Latest / AI Fire Daily / 🎙️ EP 386: Google DeepMind’s Interactive Genie 4 & Liquid AI’s Edge Revolution!

Transcript

🎙️ EP 386: Google DeepMind’s Interactive Genie 4 & Liquid AI’s Edge Revolution! · AI Fire Daily

• 0:00
  Imagine a photorealistic video of a bustling
• 0:03
  city. The streets are totally alive with the
• 0:06
  traffic and pedestrians. You type just one single
• 0:10
  sentence into a prompt box and suddenly the digital
• 0:13
  weather changes into this massive thunderstorm.
• 0:17
  The camera angle shifts to a sweeping bird's
• 0:21
  eye view and the people on screen completely
• 0:23
  alter what they're doing, all while the video
• 0:26
  keeps playing. It feels like magic, honestly.
• 0:30
  You know, it's really just relentless, staggering
• 0:33
  computational math underneath. Welcome to the
• 0:36
  Deep Dive. Today, we're looking at this massive
• 0:39
  pendulum swing in AI right now. Yeah, the whole
• 0:41
  industry is like stretching in totally opposite
• 0:43
  directions. Exactly. First, we're going to look
• 0:46
  at Google generating entire interactive universes
• 0:49
  in the cloud. Then we'll contrast that with the
• 0:52
  pretty chaotic financial reality of the AI business
• 0:57
  right now. Seriously turbulent stuff out there.
• 0:59
  And finally, we'll see why the future might actually
• 1:02
  bypass the cloud entirely to run locally on your
• 1:05
  phone. So let's start with Google DeepMind. They
• 1:09
  just officially announced Genie 4. Right, which
• 1:12
  is this fascinating leap in generative tech.
• 1:15
  Genie 4 is a highly advanced world model. Let's
• 1:19
  define that jargon really quick. A world model
• 1:21
  is an AI that learns how physical environments
• 1:23
  act and react. Perfect, yeah. It generates interactive
• 1:27
  environments in real time. So you aren't just
• 1:29
  rendering a static pre -planned video file. Right.
• 1:32
  The simulation actually stays completely active
• 1:34
  while you're prompting it. You just use natural
• 1:37
  language prompts to control it, right? Exactly.
• 1:39
  You can dynamically change the camera angle on
• 1:41
  the fly, seamlessly alter the weather, or, you
• 1:45
  know, trigger random events. The upgrades to
• 1:47
  the actual realism are frankly stunning here.
• 1:50
  Human movement looks much more natural and physically
• 1:52
  grounded. Yeah. And the physics. Collisions and
• 1:55
  aerodynamics behave much closer to actual reality
• 1:58
  now. It's like playing director on a brand new
• 2:00
  universe. The fundamental laws of physics are
• 2:02
  generated on the fly. But why does the accuracy
• 2:06
  of aerodynamics matter so much here? Because
• 2:08
  DeepMind sees a very specific enterprise use
• 2:11
  case. They view this as a safe simulation environment
• 2:13
  for robotics. Real -world robot failure is incredibly
• 2:17
  expensive. Yeah, and sometimes dangerous. Right.
• 2:19
  If a robotic arm breaks a real glass, it costs
• 2:21
  money. Right. And breaking a simulated glass
• 2:24
  is completely free. Exactly. Robots can practice
• 2:26
  highly complex tasks. They master physical mechanics
• 2:30
  inside these simulated worlds first. And it's
• 2:33
  not just visual data being simulated anymore.
• 2:36
  Genie 4 generates native audio alongside the
• 2:39
  video output. That part is so cool. It creates
• 2:42
  ambient sound that perfectly matches the physical
• 2:44
  scene. Yeah. It generates synchronized music
• 2:47
  and interactive dialogue, too. The audio and
• 2:49
  environment are generated completely intertwined.
• 2:52
  It's a massive sandbox for synthetic training
• 2:55
  data. But we should definitely note it isn't
• 2:58
  broadly released yet. It's currently a research
• 3:00
  preview strictly for trusted testers. Google
• 3:03
  says wider access is coming somewhat soon, though.
• 3:06
  Project Genie reaches Google AI Ultra users next
• 3:09
  and then API early access partners. But there
• 3:12
  are definitely still major technical limitations
• 3:14
  involved. Genie 4 currently supports very short
• 3:18
  time horizons. You only get a few minutes of
• 3:20
  continuous stable interaction. Yeah, long sessions
• 3:23
  break the illusion pretty quickly. The system
• 3:25
  really struggles to maintain consistency over
• 3:28
  extended periods. Longer sessions cause objects
• 3:31
  and visual details to drift. I mean, I still
• 3:34
  wrestle with prompt drift myself. The background
• 3:37
  just starts morphing in really weird, unpredictable
• 3:39
  ways. Oh yeah, a parked car might slowly turn
• 3:43
  into a toaster. Right. Oh, and Google is adding
• 3:47
  SynthID to all output. All generated video and
• 3:49
  audio will be permanently watermarked. Which
• 3:52
  is a highly necessary safety guardrail against
• 3:55
  potential deepfakes. We have to know what's real
• 3:57
  and what's simulated. But I wonder, how can robotics
• 4:01
  developers trust the safety training if the simulated
• 4:04
  physics start to hallucinate or drift over time?
• 4:07
  Well, that's the big bottleneck for researchers
• 4:09
  right now. Short task training works perfectly
• 4:12
  fine for simple actions. But long -term planning
• 4:15
  tasks remain highly unreliable. So they have
• 4:18
  to constantly restart it. Yeah, developers have
• 4:20
  to reset the environment. very frequently they
• 4:23
  essentially stitch together short stable simulation
• 4:25
  runs to train so the physics engine breaks down
• 4:28
  if left running too long exactly the illusion
• 4:32
  shatters eventually running these massive universe
• 4:38
  simulating models is incredibly expensive it
• 4:41
  requires an absolutely incomprehensible amount
• 4:44
  of raw compute oh yeah staggering amounts And
• 4:48
  that intense demand leads directly into a harsh
• 4:50
  financial reality check. The AI industry is experiencing
• 4:54
  extreme economic turbulence right now. Yeah,
• 4:57
  the market is shaking violently beneath the surface.
• 4:59
  We're seeing incredible financial highs and terrifying
• 5:03
  lows. Like on the boom side, this firm called
• 5:06
  Disruptive just raised an enormous fund. I saw
• 5:09
  that. Ten billion dollars, right? Yeah. Reportedly
• 5:12
  the largest venture fund in Texas history, writing
• 5:15
  checks up to $1 billion for late -stage AI and
• 5:18
  defense. Wow. Anthropic is also making very aggressive
• 5:22
  strategic market moves. They just slashed Claude
• 5:25
  pricing for U .S. government agencies. Right.
• 5:28
  A massive 40 % discount across the board. Around
• 5:31
  500 agencies already use their models. Why offer
• 5:35
  such a steep discount to the government, though?
• 5:37
  It feels like a massive loss leader strategy.
• 5:40
  Oh, it totally is. It's a highly strategic land
• 5:42
  grab for federal contracts. If you get agencies
• 5:44
  hooked on Claude's specific ecosystem, they won't
• 5:47
  switch. The switching costs just become far too
• 5:49
  high down the road. Exactly. They're embedding
• 5:51
  themselves deeply into critical federal infrastructure.
• 5:54
  This deal could ultimately save the government
• 5:56
  $1 billion. And Anthropic's potential IPO is
• 5:59
  a pretty wild story, too. It could mint billions
• 6:02
  for their early venture capitalists. But Amazon
• 6:04
  might actually emerge as the absolute biggest
• 6:06
  winner there. Right, because of those convertible
• 6:08
  notes. Yeah, Amazon holds $6 .2 billion worth.
• 6:11
  Those specific notes were valued at almost $98
• 6:14
  billion in June. That's crazy. But contrast this
• 6:17
  immense wealth with the looming industry risks.
• 6:20
  Elon Musk recently issued a very stark public
• 6:23
  warning. Yeah, the guppy level death comment.
• 6:25
  Yeah, he says over 200 GPU startups face imminent
• 6:28
  doom. He's talking about the so -called neocloud
• 6:31
  companies, these middle tier firms that lease
• 6:33
  servers to AI developers. And lenders are rapidly
• 6:36
  pulling back from these data centers. Right,
• 6:38
  because they rely almost entirely on AI infrastructure
• 6:41
  revenue. It's incredibly risky. Yeah. They bought
• 6:44
  massive amounts of highly expensive hardware.
• 6:47
  It's kind of like building skyscrapers on shifting
• 6:50
  sand. Right. Like landlords taking huge mortgages
• 6:53
  to buy GPU servers. That is a perfect analogy.
• 6:56
  Yeah. They assume the AI startups renting will
• 6:59
  always have venture capital. When that VC funding
• 7:01
  dries up, the renters immediately default. And
• 7:04
  then the neocloud landlord's entire business
• 7:06
  model collapses overnight. It's a highly precarious
• 7:09
  financial ecosystem. And we're seeing massive
• 7:12
  disruptive shakeups elsewhere, too. A meta AI
• 7:15
  director recently quit his lucrative job. Right.
• 7:17
  He's starting a brand new company called Nondoc,
• 7:20
  taking on Microsoft and Google Workspace directly.
• 7:23
  Nondoc wants all your daily work living in one
• 7:25
  place. Docs, slides, data, and AI running natively
• 7:29
  together. Meanwhile, there are intense legal
• 7:32
  and security battles happening. Elvis Presley's
• 7:35
  estate is currently suing Anthropic in court.
• 7:37
  They're actively seeking $41 million in total
• 7:40
  damages. They allege Claude was trained on 274
• 7:44
  copyrighted songs without any permission. This
• 7:46
  lawsuit represents the ultimate collision we
• 7:48
  keep seeing. Legacy copyright laws are crashing
• 7:52
  into modern synthetic training methods. The legal
• 7:54
  precedent set here will be absolutely massive.
• 7:57
  We also saw a significant concerning security
• 8:00
  patch recently. Cloudflare patched a severe bug
• 8:03
  involving their hosted GLM 5 .3 model. The model
• 8:06
  accidentally exposed other customers' recent
• 8:08
  session data, right? Yeah, between August 27th
• 8:11
  and September 9th. Cloudflare says 155 responses
• 8:14
  were directly affected. It's a stark reminder
• 8:17
  that cloud infrastructure is fundamentally vulnerable.
• 8:21
  Data leakage remains a constant terrifying threat
• 8:23
  for enterprises. But what happens to everyday
• 8:27
  AI developers if these middle -tier neocloud
• 8:31
  GPU lenders suddenly collapse? It would trigger
• 8:34
  a massive, immediate compute supply shock. Developers
• 8:38
  rent server space from these providers daily.
• 8:41
  If those servers go dark, applications break
• 8:44
  instantly. Startups would scramble for expensive,
• 8:47
  larger alternatives, and many small teams would
• 8:49
  just run out of runway. Small developers might
• 8:52
  lose their computing lifelines overnight. Exactly.
• 8:54
  That is exactly the fear. Well, we're going to
• 8:59
  take a quick break for our sponsor. But when
• 9:01
  we come back, we'll look at why developers are
• 9:03
  pivoting away from the cloud entirely. Yeah,
• 9:06
  the shift to local is huge. All right, we're
• 9:09
  back. So because relying on expensive cloud infrastructure
• 9:12
  is so risky, a massive pivot is happening. Developers
• 9:16
  heavily want powerful models that don't need
• 9:19
  the cloud. They're looking for local, secure,
• 9:22
  edge -based solutions. Which brings us to Liquid
• 9:25
  AI. They just dropped something truly fascinating
• 9:28
  for developers. It's a brand new family of models
• 9:30
  called LFM3. Whoa. Imagine spinning up an omnimodal
• 9:35
  AI entirely offline on your phone. It's a profound
• 9:39
  shift in how we approach architecture. LFM3 consists
• 9:42
  of tiny, surprisingly powerful local AI models
• 9:45
  built specifically to run smoothly on edge hardware.
• 9:48
  So they run directly on personal phones and laptops.
• 9:52
  Right. There are six models in this new LFM3
• 9:55
  family. They hover right around the 2 billion
• 9:58
  parameter range. They efficiently cover text,
• 10:01
  vision, and complex audio tasks, plus an omni
• 10:04
  model and models for decision tasks. The specs
• 10:07
  on the 1 .7 billion model are staggering. It
• 10:12
  has a massive 256K context window. Let's define
• 10:15
  context window for some quick clarity. How much
• 10:18
  information an AI can remember in one single
• 10:20
  conversation? Exactly. With a 256K window, it's
• 10:24
  huge. It can essentially read 50 dense PDF reports
• 10:27
  at once. Wow. And it fits in roughly one gigabyte
• 10:31
  of memory. Yeah, because it's running at 4 -bit
• 10:34
  precision. It's incredibly efficient for constrained
• 10:36
  mobile hardware. What does 4 -bit precision actually
• 10:38
  mean in practice, though? Think of it like downsapling
• 10:41
  a massive, high -resolution photograph. You shrink
• 10:44
  it so it fits on an old hard drive. It loses
• 10:47
  some microscopic detail, but keeps the core shapes.
• 10:50
  Exactly. The AI retains its reasoning capability
• 10:53
  while shrinking its footprint. The underlying
• 10:55
  architecture is completely new here as well.
• 10:57
  It uses a complex trihybrid system to process
• 11:01
  data. They mentioned gated convolutions, sliding
• 11:04
  window attention, and global attention. To a
• 11:07
  lot of people, that just sounds like intense
• 11:08
  jargon. It definitely does. Let's break down
• 11:12
  the sliding window attention part. Normally,
• 11:15
  an AI tries to hold an entire document in memory.
• 11:18
  That requires massive compute, which a phone
• 11:20
  simply doesn't have. Right. Mobile devices are
• 11:23
  heavily constrained by thermal limits. Exactly.
• 11:25
  So sliding window attention fundamentally changes
• 11:29
  the processing approach. Think of it like moving
• 11:31
  a magnifying glass across a page. Ah. You're
• 11:35
  only focusing on a small overlapping chunk at
• 11:38
  a time. Right. It compresses the heavy reasoning
• 11:41
  without needing massive memory overhead. That
• 11:44
  perfectly explains how they compress so much
• 11:46
  into one gigabyte. And the base model was pre
• 11:49
  -trained very extensively as well. 22 trillion
• 11:52
  tokens of training data. Yeah, with extra emphasis
• 11:55
  on highly specific practical skills. It excels
• 11:59
  at math, complex code, tool use, and agentic
• 12:02
  traces. All six models share the exact same structural
• 12:06
  backbone, right? Yep. You can easily fine -tune
• 12:09
  them on a single consumer GPU. Vision models
• 12:12
  handle dense images easily. And the Omni model
• 12:15
  parses text, images, video, and audio simultaneously.
• 12:19
  We should clarify the specific licensing model
• 12:22
  here, too. Liquid AI released these as open -weight
• 12:25
  models. They're available through Hugging Face
• 12:28
  and Liquid AI directly. And local versions are
• 12:31
  already appearing in Alama and LM Studio. But
• 12:34
  CNET actually just published a helpful glossary
• 12:37
  of 65 terms. Right. To remind us that open weight
• 12:40
  is not the same as open source. Open source implies
• 12:43
  you have full access to the training data. Open
• 12:46
  weight just gives you the final compiled model
• 12:49
  architecture. It's an incredibly important legal
• 12:52
  and operational distinction. So LFM3 is specifically
• 12:56
  optimized for powering local AI agents. Liquid
• 13:00
  says it handles daily digital tasks beautifully.
• 13:02
  It heavily excels at inbox triage and calendar
• 13:05
  management. It can handle web research and boring
• 13:09
  recurring workflows. And it does all this without
• 13:12
  sending data to the cloud. Your highly personal
• 13:14
  data stays entirely on your physical device.
• 13:18
  But I'm looking at this one gigabyte model and
• 13:20
  I'm highly skeptical. Can it actually handle
• 13:24
  the messy reality of someone's daily inbox without
• 13:27
  hallucinating? Well, it handles routine categorization
• 13:30
  incredibly well. Sorting newsletters, standard
• 13:32
  meeting requests, it has no issues there. But
• 13:35
  it definitely struggles with highly nuanced or
• 13:37
  emotional emails. No, really? So what happens
• 13:40
  then? It's explicitly programmed to route complex
• 13:43
  ambiguity directly to a human. It heavily relies
• 13:46
  on strict, predefined operational boundaries.
• 13:49
  So they handle daily chores, leaving complex
• 13:51
  thinking to the cloud. That's the exact intended
• 13:54
  workflow. Having this kind of local computing
• 13:57
  power is incredible. But to actually utilize
• 14:00
  it, developers need practical frameworks. We
• 14:03
  need reliable tools to bind this all together
• 14:05
  seamlessly. Yeah, and the tooling ecosystem is
• 14:08
  evolving rapidly alongside these models. We're
• 14:10
  seeing fascinating new practical applications
• 14:12
  emerge almost daily. Let's start with a new tool
• 14:15
  called Overmind. This software looks incredibly
• 14:17
  useful for independent developers. It literally
• 14:20
  captures your AI agent's real -world messy work,
• 14:23
  then it turns that work into highly specialized
• 14:26
  local models. You essentially record your absolute
• 14:30
  best problem -solving strategies, and the tool
• 14:32
  converts those into reliable local agents. You
• 14:36
  can train and test these models entirely yourself,
• 14:38
  run them locally for much better speed and performance.
• 14:41
  It acts like an incredibly competent, tireless
• 14:44
  digital twin. Totally. Then there's a testing
• 14:48
  platform called They Said. It focuses heavily
• 14:51
  on the actual product development cycle. Right.
• 14:54
  They Said lets you test ideas directly with real
• 14:57
  customers. You do this crucial step before you
• 14:59
  ever launch a product. And you get instant, actionable
• 15:02
  feedback directly inside your workspace. It integrates
• 15:05
  perfectly with Cloud, Cursor, or Jira workflows.
• 15:08
  Microsoft is also updating their practical enterprise
• 15:11
  tool sets. MAI Cyber 1 .1 Flash is now officially
• 15:16
  live on M -Dash. It matches the previous version's
• 15:19
  intense security capabilities entirely, but crucially
• 15:23
  reduces scanning costs by 18%. That's a massive
• 15:26
  critical margin improvement for large enterprise
• 15:28
  teams. Finally, there's a fascinating new app
• 15:32
  called Orbit. It's specifically designed to block
• 15:35
  highly distracting apps. Orbit employs a genuinely
• 15:39
  fascinating psychological hack. It combats our
• 15:42
  own vicious doom -scrolling habits directly.
• 15:45
  Yeah, you actually earn access back through specific
• 15:48
  friend challenges. Or you can make small micropayments
• 15:51
  to unlock your apps. It uses AI and micropayments
• 15:55
  to brutally enforce discipline. We're using AI
• 15:58
  to fight the dopamine algorithms created by other
• 16:00
  AIs. It literally monetizes your own lack of
• 16:03
  self -control. Which is brilliant, honestly.
• 16:07
  But looking at the dev side, which of these tools
• 16:09
  fundamentally shifts how the solo developer scales
• 16:13
  their time the most? Overmind is definitely the
• 16:16
  absolute clear winner for scale. A solo developer
• 16:19
  can suddenly automate highly complex, bespoke
• 16:22
  workflows. Instead of relying on generic cloud
• 16:25
  APIs. Right. You're building a highly customized
• 16:28
  workforce that lives on your laptop. It fundamentally
• 16:31
  changes the math of what one person can build.
• 16:34
  Overmind basically lets you clone your own best
• 16:37
  workflows locally. It's a massive, unprecedented
• 16:40
  force multiplier. We've covered a tremendous
• 16:44
  amount of ground today. Yeah, and if we step
• 16:46
  back, the narrative through line becomes incredibly
• 16:48
  obvious. We're watching the AI landscape stretch
• 16:51
  dramatically right now. Pushing outward to its
• 16:54
  absolute physical extremes. On one end, you have
• 16:57
  massive cloud -based corporate entities. Think
• 17:00
  of Google's Genie 4 simulating complex physical
• 17:03
  laws. Right. Those giant world models require
• 17:07
  immense centralized and highly expensive infrastructure.
• 17:10
  But on the other end, the cloud business is incredibly
• 17:13
  financially perilous. Neo cloud lenders are facing
• 17:16
  incredible mounting market pressure. So because
• 17:20
  of that severe risk, innovators are aggressively
• 17:22
  compressing AI. Liquid AI is actively shrinking
• 17:26
  incredible reasoning power. They basically compress
• 17:28
  a digital brain into a one gigabyte file. And
• 17:32
  it runs locally, safely, and completely privately
• 17:35
  in your pocket. The stark duality between giant
• 17:38
  and tiny is totally fascinating. It's a profound
• 17:41
  shift in our core technological philosophy. Before
• 17:45
  we go, consider this provocative final thought.
• 17:48
  Local models are getting much smarter every single
• 17:52
  month. Cloud -based world models are becoming
• 17:54
  eerily realistic and complex. How long till the
• 17:57
  tiny AI on your phone changes? What do you mean?
• 18:01
  What if it starts generating its own mini simulations?
• 18:04
  It could simulate thousands of outcomes directly
• 18:06
  inside your pocket. Oh, wow. It could test solutions
• 18:09
  for your daily life completely silently. Then
• 18:11
  it only suggests the absolute optimal path to
• 18:14
  you. So your phone just becomes a portal to thousands
• 18:16
  of simulated futures? Exactly. Just something
• 18:19
  to think about. Thank you for joining the Deep
• 18:21
  Dive today. We always appreciate your endless
• 18:24
  curiosity and your valuable time.