Latest / AI Fire Daily / 🎙️ EP 395: Anthropic’s Fable 5.2 Destroys GPT-6 & Meta’s Ultra-Cheap Muse Ember!

Transcript

🎙️ EP 395: Anthropic’s Fable 5.2 Destroys GPT-6 & Meta’s Ultra-Cheap Muse Ember! · AI Fire Daily

• 0:00
  Imagine a thousand artificial minds working together
• 0:03
  in perfect sync, building a functional operating
• 0:06
  system, and they do it in just eight days. It
• 0:10
  is a staggering reality we face right now. Right.
• 0:13
  And on the exact opposite end, we now have AI
• 0:15
  models that cost literal pennies, just running
• 0:18
  constantly in the background. It is a wild contrast.
• 0:22
  Welcome to today's deep dive. We are unpacking
• 0:25
  a massive stack of updates. We will look at Anthropic's
• 0:28
  new autonomous engineering models, then some
• 0:30
  wild industry drama involving billion -dollar
• 0:32
  power plants, plus a suite of new workflow tools,
• 0:36
  and finally Meta's race to the bottom of AI pricing.
• 0:40
  We are witnessing something profound here. We
• 0:44
  are seeing the simultaneous industrialization
• 0:46
  and commoditization of intelligence. It really
• 0:50
  is a two -front war today. I mean, the ceiling
• 0:52
  of what these models can do is getting unimaginably
• 0:55
  high. But the floor is dropping straight down
• 0:58
  to zero. Let's start with that high end. Anthropic
• 1:02
  just dropped Fable 5 .2 and Mythos 5 .2. I look
• 1:06
  at this update and it feels different. We aren't
• 1:11
  just generating text anymore. We are generating
• 1:13
  complex action. What is going on under the hood?
• 1:17
  No, they really pushed the envelope on complex
• 1:19
  coding and engineering. The raw performance specs
• 1:23
  alone are a massive leap. Outputs are about 30
• 1:26
  % faster across the board. They even included
• 1:29
  an optional mode for 2 .5 times faster token
• 1:32
  generation. And API prices actually dropped 20%,
• 1:35
  right? So we are getting much more capability
• 1:38
  for less capital. $8 per million input tokens
• 1:41
  and $40 for the output. But the benchmarks tell
• 1:44
  the real story of what that capability looks
• 1:46
  like. The benchmarks definitely paint a vivid
• 1:49
  picture. Fable 5 .2 hit 49 .1 % on Frontier Code
• 1:54
  2 .0. That comfortably beats GPT -6 Astra's 42
• 1:58
  .8%. But I want to focus on the remote labor
• 2:01
  index. It jumped massively to 41 .3%. Yeah, that
• 2:06
  is a staggering generational leap. Fable 5 .1
• 2:09
  was sitting at just 17 .9 % on that exact same
• 2:12
  index. You don't see jumps like that very often.
• 2:15
  And we have to mention Mythos 5 .2 as well. It
• 2:18
  boasts a one million token context window. Let's
• 2:21
  clarify that term really quickly. A context window
• 2:24
  is the model's short -term memory during a single
• 2:26
  conversation. Exactly. It's a great way to think
• 2:28
  about it. It holds massive amounts of data in
• 2:31
  its active memory at once. Anthropix has Fable
• 2:34
  5 .2 design three working electronic devices
• 2:36
  from scratch. It didn't just draw a picture.
• 2:39
  It mapped out the intricate circuit pathways.
• 2:42
  Wow. Yeah, and it helped researchers solve two
• 2:44
  distinct black hole physics problems. And then
• 2:47
  there is the operating system project, managing
• 2:50
  1 ,000 individual AI agents to build an OS. You
• 2:55
  have to picture what it takes to build an operating
• 2:57
  system. You need memory management, kernel architecture,
• 3:01
  a user interface. Right, it requires immense
• 3:04
  hierarchical delegation. Fable 5 .2 had to spawn
• 3:08
  an agent for the file system, another for the
• 3:10
  security protocols. Then it monitored their work,
• 3:14
  caught their errors, and integrated the code.
• 3:16
  It basically acts like a digital conductor for
• 3:20
  an orchestra of machines. It changes our fundamental
• 3:23
  relationship with automation. We used to think
• 3:26
  of AI as a tool for a specific task. You ask
• 3:29
  for code, it gives you code. But this feels entirely
• 3:32
  different. It is like hiring a lead architect
• 3:34
  instead of just a bricklayer. Anthropic is pushing
• 3:38
  AI to independently manage entire engineering
• 3:41
  projects. They're moving the industry from simple
• 3:45
  execution to complex orchestration. Though, you
• 3:48
  know, it is worth noting they heavily restrict
• 3:51
  Mythos 5 .2. It's totally locked down for verified
• 3:54
  cybersecurity and life sciences organizations.
• 3:57
  You can't just access it off the street. But
• 3:59
  Fable 5 .2 is widely available on paid plans.
• 4:03
  The ship toward orchestration feels heavy to
• 4:06
  me. Are we seeing the end of human project managers?
• 4:09
  I wouldn't write the obituary for human managers
• 4:12
  just yet. It's much more about augmenting human
• 4:15
  oversight. Humans will still set the ultimate
• 4:17
  goals and ethical boundaries. The AI simply handles
• 4:20
  the hyper -complex coordination underneath that
• 4:23
  umbrella. So it automates the management, not
• 4:26
  just the raw code. Precisely. It frees humans
• 4:29
  to focus on the destination instead of the map.
• 4:33
  But there is a fatal flaw in all of this digital
• 4:36
  capability. It doesn't matter how smart Fable
• 4:39
  5 .2 is if we literally don't have the electricity
• 4:41
  to turn it on. Which brings us to a messy physical
• 4:44
  reality. The energy bottleneck. Yeah. The digital
• 4:48
  world feels infinite, but it is anchored in concrete
• 4:51
  and steel. Anthropic is spending $3 .4 billion
• 4:54
  on natural gas turbines. They need to generate
• 4:58
  1 .7 gigawatts of power. That is an almost incomprehensible
• 5:03
  amount of electricity. It's basically a small
• 5:05
  nuclear reactor's worth of energy. It is deeply
• 5:08
  counterintuitive to me. Anthropic is fundamentally
• 5:11
  a rent -first software company. They don't want
• 5:15
  to own massive physical infrastructure. That's
• 5:18
  like a website developer buying a hydroelectric
• 5:20
  dam just to keep their servers on. Does this
• 5:28
  distract from their core mission? They literally
• 5:31
  have no other choice. The current power grid
• 5:34
  simply cannot handle the load of training these
• 5:36
  next generation models. AI's electricity problem
• 5:39
  is becoming incredibly severe. We abstracted
• 5:43
  hardware away with the cloud a decade ago. Now
• 5:46
  it's back with a vengeance. AI companies are
• 5:49
  being forced to become energy companies by pure
• 5:51
  necessity. And the geopolitical reality surrounding
• 5:54
  this hardware is just as complex. Look at the
• 5:57
  U .S. AI Safety Institute. It's meant to independently
• 6:00
  test AI safety for the public good. Yet it still
• 6:04
  lacks a permanent leader. Right. And staff are
• 6:07
  currently reporting informally straight to the
• 6:09
  White House. The bureaucracy simply cannot keep
• 6:12
  pace with the technology's development speed.
• 6:15
  Then we have some serious drama at the very top
• 6:18
  of the supply chain. Jensen Huang, the CEO of
• 6:20
  NVIDIA, suddenly pulled out of two major events.
• 6:23
  He canceled on the Eleven Lab Summit and Anthropic
• 6:26
  CEO Forum. NVIDIA is actively investigating if
• 6:29
  he was targeted by Chinese intelligence. The
• 6:32
  timing is incredibly sensitive here. NVIDIA is
• 6:36
  the absolute bedrock of this entire AI boom.
• 6:38
  They make the chips that make the magic happen.
• 6:41
  Any threat to their leadership is immediately
• 6:44
  a global security issue. Wired magazine added
• 6:47
  a strange plot twist to the situation. They reported
• 6:50
  Huang hired a high profile crisis communications
• 6:52
  firm, and he did this a full year before admitting
• 6:55
  to an affair. NVIDIA is currently investigating
• 6:58
  that entire situation as well. It's a stark reminder
• 7:01
  that fragile human beings still run these massive
• 7:03
  companies. The pressure at the top of the AI
• 7:06
  food chain must be unimaginable right now. We
• 7:10
  also have some quick updates from the broader
• 7:11
  ecosystem. Google just upgraded AI Studio for
• 7:14
  map -based applications. You can now create 3D
• 7:17
  buildings and Street View apps using simple text
• 7:19
  prompts. That dramatically lowers the barrier
• 7:22
  to entry for spatial computing. You don't need
• 7:25
  a degree in 3D modeling to build interactive
• 7:27
  worlds anymore. Brian Armstrong also revealed
• 7:29
  his personal AI agent setup recently. He calls
• 7:32
  it his second brain productivity stack. It keeps
• 7:35
  his workflows moving constantly while he focuses
• 7:37
  on larger strategies. We're seeing that pattern
• 7:39
  emerge everywhere. High -level executives are
• 7:42
  using agents to handle their digital administration.
• 7:45
  There are also rumors that GPT 6 .2 sole drops
• 7:48
  next Thursday. Leaked benchmarks suggest it might
• 7:52
  beat Opus 5 .5 at a significantly lower cost.
• 7:55
  We have to keep the hype on a tight leash there.
• 7:58
  OpenAI hasn't officially confirmed any of those
• 8:00
  rumors yet. Leaks in this industry are notoriously
• 8:04
  unreliable. The underlying theme remains clear,
• 8:07
  though. The scale of compute requires massive
• 8:09
  physical resources. Is physical energy the ultimate
• 8:12
  bottleneck for AI scale? It absolutely is. Power
• 8:16
  constraints are dictating model growth across
• 8:18
  the entire industry right now. You simply cannot
• 8:21
  train the next generation without securing the
• 8:23
  megawatts first. Data centers are waiting years
• 8:26
  for grid connections. Got it. Physical energy
• 8:29
  grids are the new silicon. Power is the new compute.
• 8:33
  We are going to take a brief pause here. I'll
• 8:35
  be right back. See you in a minute. Welcome back.
• 8:38
  We just spent some time discussing the macro
• 8:40
  -level industry realities, the gigawatts and
• 8:43
  the billion -dollar power plants. Let's bring
• 8:46
  it back down to the micro -level. How are developers
• 8:49
  actually building with these tools today? This
• 8:51
  is where the abstract concepts become highly
• 8:53
  practical. We are seeing a new wave of tools
• 8:56
  that completely change daily workflows. Let's
• 9:00
  start with Higgs Field Miracle. It turns videos
• 9:02
  into editable layers. You can move objects around,
• 9:06
  you can change colors, you can tweak fine details.
• 9:08
  It essentially treats moving video like a Photoshop
• 9:11
  document. But it's not just pasting pixels over
• 9:14
  a flat frame. It actually breaks the video down
• 9:16
  into semantic components. It calculates the 3D
• 9:19
  depth and the light sources of the original video.
• 9:22
  So when you insert or modify an object, it matches
• 9:25
  the scene's geometry perfectly. It profoundly
• 9:28
  changes our relationship with recorded reality.
• 9:31
  Next, we have MCP1. It connects all your MCP
• 9:34
  servers through one secure URL. It works seamlessly
• 9:37
  across Cloud, Cursor, ChatGPT, and Codex. This
• 9:42
  is a massive deal for basic interoperability.
• 9:45
  If you are a developer listening to this, you
• 9:47
  know the ecosystem is incredibly fragmented right
• 9:50
  now. You have five different AI tools running
• 9:52
  in isolated silos. MCP1 acts as a universal translator
• 9:57
  for those different environments. And it is completely
• 9:59
  free for individual users. It addresses a very
• 10:02
  real pain point. I still wrestle with prompt
• 10:04
  drift myself. It is incredibly frustrating when
• 10:08
  that happens. Prompt drift is when an AI changes
• 10:10
  its answers slowly over time. It requires constant
• 10:14
  vigilance just to keep your tools working predictably.
• 10:16
  Having unified protocols helps manage that underlying
• 10:19
  chaos. We also saw the release of the new GitHub
• 10:22
  PR overview. It provides a visual glance at everything
• 10:26
  blocking a pull request. Yeah, it instantly highlights
• 10:29
  failed checks and missing approvals. It drastically
• 10:32
  reduces the cognitive load for developers who
• 10:34
  are just trying to merge code safely. These tools
• 10:37
  all seem aimed at reducing systemic friction.
• 10:39
  Will unified tools like MCP1 prevent AI ecosystems
• 10:43
  from totally fragmenting? They certainly help
• 10:45
  stabilize things. Historically, interoperability
• 10:48
  protocols usually win out in the long run. Developers
• 10:51
  simply refuse to be locked into a single walled
• 10:53
  garden forever. They want flexibility. So it
• 10:56
  essentially builds bridges between competing
• 10:59
  AI models? Exactly. It lets the best tools actually
• 11:01
  talk to each other. This brings us to a fascinating
• 11:04
  pivot in our deep dive. We have these incredibly
• 11:07
  powerful workflow tools. But to be useful, they
• 11:10
  need to run constantly in the background. Meta
• 11:12
  just launched Muse Ember to do exactly that.
• 11:15
  This is Meta's smallest and cheapest AI model
• 11:18
  yet. It is entirely designed for repetitive tasks.
• 11:22
  tool calls and high volume automation. It isn't
• 11:25
  trying to write poetry. It's trying to sort data.
• 11:28
  The pricing structure is almost hard to comprehend.
• 11:32
  It is 10 cents per million input tokens and 50
• 11:37
  cents per million output tokens for standard
• 11:40
  requests. To put that in perspective, a million
• 11:43
  tokens is roughly a dozen thick novels and it
• 11:47
  processes that for a dime. It makes intelligence
• 11:50
  virtually free. And despite that tiny cost, it
• 11:54
  still maintains a 1 million token context window.
• 11:57
  Plus, it handles multimodal inputs effortlessly.
• 12:00
  It can process text, images, audio, and video
• 12:03
  interchangeably. The benchmarks are what really
• 12:06
  surprised me here. It scored 37 .8 % on an automation
• 12:10
  bench. That completely crushes GPT -6 Luna's
• 12:13
  20 .2%. And it easily beats Claude Haiku 5 .5's
• 12:18
  13 .9%. The raw cost efficiency is the real story,
• 12:22
  though. It costs about 14 cents to run a standardized
• 12:25
  benchmark task. Meta's more powerful Muse Spark
• 12:28
  1 .3 costs $2 .39 for the exact same task. Ember
• 12:33
  is roughly 17 times cheaper for nearly the same
• 12:37
  overall score. Meta offers a deeply integrated
• 12:40
  family of models now. Muse Wonder handles the
• 12:43
  complex, high -level strategic planning. Then
• 12:45
  Ember executes thousands of smaller, tedious
• 12:48
  tasks at that rock -bottom cost. They're all
• 12:50
  available directly on Google Cloud now. It's
• 12:52
  a brilliant structural strategy. You don't need
• 12:55
  a supercomputer to filter your inbox or parse
• 12:57
  a spreadsheet. Use the cheap model for the heavy
• 13:00
  lifting. Ember does have some limitations, of
• 13:03
  course. It still trails Cloud Haiku 5 .5 on highly
• 13:06
  demanding software engineering tests. But the
• 13:08
  sheer volume changes the entire equation. Whoa,
• 13:11
  imagine scaling to a billion queries. They want
• 13:15
  ambient intelligence running invisibly in the
• 13:17
  background of everything we do. You won't even
• 13:20
  know it is there. Is a cheap, reliable model
• 13:22
  actually more disruptive than a genius -level
• 13:25
  one? Well, if you look at computing history,
• 13:28
  yes. Ubiquity changes the broader market much
• 13:30
  faster than raw capability. When intelligence
• 13:33
  costs nothing, you embed it in absolutely everything.
• 13:37
  You put it in microwaves, in traffic lights,
• 13:39
  in every single email client. Cheap AI agents
• 13:42
  running millions of tasks reshapes everything.
• 13:45
  It shifts the entire economic foundation of digital
• 13:48
  labor. We have covered a vast amount of ground
• 13:51
  today. Let's slowly synthesize what we are actually
• 13:54
  looking at. We are seeing a profound barbell
• 14:00
  effect taking hold in AI. The middle ground is
• 14:02
  completely disappearing from the market. Exactly.
• 14:05
  On one end, we have Anthropic. They are building
• 14:08
  god -tier models that orchestrate a thousand
• 14:10
  independent agents. They are designing complex
• 14:13
  electronics from scratch. And they require multi
• 14:17
  -billion dollar natural gas power plants just
• 14:19
  to turn the servers on. It is the ultimate centralization
• 14:23
  of raw intellectual power and physical capital.
• 14:26
  And on the exact opposite end of the spectrum,
• 14:29
  we have Meta. They are relentlessly driving the
• 14:32
  cost of intelligence down to fractions of a penny.
• 14:35
  They are making it small enough to run our mundane
• 14:37
  daily developer tools. The spectrum of intelligence
• 14:41
  is stretching in both directions simultaneously.
• 14:44
  It forces you to ask where this all ultimately
• 14:46
  intersects. If AI models are now smart enough
• 14:50
  to coordinate 1 ,000 workers to build operating
• 14:52
  systems and cheap enough to run constantly in
• 14:55
  the background, how long until they start designing
• 14:59
  their own power plants to solve their own electricity
• 15:02
  bottlenecks? That is a fascinating reality to
• 15:05
  ponder. Thank you for joining us on this deep
• 15:07
  dive.