47
  distinction to understand. It didn't plug into
• 4:49
  the underlying code of the game, right? Right.
• 4:51
  It didn't read the game's memory to know where
• 4:53
  the enemies were. It just looked at a grid of
• 4:55
  RGB pixel values on a screen right recognized
• 4:59
  like oh that cluster green pixels is a tree that
• 5:01
  cluster red is my character calculated the optimal
• 5:04
  spatial path and then outputted a logical keystroke
• 5:06
  commit which is wild it translates a two -dimensional
• 5:09
  image matrix into actual spatial and strategic
• 5:13
  reasoning which makes the restricted model mythos
• 5:15
  5 even wilder right because anthropics internal
• 5:18
  experts are using that unrestricted model to
• 5:22
  accelerate drug design by a factor of 10 wow
• 5:25
  Ten times. Yeah! They are plugging it into bioinformatics
• 5:28
  tools. And bioinformatics is essentially computational
• 5:32
  biology, using software to simulate how proteins
• 5:35
  fold in three -dimensional space to block a virus.
• 5:39
  And Methos 5 isn't just running the software,
• 5:41
  it's acting as the lead scientist. So it's making
• 5:43
  the actual choices. Exactly. It chooses the binding
• 5:46
  sites, it runs the simulation, analyzes the failures,
• 5:49
  and then adjusts its own hypothesis. Like, out
• 5:51
  of 14 protein targets, it independently generated
• 5:54
  strong candidates for nine of them. That's incredible
• 5:57
  and you know to understand why this massive technological
• 5:59
  leap is happening exactly right now, you really
• 6:02
  have to look at the financial architecture of
• 6:04
  Silicon Valley. Anthropic and their main competitor
• 6:06
  OpenAI have both quietly filed S1 paperwork with
• 6:10
  the SEC. Meaning the race to a multi -billion
• 6:13
  dollar initial public offering is officially
• 6:15
  on. Precisely. And S1 is essentially the document
• 6:18
  you present to Wall Street to prove your company
• 6:20
  is worth investing in. You have to tell a compelling
• 6:23
  narrative. Right. And Anthropic Currently has
• 6:26
  a leaner computing cost structure than its rivals
• 6:29
  So by dropping fable five and mythos five today
• 6:33
  that I mean unequivocally claimed the top spot
• 6:36
  on nearly every performance benchmark They are
• 6:39
  aggressively positioning themselves as the undisputed
• 6:42
  kings of automated labor right before they ask
• 6:45
  the public markets for billions of dollars. You
• 6:47
  know, I'm looking at this two tiered release
• 6:49
  and it feels like Anthrobic has engineered the
• 6:52
  most powerful combustion engine in human history.
• 6:54
  OK, I like that. But instead of putting it in
• 6:56
  one chassis, they built two. So for the public,
• 6:58
  they are selling a street legal sports car. That's
• 7:02
  Fable 5. It's incredibly fast, but there's a
• 7:04
  heavy governor bolted onto the engine, so you
• 7:05
  can't just, you know, go 300 miles per hour on
• 7:08
  the interstate. Right. But then for professional
• 7:10
  drivers on a closed track, they offer the Formula
• 7:12
  One version, Mythos 5. And it's the exact same
• 7:15
  engine, just with the restrictions ripped off.
• 7:18
  Yeah, the underlying neural network is identical.
• 7:20
  The only difference is the software guardrails
• 7:22
  dictating how the engine is allowed to express
• 7:24
  its power. But wait, if the engine is the exact
• 7:26
  same, isn't releasing the... Street legal version
• 7:30
  to millions of people incredibly reckless like
• 7:33
  won't an adversary or some competing lab eventually
• 7:35
  figure out how to bypass the governor Access
• 7:38
  the raw weights of the model and just unleash
• 7:40
  that f1 speed on the open road Yeah, that is
• 7:43
  the exact security nightmare keeping AI researchers
• 7:46
  awake at night And it is the entire reason anthropic
• 7:50
  has built an incredibly aggressive containment
• 7:52
  system into fable five I mean they know this
• 7:55
  engine is capable of walking a user through the
• 7:57
  synthesis of a bio weapon or Executing an autonomous
• 8:02
  cyber attack. Yeah. So how do you physically
• 8:04
  prevent a supercomputer from handing over those
• 8:06
  blueprints? You don't just train it to politely
• 8:08
  refuse you build a mechanical trap door the opus
• 8:11
  fallback, right? Fable five utilizes a completely
• 8:14
  separate layer of software called a safety classifier
• 8:17
  Before Fable 5 even begins to process your prompt,
• 8:20
  this tiny, ultra -fast classifier reads your
• 8:23
  request and scores it for biological, chemical,
• 8:26
  or cybersecurity risk. And if your prompt triggers
• 8:29
  a high -risk score, the system physically disconnects
• 8:32
  your input from Fable 5. It intercepts the data
• 8:35
  and reroutes it to an older, heavily neutered
• 8:38
  model called Claude Opus 4 .8. I love the mechanics
• 8:42
  of this so much. It's like hiring a world -class
• 8:44
  master hacker to audit your home security system.
• 8:47
  They're brilliant, analyzing all your network
• 8:49
  vulnerabilities, but the very second you ask
• 8:51
  them, hey, how would I use these exact same techniques
• 8:53
  to break into the bank down the street, the system
• 8:55
  puts you on hold and patches you through to a
• 8:57
  mall security guard. Exactly. I mean, Opus 4
• 9:00
  .8 isn't useless, but in this context, it's the
• 9:02
  mall cop. It can tell you what time the bank
• 9:05
  opens, but it fundamentally lacks the cognitive
• 9:07
  architecture to draw up the blueprints to the
• 9:10
  vault. And the stats back up how robust this
• 9:13
  trapdoor is, too. External security testers spent
• 9:16
  over a thousand hours trying to jailbreak this
• 9:18
  system. And jailbreaking is when you use linguistic
• 9:23
  tricks, like telling the AI to pretend it's a
• 9:26
  character in a sci -fi movie to bypass the safety
• 9:28
  filters. They fail to find a single universal
• 9:31
  exploit. Fable 5 currently has a 0 % success
• 9:35
  rate in assisting with offensive cyber tasks
• 9:37
  because the classifier cuts the cord before the
• 9:41
  model can even think about the malicious code.
• 9:43
  Anthropic is also implementing a strict 30 -day
• 9:46
  data retention policy on all prompts sent to
• 9:49
  Mythos class models, which includes Fable 5.
• 9:52
  They are actively monitoring the trafficked spot
• 9:54
  novel multi -prompt attacks. If someone discovers
• 9:57
  a new linguistic trick to jimmy the lock on that
• 10:00
  trapdoor, Anthropic wants to analyze the attempt
• 10:03
  and patch the vulnerability in real time. But
• 10:05
  this containment strategy comes with a massive
• 10:07
  cost to the user experience. Anthropic openly
• 10:10
  admits these safety classifiers are tuned conservatively.
• 10:12
  meaning they trigger false positives in less
• 10:14
  than 5 % of sessions. So if you're a software
• 10:17
  engineer listening to this on your commute right
• 10:19
  now, think about your current workflow. 5 % sounds
• 10:23
  like a rounding error, right? Sure. But if you
• 10:25
  are running a multi -day autonomous coding session
• 10:28
  where the AI is compiling and testing its own
• 10:31
  work over thousands of iterations, a single false
• 10:34
  positive derails the entire chain. Yeah, that
• 10:37
  would be incredible frustrating. You ask an innocent
• 10:39
  question about network security architecture,
• 10:41
  the trapdoor opens, and suddenly your genius
• 10:43
  F1 engine is replaced by the mall cop, completely
• 10:47
  ruining the project. Well, Anthropic is actively
• 10:50
  choosing a degraded user experience in exchange
• 10:52
  for a reduction in global risk. They argue that
• 10:56
  protecting the power grid from an automated cyber
• 10:58
  attack is worth inconveniencing a developer.
• 11:02
  Yeah, that makes sense. And while developers
• 11:03
  are annoyed by the false positives, at least
• 11:05
  that containment system is transparent. You get
• 11:08
  a notification. You know you've been bumped to
• 11:09
  Opus 4 .8. Which brings us to the core of today's
• 11:13
  tech outrage. Let's shift from the transparent
• 11:16
  guardrails designed to stop hackers to the invisible
• 11:19
  guardrails designed to stop competitors. Oh yeah,
• 11:22
  this is where we cross the line from global security
• 11:24
  into corporate strategy. Anthropic has embedded
• 11:27
  a second layer of safeguards into Fable 5, but
• 11:30
  these aren't looking for bioweapons. They are
• 11:32
  looking for distillation. Okay, let's define
• 11:35
  distillation because it is the entire business
• 11:37
  model for thousands of tech startups right now.
• 11:40
  Definitely. Distillation is when you take a massive
• 11:43
  like a hundred million dollar artificial intelligence
• 11:45
  and you use it to tutor your own smaller cheaper
• 11:49
  AI. You ask the genius AI thousands of complex
• 11:53
  questions and you feed its perfect answers into
• 11:56
  your tiny model until your tiny model learns
• 11:58
  how to mimic the genius. Right. And Anthropic
• 12:00
  obviously does not want you using their billions
• 12:02
  of dollars of research to build a competing product,
• 12:05
  what they call a frontier LLM, obviously. So
• 12:08
  if their system suspects your prompts are being
• 12:10
  used for distillation, they do not reroute you
• 12:13
  to Opus 4 .8. They do not send you a notification
• 12:15
  or ban your account. Instead, they use hidden
• 12:18
  techniques like prompt modification, steering
• 12:20
  vectors, and parameter -efficient fine -tuning,
• 12:23
  or PEVFT. So, mechanically, what does that actually
• 12:26
  look like for the user? Think of PAT as a hyper
• 12:29
  -specific, temporary lobotomy. Instead of retraining
• 12:33
  the billions of mathematical parameters that
• 12:35
  make up the AI's brain, which is impossible to
• 12:38
  do on the fly, and Thropic injects a tiny lightweight
• 12:42
  adapter layer over the model's outputs. Combine
• 12:45
  that with steering vectors, which literally add
• 12:47
  mathematical weights that push the AI's thoughts
• 12:50
  away from high -level elegant code and force
• 12:53
  it toward generic unhelpful boilerplate. So they
• 12:56
  are silently sabotaging their own model's performance
• 12:59
  in real time. It actively degrades its own intelligence.
• 13:03
  Correct. The AI intentionally lowers the mathematical
• 13:06
  probability of outputting the correct syntax.
• 13:08
  It gives you slightly flawed advice, or it entirely
• 13:12
  misses an obvious solution and you're never notified
• 13:14
  that this restriction has been activated. The
• 13:16
  fallout in the developer community today is just
• 13:18
  massive. There is a viral post making the rounds
• 13:21
  by a developer named Jonathan Reddy, and he exposes
• 13:24
  this invisible nerfing as a catastrophic supply
• 13:26
  chain risk for modern businesses. Oh, I saw that.
• 13:29
  Yeah. Reddy points out that the old definition
• 13:31
  of frontier AI research is basically dead. The
• 13:35
  line between building a rival supercomputer and
• 13:37
  just building a normal startup product has completely
• 13:40
  vanished. Yeah, because five years ago, if you
• 13:42
  started a company, you wrote some basic code,
• 13:45
  spun up a database, and you were good to go.
• 13:47
  Today, standard software development involves
• 13:50
  training your own AI tools. Startups routinely
• 13:53
  build embedding models, which are basically algorithms
• 13:56
  that translate text into numbers so an AI can
• 13:59
  understand the semantic similarity between two
• 14:01
  concepts. Right. And they build re -rankers.
• 14:04
  which are secondary models that sort those search
• 14:06
  results. This isn't supervillain research, this
• 14:08
  is just how you build a functional search bar
• 14:10
  in 2026. And Reddy has a small bootstrapped travel
• 14:13
  application called Wanderfugal. He is constantly
• 14:17
  training his own Imjong algorithms to make his
• 14:19
  app's recommendation engine better. So if Jonathan
• 14:21
  asks Fable 5 for help debugging his machine learning
• 14:24
  pipeline, and Fable 5 suddenly spits out terrible
• 14:27
  broken code, he has absolutely no way of knowing
• 14:30
  why. Right. Did he write a bad prompt? Did he
• 14:34
  misunderstand the math or did anthropics invisible
• 14:37
  safeguards secretly activate? label his tiny
• 14:41
  travel app as a competing frontier model, and
• 14:44
  proactively ruin his project. Exactly. If you
• 14:47
  are building digital infrastructure, your tools
• 14:50
  must be deterministic. If a tool fails, it has
• 14:53
  to fail transparently so you can fix the error.
• 14:56
  If a developer platform is explicitly allowed
• 14:58
  to stop optimizing for your success without telling
• 15:01
  you, the trust in that entire ecosystem just
• 15:04
  evaporates. It goes right back to the table saw
• 15:06
  analogy. Imagine the manufacturer programs the
• 15:09
  saw so that if it suspects you are cutting wood
• 15:11
  for a competing tool factory, it quietly reduces
• 15:14
  its own RPMs. The blade slows down just enough
• 15:17
  to splinter the wood. You're standing there wondering
• 15:19
  if you bought bad lumber, completely unaware
• 15:21
  the tool itself is actively working against you.
• 15:23
  And this really forces us to look back at that
• 15:25
  upcoming IPO. What is the actual motivation for
• 15:28
  this silent sabotage? And Tropic claims they
• 15:30
  are doing this to prevent authoritarian nation
• 15:32
  states from using Fable 5 to spin up rival, unaligned
• 15:35
  intelligence systems in a matter of weeks. Which
• 15:38
  is fair. And frankly, yeah, that is a legitimate
• 15:40
  national security threat. And the Tropic has
• 15:43
  publicly stated they face massive daily extraction
• 15:46
  attempts from foreign adversaries. But the timing
• 15:49
  is just impossible to ignore. Is this actually
• 15:51
  about global security, or are they quietly suffocating
• 15:55
  everyday developers to protect their own market
• 15:57
  share and technical moat right before they ask
• 16:00
  Wall Street for a multi -billion dollar valuation?
• 16:02
  That is the big question. I mean, if it's really
• 16:05
  about security, why gaslight the developer? If
• 16:08
  they break the terms of service, ban the account,
• 16:11
  or route them to the mall cop. Making the AI
• 16:14
  act unintelligently in secret is a wildly risky
• 16:17
  choice for a company trying to convince Fortune
• 16:19
  500 companies to rely on them. It highlights
• 16:22
  the immense, almost completely unchecked power
• 16:24
  these three or four major AI labs hold right
• 16:27
  now. They aren't just selling software. They
• 16:29
  are unilaterally deciding the physics of the
• 16:31
  new digital economy. It's an incredible tension
• 16:34
  to navigate, and we're going to tie all these
• 16:35
  threads together in just a moment. But first...
• 16:37
  Before we wrap up, a quick thanks to Old Glory,
• 16:40
  an iconic music and sports fan merch store. Perfect
• 16:43
  for the techie in your life. Use promo code SNARFL
• 16:47
  for 15 % off at oldglory .com. So if we look
• 16:51
  at the big picture today, we have really crossed
• 16:53
  a threshold. We now have access to systems that
• 16:56
  can digest a 50 million line code base, understand
• 17:00
  the three -dimensional physics of a video game
• 17:02
  from a screenshot, and act as autonomous biological
• 17:05
  scientists. It's staggering. They are compressing
• 17:08
  months of human labor into hours. But that immense
• 17:11
  world -altering capability comes bolted to some
• 17:14
  incredibly heavy training wheels. And as we've
• 17:16
  seen, some of those training wheels are highly
• 17:18
  visible. The Opus 4 .8 fallback and the safety
• 17:20
  classifiers actively, transparently stop bad
• 17:23
  actors from designing bioweapons or exploiting
• 17:25
  zero -day cyber vulnerabilities. But the invisible
• 17:28
  training wheels, the steering vectors, and PEFT
• 17:31
  models that quietly degrade the AI's logic to
• 17:34
  protect a corporate business model, they introduce
• 17:36
  a massive layer of paranoia into modern engineering.
• 17:38
  Which leaves us with a lingering question to
• 17:41
  mull over as you head into the rest of your day.
• 17:43
  If tech companies can now quietly and invisibly
• 17:46
  degrade the performance of their own products
• 17:49
  to protect their business modes under the umbrella
• 17:51
  of safety, well... What happens when these exact
• 17:54
  same black box AI models are fully integrated
• 17:58
  into our critical infrastructure? Yeah, that's
• 18:00
  a scary thought. When Fable 6 or 7 is actively
• 18:03
  running our municipal power grids, or triaging
• 18:06
  patients in a hospital system, who is monitoring
• 18:09
  the monitors when the sabotage is completely
• 18:11
  invisible. We're here to help at SNARFL, so reach
• 18:13
  out on our website if you have any more questions.
• 18:18
  That's a wrap on today's top stories. Remember,
• 18:21
  tech doesn't have to be complicated, and we're
• 18:23
  here to help connect the dots. If you've got
• 18:26
  questions or want to learn more, visit us at
• 18:28
  snarfle .com. Thanks for listening, and we'll
• 18:32
  catch you next time.