Latest / AI Newspaper / Meta and Apple Team Up to Run a Frontier Open AI Model on Two Desktops

Transcript

Meta and Apple Team Up to Run a Frontier Open AI Model on Two Desktops · AI Newspaper

• 0:00
  Welcome to the AI Newspaper, your daily update
• 0:02
  on everything happening with AI. Visit AINewspaper
• 0:05
  .org to see today's stories and sign up for email
• 0:08
  alerts. So imagine a supercomputer capable of
• 0:12
  outsmarting the world's best artificial intelligence
• 0:14
  models. Right. Now imagine that exact same supercomputer
• 0:17
  is just sitting quietly on your desk, right next
• 0:20
  to your coffee mug, and it's using less electricity
• 0:23
  than your kitchen toaster. It is pretty wild
• 0:26
  to picture. It really is. Because for the past
• 0:28
  few years, the frontier of AI has been dominated
• 0:30
  by this one very specific image, right? Massive,
• 0:34
  sprawling data centers out in the desert, acres
• 0:36
  of humming servers that consume the power of
• 0:38
  a small city. Yeah, that cloud infrastructure
• 0:40
  was definitely the undisputed battlefield. Exactly.
• 0:43
  But we are witnessing a massive paradigm shift
• 0:47
  happening right now. The era of the cloud monopoly
• 0:49
  might have actually just met its match. And the
• 0:52
  contest is moving straight to the hardware sitting
• 0:54
  on your desk. It's true. It really represents
• 0:56
  a fundamental relocation of where the most complex
• 0:59
  computing happens in our world. Because, I mean,
• 1:02
  when you change where the compute lives, you
• 1:05
  change the entire dynamic of the industry. The
• 1:07
  power is moving from those centralized data farms
• 1:09
  back to the edge, back to the user. And the catalyst
• 1:12
  for this massive shift is today's focus. Meta
• 1:16
  just launched a brand new model called MuseSpark
• 1:18
  Open. And let me just give you the scale of this
• 1:21
  thing because it is an absolute behemoth. 816
• 1:25
  billion parameters. Which is massive. Huge. But
• 1:29
  the plot twist here, the thing that makes this
• 1:31
  an entirely new ballgame, is that Meta explicitly
• 1:34
  designed and optimized this massive model to
• 1:37
  run locally on Apple desktop computers. Yeah.
• 1:40
  And that creates a multilayered earthquake across
• 1:43
  the tech sector when you look closely at the
• 1:44
  moving parts. I mean, on one level, you are watching
• 1:47
  a major U .S. tech giant make a highly calculated,
• 1:50
  very aggressive play to reclaim open weight dominance
• 1:53
  from a wave of incredibly successful Chinese
• 1:55
  labs. Right. But then on another level, it's
• 1:58
  a profound technical attempt to decentralize
• 2:00
  massive AI computing. And perhaps most surprisingly,
• 2:03
  for anyone who follows Silicon Valley history
• 2:06
  anyway, this launch represents a sudden stunning
• 2:09
  truce between two of the industry's most bitter
• 2:12
  historical rivals. Oh, I definitely want to get
• 2:14
  into the drama of that truce in a bit because
• 2:16
  the history there is just wild. But before we
• 2:19
  talk about the companies, I really want to understand
• 2:22
  the actual physics of this. Sure. Because squeezing
• 2:24
  an 800 billion parameter model onto a desktop,
• 2:27
  it isn't just a software update. Feels like a
• 2:30
  completely different way of thinking about hardware.
• 2:32
  How is that even physically possible? Well, the
• 2:35
  secret lies in the architecture of MuseSpark
• 2:37
  Open itself. So the model has 816 billion total
• 2:41
  parameters, but it utilizes a mixture of expert
• 2:44
  structure. OK, meaning it doesn't use all of
• 2:47
  those parameters at the same time. Precisely.
• 2:49
  Out of those 816 billion parameters, only 42
• 2:53
  billion are actively engaged for any given token
• 2:55
  during inference. Oh, wow. That's a huge reduction.
• 2:59
  It is. Instead of firing up the entire neural
• 3:01
  network for every single word it generates, the
• 3:04
  system acts like a hyper -efficient router. It
• 3:08
  looks at the prompt and routes the computational
• 3:10
  problem only to the specific expert subnetworks
• 3:13
  that were trained for that specific task. I see.
• 3:16
  So if you ask it to code in Python, it doesn't
• 3:18
  bother waking up the neural pathways trained
• 3:20
  on French poetry. So it's basically preserving
• 3:23
  an enormous amount of computational energy by...
• 3:26
  just keeping the irrelevant parts of its brain
• 3:28
  completely dormant. Exactly. That routing efficiency
• 3:31
  is step one. Step two is its memory capacity.
• 3:35
  Meta actually built this with a one million token
• 3:38
  context window. Wait, one million? So to put
• 3:41
  that into perspective for everyone, a million
• 3:43
  tokens is roughly equivalent to a few thick novels.
• 3:48
  Easily, yeah. If you're a developer, you could
• 3:50
  literally feed this model your entire massive...
• 3:54
  corporate code base, plus a whole stack of technical
• 3:56
  documentation all at once. Or, I mean, even hours
• 4:00
  of raw video, considering it's natively multimodal.
• 4:03
  Right. And that multimodal aspect is critical
• 4:05
  here. It processes text, images, audio and video
• 4:09
  natively without needing any separate translator
• 4:11
  modules. Meta built this thing specifically for
• 4:15
  complex software engineering and long horizon
• 4:18
  tasks. Like an agent. Yes. This is what we call
• 4:21
  an agentic model. It doesn't just act like a
• 4:24
  search engine that spits out an answer to a prompt.
• 4:26
  It can actually formulate a long -term plan,
• 4:29
  use external software tools, debug its own code
• 4:32
  when it hits an error, and execute workflows
• 4:35
  that require dozens of sequential steps. Which
• 4:37
  is incredible, but it brings me right back to
• 4:40
  the physical constraint problem. An 816 billion
• 4:43
  parameter model. I mean, even if it's only using
• 4:45
  42 billion active parameters at a time. it still
• 4:50
  takes up a massive amount of physical storage
• 4:52
  space just to exist on a drive. Oh, absolutely.
• 4:55
  So how does a desktop computer hold a file that
• 4:58
  large in its active memory? That's where quantization
• 5:00
  comes into play. You can think of quantization
• 5:03
  as a highly sophisticated form of compression.
• 5:06
  Okay. In a standard AI model, the parameters,
• 5:08
  you know, the numerical weights that define how
• 5:11
  the model thinks, they are stored in very high
• 5:14
  precision formats. Quantization... mathematically
• 5:18
  rounds those numbers down to a lower precision.
• 5:20
  Got it. It shrinks the physical size of the model
• 5:23
  significantly while retaining almost all of the
• 5:26
  model's actual reasoning capabilities. But even
• 5:30
  compressed, you still need serious iron to run
• 5:33
  it. Right, I'm assuming my standard laptop would
• 5:35
  just instantly crash if I tried to open this
• 5:37
  thing. Oh, a standard laptop wouldn't stand a
• 5:39
  chance. Meta and Apple optimized this for a highly
• 5:42
  specific, very high -end setup. You need to connect
• 5:45
  two to four Apple Mac Studios powered by the
• 5:48
  brand new M5 Ultra chip. Okay, so you do need
• 5:51
  multiple desktop machines. You do. But the reason
• 5:54
  this setup works on Apple hardware, as opposed
• 5:56
  to standard PC desktops, is Apple's unified memory
• 6:00
  architecture. How is that different? Well, traditional
• 6:03
  computers separate the CPU's memory from the
• 6:05
  graphics card's memory. Moving data back and
• 6:08
  forth between them creates a massive bottleneck.
• 6:11
  But Apple's M5 Ultra uses unified memory, meaning
• 6:14
  the integrated GPU has direct access to up to
• 6:17
  512 gigabytes of RAM per single machine. And
• 6:21
  a memory bandwidth of 1 .2 terabytes per second,
• 6:23
  right? Which means it can read that data almost
• 6:25
  instantly. Yeah, the memory bandwidth is the
• 6:28
  real unsung hero here. Generation speed in AI
• 6:30
  is almost always bottlenecked by how fast the
• 6:33
  chip can read the weights from memory, not by
• 6:35
  the processing cores themselves. 1 .2 terabytes
• 6:38
  per second allows the M5 Ultra to pull the necessary
• 6:41
  slices of the model into the processor at breakneck
• 6:43
  speeds. Okay, let's unpack this logistics -wise,
• 6:47
  though. You mentioned connecting two to four
• 6:49
  of these Mac studios together. Right. If you
• 6:51
  are linking completely separate desktop computers
• 6:54
  with physical cables, how do they actually share
• 6:57
  a single AI brain without... massive lag. I've
• 7:00
  heard terms like pipeline parallelization and
• 7:03
  tensor parallelization thrown around a lot, but
• 7:05
  I struggle to visualize the difference. It can
• 7:08
  be confusing, yeah. Is pipeline parallelization
• 7:11
  basically like an assembly line where one Mac
• 7:14
  does the first half of the thought process and
• 7:16
  hands it to the next, while tensor parallelization
• 7:19
  is like breaking a single widget into pieces
• 7:22
  so both Macs can build it at the exact same time?
• 7:24
  That is a perfect analogy, actually. Yes. The
• 7:27
  assembly line is exactly how pipeline parallelization
• 7:29
  works. Mac A handles the first set of neural
• 7:32
  layers, does some math, finishes its part of
• 7:35
  the thought, and hands the result over to Mac
• 7:37
  B. But I'm guessing that creates a bottleneck
• 7:39
  if Mac B just has to sit around waiting for Mac
• 7:42
  A to finish its part of the assembly line. It
• 7:44
  creates significant latency. And that's why this
• 7:47
  new setup uses tensor parallelization instead,
• 7:50
  like you said, breaking the widget into tiny
• 7:52
  pieces. multiple Macs are building their respective
• 7:55
  parts of that exact same mathematical layer at
• 7:59
  the exact same time. So they are splitting the
• 8:02
  actual mathematical matrices across the machine
• 8:04
  simultaneously. They are. But as you might imagine,
• 8:08
  tensor parallelization is notoriously difficult
• 8:11
  across separate physical machines because it
• 8:14
  requires constant instantaneous communication.
• 8:16
  I would think so. The machines have to synchronize
• 8:19
  their math back and forth for every single token
• 8:21
  they generate. If there is even a microsecond
• 8:24
  of delay in the networking cable, the whole process
• 8:26
  just grinds to a halt. So how did they solve
• 8:29
  the network delay between two separate desktop
• 8:31
  computers then? With a very quiet but incredibly
• 8:34
  important software update. Apple's Mac OS 26
• 8:38
  .2 introduced a low -latency RDMA driver. RDMA
• 8:42
  stands for Remote Direct Memory Access. And they
• 8:45
  are running this over Thunderbolt 5 cables. Wait,
• 8:48
  does RDMA mean the computers are bypassing the
• 8:51
  main processor entirely when they talk to each
• 8:53
  other? That's the key right there. Normally,
• 8:56
  when two computers send data, it has to go through
• 8:58
  the operating system, the CPU, the network stack.
• 9:00
  It's a traffic jam. Right. RDMA allows the GPU
• 9:03
  of one Mac Studio to reach directly into the
• 9:05
  memory of the second Mac Studio, completely bypassing
• 9:08
  the operating system overhead. Combined with
• 9:10
  Apple's new MLX framework, it creates a seamless
• 9:13
  mesh. That is crazy. Yeah. To the software, those
• 9:16
  four separate Mac studios appear as one single
• 9:19
  massive supercomputer. That is just brilliant
• 9:21
  engineering. But here is the stat that completely
• 9:24
  caught me off guard. You're talking about a desktop
• 9:26
  supercomputer rivaling data centers. What is
• 9:30
  the actual power draw in this setup? A single
• 9:33
  Mac studio with an M5 Ultra peaks at about 300
• 9:36
  watts during inference. Wait, 300 watts? My kitchen
• 9:40
  toaster uses roughly 1 ,200 watts. Are you saying
• 9:44
  a desktop computer is running a frontier scale
• 9:47
  AI model on a fraction of the electricity it
• 9:49
  takes to toast a bagel? I am. I know. The efficiency
• 9:53
  scaling is profound when you look at the industry
• 9:55
  standard. I mean, a traditional server rack filled
• 9:58
  with NVIDIA GPUs requires thousands of watts,
• 10:01
  plus massive industrial cooling systems. Yeah,
• 10:03
  they use an insane amount of power. Yet on a
• 10:06
  cluster of four Mac studios running tensor parallelization,
• 10:09
  pulling barely over a kilowatt combined, MuseSpark
• 10:12
  Open generates over 36 tokens per second. 36
• 10:15
  tokens per second is faster than the average
• 10:18
  human can read. By a wide margin. You literally
• 10:21
  have frontier level capabilities running efficiently
• 10:23
  on standard wall outlets. Okay, so the engineering
• 10:27
  is a marvel. But once you realize that Meta has
• 10:31
  figured out how to put massive, secure, local
• 10:33
  AI power on a desk with toaster -level efficiency,
• 10:36
  the immediate next question is why? Right. Why
• 10:41
  pivot so hard to local desktop architecture when
• 10:43
  Meta already owns some of the largest cloud data
• 10:46
  centers on Earth? Well, if we connect this to
• 10:48
  the bigger picture, we have to look at the global
• 10:50
  geopolitical landscape of AI, specifically the
• 10:53
  open -weight race. Okay. Because for the past
• 10:56
  couple of years, the momentum in open -weight
• 10:58
  models hasn't been coming out of Silicon Valley.
• 11:00
  Chinese developers have largely dominated the
• 11:02
  top of the leaderboards for models you can download
• 11:04
  and run yourself. We've definitely seen that
• 11:07
  trend. Models like Alibaba's Quen, Moonshot AI's
• 11:10
  Kimi K3, and Zenon .ai's GLM have been setting
• 11:14
  an incredibly fast pace lately. They kind of
• 11:17
  became the go -to resources for developers globally
• 11:19
  who wanted powerful tools without being locked
• 11:22
  into an open AI or Google Cloud API. Which is
• 11:25
  exactly the narrative Meta is trying to disrupt
• 11:27
  here. They are heavily pushing the branding of
• 11:30
  Muse Spark Open as the most capable U .S. open
• 11:34
  weight model to date. Yeah, they've been very
• 11:36
  vocal about that. And they aren't just making
• 11:38
  claims either. The benchmark evaluations actually
• 11:41
  back up the marketing. Muse Spark Open is rivaling
• 11:44
  or outright outperforming those top tier Chinese
• 11:47
  models, particularly in coding, complex agentic
• 11:50
  workflows, and... crucially in safety metrics.
• 11:53
  And if you're a developer, a startup founder,
• 11:55
  or an enterprise CTO, this is where it impacts
• 11:58
  you directly. Think about your core infrastructure.
• 12:01
  Do you want your proprietary code base or your
• 12:04
  sensitive patient healthcare data relying on
• 12:07
  a foreign open weight model or a domestic one?
• 12:10
  Exactly. But I have to push back a little here.
• 12:13
  Does Meta's sudden intense focus on local hardware
• 12:16
  actually give them a... strategic edge over those
• 12:19
  Chinese labs. Couldn't a foreign lab just deploy
• 12:21
  a great model on the cloud? They could, but controlling
• 12:25
  where a model runs is rapidly becoming just as
• 12:28
  critical as the capabilities of the model itself.
• 12:30
  The cloud is a massive vulnerability for enterprise
• 12:33
  data. True. By ensuring developers can run a
• 12:35
  highly capable, entirely domestic U .S. model
• 12:38
  locally, Meta is offering a massive security
• 12:40
  and privacy incentive. I mean, If you run MuseSpark
• 12:44
  Open on a Mac studio sitting under your desk,
• 12:46
  no data ever has to leave your building. Right.
• 12:48
  It never touches a cloud server, not Meta's,
• 12:50
  not Amazon's, not anyone's. The ultimate enterprise
• 12:53
  privacy feature is literally just being able
• 12:55
  to unplug the Ethernet cable. Exactly. Air gapping
• 12:59
  your AI is the gold standard for security. And
• 13:03
  Meta is heavily leaning into that safety aspect.
• 13:06
  In their preparedness testing against cybersecurity
• 13:07
  jailbreaks, which is where attackers try to force
• 13:10
  the model to write malicious code or bypass internal
• 13:13
  safeguards, MuSpark Open refused 99 .2 % of those
• 13:17
  attacks. Wow. How does that compare to the competition?
• 13:20
  Well, in the exact same testing setup, China's
• 13:23
  Kimi K3 refused only 75 .5%. That's a big gap.
• 13:27
  It is. So Meta's pitch to the enterprise market
• 13:30
  isn't just that they have a smart model. The
• 13:32
  pitch is that they have a secure, robust, locally
• 13:35
  executable foundation that a bank or a hospital
• 13:38
  can actually trust inside their own firewalls.
• 13:40
  It's a highly compelling pitch. But to win this
• 13:43
  global open weight race, to actually make that
• 13:46
  local execution strategy work in the real world,
• 13:49
  Meta needed the absolute best local consumer
• 13:51
  hardware available to prove the model was viable.
• 13:54
  They did. They needed massive unified memory
• 13:57
  bandwidth on a machine a developer could actually
• 13:59
  just order online. And that unavoidable hardware
• 14:03
  requirement forced Meta to the negotiating table
• 14:06
  with an old, very bitter enemy. Oh, and calling
• 14:10
  the history between Meta and Apple bitter is
• 14:12
  almost understating the friction there. The relationship
• 14:16
  between these two companies has been notoriously
• 14:18
  toxic for a very long time. Yeah. For years,
• 14:20
  former Apple CEO Tim Cook and Mark Zuckerberg
• 14:23
  clashed very publicly. They were taking personal
• 14:26
  shots at each other's business models constantly.
• 14:28
  Oh, yeah. Cook repeatedly criticized Meta's approach
• 14:31
  to user privacy. And then Zuckerberg fired back.
• 14:34
  famously saying Apple hadn't invented anything
• 14:36
  great since the iPhone and that their products
• 14:39
  were just overpriced. The animosity was highly
• 14:42
  visible, but it wasn't just rhetorical posturing.
• 14:45
  You know, it became deeply financial. Back in
• 14:48
  2021, Apple introduced app tracking transparency
• 14:51
  to iOS devices. Right. The ATT update. Yeah.
• 14:54
  That simple pop up asking users if they wanted
• 14:57
  to be tracked completely decimated Meta's ability
• 14:59
  to serve highly targeted digital ads across mobile
• 15:02
  devices. We really. cannot overstate the financial
• 15:05
  impact of that one software update. Meta estimated
• 15:09
  that Apple's privacy change cost them roughly
• 15:11
  $10 billion in lost revenue in 2022 alone. $10
• 15:17
  billion. That's a staggering number. Tech giants
• 15:21
  don't just hug it out after a targeted hit to
• 15:24
  their core business model like that. Typically,
• 15:26
  no. But the landscape at Apple has shifted dramatically,
• 15:30
  primarily because of a regime change. John Ternes,
• 15:33
  who comes from a deep hardware engineering background,
• 15:35
  took over as Apple's new CEO just last month.
• 15:38
  Right. That shift at the very top of Apple, combined
• 15:41
  with Meta strategically hiring former Apple director
• 15:43
  Daniel Gross onto their AI team last year, seemingly
• 15:46
  paved the way for this unprecedented collaboration.
• 15:49
  The cultural thaw is fascinating to watch. Meta's
• 15:54
  AI chief, Alexander Wang, was remarkably blunt
• 15:57
  about the whole thing. He actually joked to the
• 15:59
  press. He said, we're excited for mommy and daddy
• 16:01
  to be playing nice again. Yeah, that was a funny
• 16:03
  quote. You even had Mark Zuckerberg publicly
• 16:06
  thanking Ternus on X, posting a picture of the
• 16:09
  Muse model running on a Mac. And Ternus replied
• 16:11
  by praising the boundary -pushing nature of the
• 16:14
  Mac studio. It's just a complete 180 -degree
• 16:17
  turn. It is. But it raises an important question
• 16:21
  about Apple's broader strategy in the AI space.
• 16:23
  Right, because there's a glaring contradiction
• 16:25
  here. Apple already has major, highly publicized
• 16:28
  partnerships with OpenAI to integrate chat GPT
• 16:31
  into Siri, and with Google to use Gemini inside
• 16:33
  Apple Intelligence. Exactly. So what does this
• 16:37
  all mean for Apple? Are they just frantically
• 16:39
  hedging their bets because their own AI efforts
• 16:41
  have faced criticism? Or is this a calculative
• 16:43
  move to challenge NVIDIA's dominance by proving
• 16:46
  Apple Silicon is the ultimate hardware for local
• 16:48
  AI? I'd lean heavily toward the latter. I think
• 16:51
  this is a massive strategic flex for Apple against
• 16:54
  Nvidia's total dominance. Apple is looking at
• 16:57
  the AI landscape and realizing they hold a very
• 17:00
  unique advantage. They control the most powerful
• 17:03
  consumer silicon on the market. By hosting Meta's
• 17:06
  massive frontier class models natively on Apple
• 17:10
  Silicon using their MLX framework, Apple proves
• 17:13
  a vital concept to the entire developer ecosystem.
• 17:16
  They are showing developers that they don't necessarily
• 17:18
  need to rent expensive time in an NVIDIA -dominated
• 17:21
  cloud data center. You can bypass the cloud entirely
• 17:24
  if you just buy Apple hardware. The symbiosis
• 17:26
  is brilliant. Meta actually bought hundreds of
• 17:29
  Mac computers earlier this year just so their
• 17:31
  super intelligence team could experiment and
• 17:34
  optimize the MLX framework. Through this truce,
• 17:36
  Meta gets incredibly powerful distribution for
• 17:39
  its open -weight U .S. model. And Apple gets
• 17:42
  to sell extremely high -margin hardware. Which
• 17:45
  brings us to the reality check of this whole
• 17:47
  situation. The hardware synergy sounds incredible,
• 17:51
  but bringing this power to the desktop comes
• 17:53
  with significant barriers to entry for the average
• 17:56
  user. Let's look at the fine print, starting
• 17:59
  with the hardware paywall. Right, because as
• 18:02
  we established, you cannot run MuseSpark Open
• 18:04
  on a base -level MacBook. The barrier to entry
• 18:08
  requires enterprise -level spending. Let me...
• 18:11
  Break down the math for you. A single Mac Studio
• 18:13
  equipped with an M5 Ultra chip and 256 gigabytes
• 18:17
  of unified memory currently costs at least $10
• 18:20
  ,799. And that's just the starting point. Exactly.
• 18:25
  And remember, the 512 gigabyte version required
• 18:27
  for optimal use isn't even out until later this
• 18:30
  month, which will obviously cost more. Plus,
• 18:32
  you need between two and four. of these machines
• 18:36
  linked together to run the model effectively.
• 18:38
  You are looking at tens of thousands of dollars.
• 18:41
  Yeah, it definitely prices out the casual coding
• 18:44
  hobbyist. This hardware setup targets well -funded
• 18:47
  startups, enterprise developers, and university
• 18:50
  research labs. But the upfront hardware cost
• 18:53
  isn't the only caveat we need to examine. We
• 18:56
  also have to look closely at the software side,
• 18:58
  specifically the word open in Muse Spark Open.
• 19:02
  Because open weight does not mean open source
• 19:05
  in the traditional unrestricted sense of the
• 19:08
  word. It doesn't. Meta is releasing the weights
• 19:11
  under a specific agreement called the Muse model
• 19:13
  license. Now, for the vast majority of independent
• 19:16
  developers, it is completely free to use. However,
• 19:20
  there is a major catch. Here's where it gets
• 19:22
  really interesting. If a company makes over $20
• 19:25
  million in annual revenue and they want to offer
• 19:28
  the model as a model as a service. Meaning they
• 19:31
  want to provide third parties access to its outputs
• 19:34
  via a paid API? Correct. They must explicitly
• 19:38
  notify Meta and enter into a separate customized
• 19:41
  commercial agreement. Which completely reframes
• 19:44
  Meta's entire strategy here. I mean, it's like
• 19:46
  Ford giving you the highly detailed blueprints
• 19:48
  to a world -class racing engine for free, but
• 19:51
  you can only actually build it if you buy $40
• 19:54
  ,000 worth of proprietary tools. And if you start
• 19:57
  winning too many paid races with it, Ford taps
• 20:00
  you on the shoulder and demands a cut. Is this
• 20:03
  truly open AI? Or is it just a clever strategy
• 20:06
  to get elite developers to tinker, get locked
• 20:09
  into the meta ecosystem, and eventually upgrade
• 20:11
  to meta's proprietary paid models like Muse Wonder?
• 20:15
  What's fascinating here is that the Ford analogy
• 20:17
  holds up perfectly, especially when you look
• 20:20
  at what meta's own leadership is saying. Oh,
• 20:22
  really? Yeah. Alexander Wang openly stated that
• 20:25
  these are not their most capable models. He framed
• 20:29
  this release as a way for developers to tinker
• 20:31
  at a small scale before ramping up. It is a highly
• 20:35
  calculated move to seed the market. The strategy
• 20:38
  is to let developers build their agents, integrate
• 20:40
  into the meta architecture, and then when those
• 20:43
  startups grow and need enterprise scaling, they
• 20:45
  are already locked into Meta's ecosystem. It's
• 20:48
  the classic Silicon Valley playbook, just executed
• 20:51
  at a scale we've never really seen before. Exactly.
• 20:55
  Which leaves us with a final thought to ponder
• 20:57
  as we watch this local compute revolution unfold.
• 21:00
  Yeah. Today, running a frontier class model locally
• 21:03
  requires tens of thousands of dollars in desktop
• 21:05
  hardware. But if history teaches us anything,
• 21:08
  it's that technology always shrinks. Relentlessly.
• 21:12
  That's very true. If the AI contest is permanently
• 21:15
  shifting from who owns the biggest server farm
• 21:17
  to who controls the silicon on the edge, what
• 21:20
  happens to the balance of global power when this
• 21:22
  exact toaster -level computing efficiency scales
• 21:26
  down to the smartphone in your pocket. Thanks
• 21:29
  for listening to the AI Newspaper. Visit AINewspaper
• 21:32
  .org to catch up on the latest updates and sign
• 21:34
  up for free email alerts.