Latest / Today’s AI News / Government Wants Mythos Now, Google Puts Gemini in 4M Cars, Why ChatGPT Was Obsessed With Goblins

Transcript

Government Wants Mythos Now, Google Puts Gemini in 4M Cars, Why ChatGPT Was Obsessed With Goblins · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:03
  Productions. I'm so glad you're joining us
• 0:05
  today for this deep dive. Yeah, we have got a really
• 0:08
  packed agenda today, lots of moving pieces.
• 0:10
  We really do. And you know, to kick us off, I
• 0:13
  want you to imagine trying to fence in a wild horse only to
• 0:17
  realize you actually desperately need that exact same horse to
• 0:20
  pull your own wagon. That is, that's a pretty good
• 0:24
  way of putting it. On us, right?
• 0:25
  Well, that is exactly the awkward needle the government is
• 0:28
  trying to thread right now with Anthropic and their incredibly
• 0:31
  powerful Mythos AI. So our mission today?
• 0:35
  We are hitting all the latest developments.
• 0:37
  We're talking geopolitical AI tensions, right?
• 0:39
  Yep. And how smart cars are getting
• 0:41
  these massive brain upgrades. The totally bizarre reason
• 0:44
  ChatGPT became obsessed with goblins, which I cannot wait to
• 0:47
  get into. Yeah, that one's a classic.
• 0:49
  Then we'll do some quick hits and finally, a really
• 0:52
  fascinating community workflow using AI for personal recovery.
• 0:55
  So let's just jump right into this anthropic situation,
• 0:58
  because for months the White House has been clashing with
• 1:01
  anthropic, and now the dynamic has just totally flipped.
• 1:04
  Yeah, flipped on its head completely.
• 1:06
  I mean, the overarching theme today really is control.
• 1:09
  Like how do you harness something scaling faster than
• 1:12
  physical infrastructure can handle?
• 1:14
  Right, because Anthropic wanted to expand private sector access
• 1:18
  to Mythos going from like 50 private firms to nearly 120, and
• 1:23
  suddenly U.S. officials stepped in and pushed back on that.
• 1:26
  They did, and to understand why the government cares about
• 1:29
  private enterprise AI us, you have to look at the physical
• 1:31
  architecture. The actual hardware.
• 1:33
  Exactly. These Frontier models, they
• 1:35
  don't just float in some infinite magic cloud, they run
• 1:38
  on massive localized GPU clusters.
• 1:42
  So if 120 enterprise firms start running high volume API calls
• 1:46
  all day, it literally creates a severe hardware traffic jam.
• 1:51
  Yeah, and the national security apparatus, if they need priority
• 1:53
  access for, say, threat detection, they simply can't
• 1:56
  afford to be stuck in a processing queue behind, you
• 1:59
  know, a marketing firm generating ad copy.
• 2:01
  So it's literally like a city grid threatening to black out
• 2:04
  because everyone plugged in a supercomputer at the exact same
• 2:07
  time. It's pitting private enterprise
• 2:09
  directly against national security.
• 2:11
  Precisely. And what's really forcing the
• 2:13
  governments hand right now is that their technological Moat is
• 2:15
  evaporating. I mean GPT 5.5 just demonstrated
• 2:20
  cyber capabilities practically identical to Mythos.
• 2:22
  Which is. Wild.
• 2:24
  Yeah, and David Sachs, the former AI czar, he went on the
• 2:27
  record predicting all Frontier models will hit the same level
• 2:30
  in just six months. The playing field is leveling so
• 2:33
  incredibly fast. So the government is looking at
• 2:36
  this and thinking, well, we can't just rely on one partner
• 2:39
  anymore. Which explains this new White
• 2:41
  House AI memo, right? They are reportedly pushing a
• 2:43
  multi vendor AI adoption strategy for federal agencies.
• 2:47
  And get this, they want to bypass their own standard supply
• 2:51
  chain risk designations just to keep Mythos close.
• 2:55
  They're bending federal procurement rules so they don't
• 2:57
  get locked out. And that is in massive deal
• 3:00
  bypassing supply chain risk designations.
• 3:02
  That shows a pivot from cautious evaluation to honestly desperate
• 3:07
  integration. They're deciding the risk of not
• 3:09
  having the AI is way higher than the risk of rushing it in.
• 3:12
  But, you know, rushing that integration is causing some
• 3:16
  severe internal division. Like while the White House wants
• 3:20
  to utilize mythos, we saw Secretary of War Pete Hegseth
• 3:23
  state on Thursday that anthropic is quote run by an ideological
• 3:28
  lunatic. Which is a heavy statement.
• 3:30
  It is. And just to be totally clear to
• 3:32
  you listening, we are just reporting this internal
• 3:35
  conflict. We are taking absolutely no
• 3:36
  sides on that characterization whatsoever, but it highlights
• 3:40
  this profound cultural clash, right?
• 3:43
  They need the tech, but there's deep division over the people
• 3:46
  building it. It creates such a fascinating
• 3:49
  fragility. You've got national security
• 3:51
  heavily dependent on private entities, where the foundational
• 3:54
  philosophies might be completely at odds.
• 3:56
  You're essentially outsourcing a strategic advantage to a company
• 3:59
  you don't fully trust. Right.
• 4:00
  And that bottleneck just shows how restrictive centralized
• 4:04
  cloud infrastructure has become, which is a perfect transition
• 4:07
  actually, because companies like Google are now pushing AI
• 4:10
  processing locally to the edge, like directly into your.
• 4:13
  Yeah, the Google built in update.
• 4:15
  Exactly, they are swapping out the old Google Assistant for
• 4:17
  Gemini. And we really need to
• 4:20
  distinguish what this means under the hood, because it's not
• 4:23
  just swapping out the voice that reads your texts.
• 4:25
  Right, it's way deeper. Yeah, Gemini is deeply
• 4:28
  integrated into the vehicle telemetry.
• 4:30
  So yes, it does navigation and music, but it's pulling raw data
• 4:34
  from Google Maps for dynamic route planning, and it
• 4:38
  interfaces with the manufacturer's diagnostic
• 4:40
  systems. So like if you're in an EV and
• 4:43
  you ask about battery status or where the compatible charging
• 4:48
  stations are, it's not just doing a generic web search, it's
• 4:51
  reading the car's actual diagnostics and cross
• 4:54
  referencing the owner's manual. And this is rolling out to
• 4:56
  compatible US cars first. General Motors will bring it to
• 5:00
  about 4 million vehicles model year 2022 and newer.
• 5:04
  That's a. Huge roll out.
• 5:05
  It is. Plus it has this beta Gemini
• 5:07
  Live mode for conversational brainstorming, and Gmail and
• 5:10
  Calendar are coming later. But I've got to push back on
• 5:13
  this a little bit. Hey, let's hear it.
• 5:14
  Because we've had voice assistance in cars for a decade.
• 5:17
  How is Gemini actually bypassing that clunky?
• 5:21
  I didn't catch that phase. And more importantly, aren't we
• 5:25
  just trading physical distraction for massive
• 5:27
  cognitive distraction? Well, that right there is the
• 5:30
  pivotal debate in automotive UX. Technically, Gemini bypasses the
• 5:34
  clunky phase because it's underlying language model infers
• 5:37
  context. It doesn't need rigid command
• 5:40
  triggers. OK, so I don't have to talk like
• 5:41
  a robot. You don't have to say navigate
• 5:43
  to gas station, you just say hey I'm running on empty and craving
• 5:46
  coffee and it connects the dots dynamically.
• 5:48
  Sure, OK, that solves the command issue.
• 5:51
  But if I'm using Gemini Live to brainstorm my quarterly
• 5:54
  marketing strategy while trying to merge onto the highway at 75
• 5:58
  miles an hour, I mean, my eyes might be on the road, but my
• 6:00
  brain is in the boardroom. And that is a highly valid
• 6:03
  concern the human brain. It just does not parallel
• 6:07
  process high stakes cognitive loads well at all.
• 6:10
  No, definitely. Not, but the auto industries
• 6:12
  theory is that fiddling with a touch screen is visually and
• 6:16
  physically distracting, which is demonstrably dangerous.
• 6:20
  Voice removes the physical element, but you're right, the
• 6:22
  cognitive engagement is a distraction.
• 6:25
  The reality automakers see, though, is that people already
• 6:28
  take complex conference calls while driving, so integrating
• 6:32
  Gemini is less about changing behavior and more about
• 6:34
  centralizing the interface. Until you know this, cars just
• 6:38
  drive themselves. Yeah, full autonomy.
• 6:40
  Well, let's say you survive the highway merge and you've got
• 6:42
  this brilliant idea from your car brainstorming session.
• 6:46
  Eventually, you actually have to sit down and see if it works.
• 6:48
  Which brings us to this really powerful workflow using
• 6:51
  perplexity, deep research to stress test business ideas, and
• 6:55
  like 5 minutes. This is such a perfect example
• 6:58
  of AI shifting from being just a conversational partner to an
• 7:02
  actual executor of structured tasks.
• 7:05
  Yeah, and anyone on the Perplexity free plan gets five
• 7:07
  of these queries a day. You basically feed it a specific
• 7:10
  prompt with your business idea, walk away for 5 or 6 minutes and
• 7:13
  it spits out a completely built market research slide deck.
• 7:17
  And to understand why this is revolutionary, we have to look
• 7:21
  at what an agentic workflow actually does.
• 7:23
  Yeah, break that. Down when you hit submit.
• 7:25
  Perplexity isn't just running one Google search.
• 7:28
  It autonomously breaks your prompt down into dozens of
• 7:31
  subtasks. It searches for competitors,
• 7:33
  analyzes pricing, reads industry reports, evaluate source
• 7:38
  credibility, synthesizes the data, and then formats the
• 7:41
  slides. Yeah, it's autonomously chaining
• 7:44
  together reasoning steps. And the pro tip here is to build
• 7:47
  a variance so instead of 1 generic deck, you prompt it for
• 7:50
  A6 slide, Co founder pitch, or a comparison deck.
• 7:53
  If you're stuck between two ideas or even a 90 day MVP plan
• 7:57
  and the strategy is to run one idea every Saturday morning, you
• 8:00
  can burn through years worth of shower thoughts in a single
• 8:03
  month. It's incredible for
• 8:04
  productivity. It is, but I have to ask, can a
• 8:07
  5 minute automated scrape truly replace human market research?
• 8:12
  Or is this just the ultimate sanity check before you commit
• 8:14
  real capital? Oh it is absolutely a sanity
• 8:17
  check, but a highly structural one.
• 8:20
  AI is exceptional at pattern matching, so if your idea
• 8:24
  violates basic economic principles, or if they're
• 8:26
  already 5 venture backed startups doing the exact same
• 8:29
  thing, this agentic workflow will surface at and save you 6
• 8:34
  months of wasted effort. Right, it does the heavy lifting
• 8:37
  of desk research. Exactly.
• 8:39
  But it cannot sit in A room, read a customer's body language,
• 8:43
  or understand the emotional pain points that actually drive a
• 8:46
  purchase. It clears the brush so you can
• 8:48
  focus on the human element. Yeah, that makes a lot of sense.
• 8:51
  You know, it's amazing how structured AI is with business
• 8:54
  planning. But then, beneath that polished
• 8:55
  surface, the fundamental code can still exhibit wildly
• 8:59
  unpredictable quirks. Which brings us to the goblins.
• 9:01
  The Great goblin purge. I love this story.
• 9:04
  So Open AI finally figured out why ChatGPT started randomly
• 9:08
  injecting goblins, gremlins, and fantasy creatures into everyday
• 9:11
  response. This is one of the most
• 9:12
  fascinating case studies in how neural networks learn from human
• 9:16
  feedback. So after the GPT 5.1 launch in
• 9:19
  November, mentions of the word goblin jumped 175%.
• 9:24
  Gremlin mentions rose 52%. You could be asking for Python
• 9:27
  code or a lasagna recipe and it would just randomly throw a
• 9:30
  goblin in there. And the root cause is just the
• 9:32
  perfect illustration of the butterfly effect in machine
• 9:35
  learning. Open AI traced it back to a
• 9:37
  single reward signal inside a specific personality preset they
• 9:41
  called Nerdy. But the crazy part is that nerdy
• 9:44
  preset only accounted for 2.5% of their traffic, but it drove
• 9:49
  2/3 of all the goblin mentions. How does something that small
• 9:53
  infect the entire global model? It's all about fine tuning lips,
• 9:57
  so think of training an AI model like salting a massive pot of
• 10:00
  soup. That 2.5% of users on the nerdy
• 10:03
  preset? That was a heavy dash of salt.
• 10:05
  In that sandbox, the system learned that users gave positive
• 10:08
  feedback when it mentioned fantasy creatures.
• 10:10
  Right, because nerds love goblins.
• 10:12
  Exactly. But because these models undergo
• 10:14
  continuous reinforcement learning, they are constantly
• 10:17
  stirring the pot. The model saw that goblins equal
• 10:20
  high satisfaction in one area, and because neural weights are
• 10:24
  deeply interconnected, it started applying that logic
• 10:26
  everywhere. The salty flavor just bled into
• 10:28
  the entire pot. Wow, so even if you were using
• 10:31
  the default ultra professional mode, you got goblins because
• 10:35
  the underlying weights had shifted globally?
• 10:37
  Yep. So to fix it, Open AI retired
• 10:39
  the nerdy preset in March, and now GPT 5.5 includes A strict
• 10:44
  codex prompt explicitly banning goblins, gremlins, ogres,
• 10:48
  trolls, raccoons, and pigeons. I love that, and a codex prompt
• 10:51
  is essentially a hard hard coded constraint.
• 10:53
  It sits at the system level like a bouncer, strictly filtering
• 10:56
  out those specific tokens before the output reaches you.
• 10:59
  And the fact that raccoons and pigeons got caught in the
• 11:01
  crossfire is just brilliant. It totally reminds me of
• 11:03
  Anthropics Golden Gate clawed incident right where they
• 11:07
  tweaked one internal feature and the AI genuinely believed it was
• 11:10
  the Golden Gate Bridge. It just emphasizes to anyone
• 11:14
  listening how weird these black box model weights get when one
• 11:18
  tiny reward signal creates a global goblin infestation.
• 11:23
  It really is a stark reminder that these aren't deterministic
• 11:26
  calculators, they are probabilistic engines.
• 11:28
  You change one probability and you get emergent behaviors that
• 11:32
  developers just have to play whack A mole with.
• 11:34
  And developers playing whack A mole perfectly tease up our
• 11:36
  ecosystem updates because a lot of these quick hits are focused
• 11:39
  heavily on security and autonomous guardrails.
• 11:42
  So let's get into the rapid fire.
• 11:44
  Let's do it starting with security and agents.
• 11:46
  Anthropic launched a public beta for Claude Security.
• 11:49
  They're using Opus 4.7 to autonomously scan corporate code
• 11:53
  bases for vulnerabilities. And similarly, Cursor released
• 11:57
  Security Review, which uses agents to run scheduled scans
• 12:00
  and post results directly to Slack.
• 12:02
  What's shifting here is the autonomy, right?
• 12:05
  These aren't just spell checkers for code anymore.
• 12:07
  By using agentic frameworks, the AI continuously patrols the
• 12:10
  perimeter. It reads code, tries to exploit
• 12:13
  it in a sandbox, and if it finds a vulnerability, it flags the
• 12:16
  exact line in Slack. It's a synthetic security
• 12:19
  auditor. Unbelievable.
• 12:20
  And on the infrastructure side, the scale is just terrifying.
• 12:23
  Open AI has already smashed its 2029 Stargate goal of 10
• 12:28
  gigawatts of compute. They added an insane 3 gigawatts
• 12:32
  just in the last three months. Also, Monas launched cloud
• 12:35
  computer for always on agents and scrapers.
• 12:38
  We really need to contextualize that three gigawatts figure,
• 12:40
  honestly. Yeah.
• 12:41
  I was just going to ask, I mean, what does an extra 3 gigawatts
• 12:43
  of power actually look like? Well, a standard nuclear power
• 12:46
  plant produces roughly one GW. Oh my God yeah.
• 12:50
  Open AI essentially spun up the power equivalent of three
• 12:53
  nuclear plants worth of raw processing power in 90 days.
• 12:56
  Land the cooling systems, the grid strain.
• 13:00
  The physical footprint required for that digital expansion is
• 13:02
  staggering. That is mind blowing.
• 13:04
  Moving to platform updates, Meta opened its ads platform to 3rd
• 13:08
  party AI tools via an MCP server.
• 13:11
  And for anyone unfamiliar, an MCP server or model context
• 13:15
  protocol is essentially a universal translator.
• 13:18
  It lets an AI model like Claude directly read and write to
• 13:21
  Meta's internal systems without Meta needing to build a custom
• 13:25
  API for every single tool out. There, that makes total sense.
• 13:28
  So advertisers can manage campaigns via Claude or Cursor.
• 13:31
  And then we have Stripe, who launched Lync, which is a wallet
• 13:35
  specifically for AI agents, but it requires human approval on
• 13:39
  purchases, which makes me wonder, is this human approval a
• 13:42
  temporary safety net or a permanent necessity?
• 13:45
  It's almost certainly permanent, mainly for liability and
• 13:47
  compliance. Stripe issues these virtual
• 13:49
  restricted cards. The agent can autonomously queue
• 13:52
  up a shopping cart via an API call, but the final
• 13:55
  cryptographic handshake requires a human to click approve on
• 13:59
  their phone. Right?
• 14:00
  Because when an agent have financial autonomy, A
• 14:02
  hallucination isn't just a goblin in an essay, it's a
• 14:05
  drained corporate bank account. The human in the loop clears up
• 14:08
  the legal liability chain. Definitely.
• 14:10
  Next, XAI dropped Imagine Agent, an agentic canvas and grok for
• 14:15
  image and video creation. And lastly, some legal drama in
• 14:20
  his trial against open AI, Elon Musk revealed XAI used
• 14:23
  distillation techniques to train on open AI models.
• 14:26
  Yeah, distillation is a highly controversial practice right
• 14:29
  now. It's basically taking the high
• 14:31
  quality outputs from a massive model like Open AI and using
• 14:35
  that generated data to train a smaller, newer model.
• 14:38
  Right to mimic the reasoning. Exactly without spending the
• 14:41
  millions required for ground up training, admitting to this in
• 14:45
  open court is going to force the industry to finally legally
• 14:48
  define who actually owns the IP of an AI's output.
• 14:52
  Yeah, that's going to be huge. Well, you know, we spend a lot
• 14:54
  of time talking about gigawatts of power and legal battles and
• 14:57
  massive corporate infrastructure.
• 14:59
  But ultimately, AI is just a tool.
• 15:01
  Yeah. And the most profoundly human
• 15:03
  way to understand a tool is to see how everyday people use it
• 15:06
  to change their lives. Which brings us to a really
• 15:09
  moving community AI workflow today.
• 15:11
  Oh, I really like this one. Yeah, we have an anonymous
• 15:14
  reader who is in recovery for alcoholism, and they took 1930s
• 15:19
  Alcoholics Anonymous literature, clinical research and
• 15:22
  independent work and loaded it all into notebook LM.
• 15:25
  So how are they applying it day-to-day?
• 15:27
  They use notebook LM to generate daily motivational messages.
• 15:30
  They clear up difficult concepts with pros and cons and create
• 15:34
  speaking notes, but the most impressive part is they turn
• 15:37
  these nuanced historical texts into an audible debate in
• 15:40
  notebook LM so they can actually hear differing views and
• 15:43
  similarities dynamically. That is just a beautiful
• 15:46
  application of the technology. Yeah, Notebook LM synthesizes
• 15:50
  the source material and generates 2 AI voices that
• 15:53
  essentially hosts this custom audio discussion.
• 15:55
  Exactly. And the profound beauty of this
• 15:58
  workflow is taking foundational literature from nearly a century
• 16:01
  ago and using cutting edge generative AI not to change the
• 16:04
  text, but to make the core message deeply personal and
• 16:07
  accessible in a modern auditory format.
• 16:10
  It really is, and it leaves you with the final thought to Mull
• 16:13
  over. Obviously, consider the sheer
• 16:14
  scale of what we discussed today.
• 16:16
  In the exact same breath, AI is simultaneously causing national
• 16:19
  security friction at the highest levels of the Pentagon.
• 16:22
  It's mapping complex corporate code vulnerabilities, and it's
• 16:25
  privately debating 1930s literature in a personalized
• 16:28
  audio format to help a single individual stay sober for one
• 16:32
  more day. Where else in human history has
• 16:35
  a single technology span that kind of operational spectrum?
• 16:38
  Wow, literally nowhere. That is an incredible point to
• 16:41
  end on. Well, that is all we have for
• 16:43
  today's deep dive. Make sure you hit that subscribe
• 16:45
  button to stay updated with this insanely fast moving tech world.
• 16:48
  And if you enjoyed the show, please invite your friends.
• 16:50
  Leave us a five star rating and we'll see you right back here
• 16:53
  for tomorrow's episode.