Latest / AI Fire Daily / 🎙 EP 374: Karpathy's 4-Tier AI Gap & Sakana AI Introduces MASS Self-Training Agent Teams

Transcript

🎙 EP 374: Karpathy's 4-Tier AI Gap & Sakana AI Introduces MASS Self-Training Agent Teams · AI Fire Daily

• 0:00
  What if the AI you use every day isn't the real
• 0:03
  AI? Let's unpack this. Yeah, that is a wild thought
• 0:08
  to start with. Welcome to today's deep dive into
• 0:11
  the sources. We are examining a fascinating stack
• 0:14
  of insights today. They reveal a massive hidden
• 0:17
  disconnect in our world. You are really going
• 0:19
  to want to hear this. We are unpacking why most
• 0:22
  of the globe misses the real revolution. We will
• 0:25
  explore how AI is quietly infiltrating your local
• 0:28
  devices. And why tech giants are suddenly begging
• 0:31
  for an emergency break. Finally, we examine a
• 0:34
  mind -bending breakthrough. AI is forming virtual
• 0:37
  teams with its clones to teach itself. It is
• 0:41
  a profound split in our daily reality. The public
• 0:44
  and the experts are living on different planets.
• 0:46
  Let us start with how people actually use this
• 0:48
  tech. We are looking at Andrej Karpathy's breakdown
• 0:51
  of the landscape. Right. He calls this phenomenon
• 0:54
  the great AI usage gap. It explains why your
• 0:57
  experience feels so different from the headlines.
• 0:59
  People are effectively experiencing entirely
• 1:02
  separate tiers of capability. The numbers in
• 1:05
  this report are honestly staggering. About 6
• 1:08
  billion people have basically never touched a
• 1:11
  language model. That is roughly 75 % of the human
• 1:14
  population. Just zero interaction at all. Yeah,
• 1:18
  global access remains heavily restricted by basic
• 1:21
  infrastructure. Then you look at the next tier
• 1:24
  of users. Roughly 1 to 2 billion people use free
• 1:27
  chatbots occasionally. They might write a quick
• 1:30
  email, you know, or they ask for a simple dinner
• 1:32
  recipe. Right. They treat it like a slightly
• 1:35
  better search engine. And that completely warps
• 1:37
  their perception of the technology. They think
• 1:39
  it is just a basic text generator. You are probably
• 1:42
  using it the same way I am. But using free bots
• 1:45
  for small questions feels so limiting. It really
• 1:48
  does. It's like driving a Ferrari 10 miles an
• 1:51
  hour. You're just crawling around a grocery store
• 1:53
  parking lot. I still wrestle with prompt drift
• 1:56
  myself, honestly. Uh, yeah. AI slowly forgetting
• 1:59
  your original instructions over time. Exactly.
• 2:02
  It gets incredibly frustrating. We all deal with
• 2:05
  it. But Karpathy explains why this gap exists.
• 2:08
  Short questions do not stress frontier models
• 2:11
  at all. Frontier models being the most advanced,
• 2:14
  powerful AI systems available. Right. The real
• 2:16
  magic happens on highly complex projects. We
• 2:20
  are talking about tasks that run for hours or
• 2:23
  even days. Right. And look at the professional
• 2:26
  usage tier. Only about 20 million people use
• 2:29
  frontier models for real work. That is just 0
• 2:32
  .2 % of the world. Yeah. They use it for advanced
• 2:35
  coding or heavy math. And at the very top of
• 2:38
  this pyramid, just 5 ,000 elite researchers.
• 2:41
  Only they get access to the raw, unrestricted
• 2:44
  systems. Because that is where the models actually
• 2:46
  shine brightest. They excel in purely digital,
• 2:49
  verifiable domains right now. Areas having a
• 2:52
  provable right or wrong answer. So the free versions
• 2:55
  give you a distorted reality. You cannot judge
• 2:58
  the technology by the chat window. You really
• 3:00
  can't. Internal systems offer a drastically different
• 3:03
  intelligence. I am trying to wrap my head around
• 3:06
  this. Why do verifiable domains matter so much
• 3:09
  for these models? Having a provable right or
• 3:12
  wrong answer changes everything. The system can
• 3:15
  definitively learn from its specific mistakes.
• 3:17
  It does this entirely without any human guidance.
• 3:20
  So math gives AI a clear, undeniable right answer
• 3:25
  to learn from. Exactly. It creates a perfect
• 3:28
  loop for reinforcement learning. Because simple
• 3:31
  chat windows do not stress these models, things
• 3:34
  are shifting. Developers want these systems to
• 3:36
  show their true value. So they are ripping AI
• 3:39
  right out of the browser. They are embedding
• 3:41
  it directly into our daily workflows. We are
• 3:43
  seeing entirely new professions emerge from this.
• 3:46
  A former McKinsey analyst just coined a fascinating
• 3:49
  term. They call this emerging role the marketing
• 3:52
  engineer. It perfectly blends marketing strategy,
• 3:55
  coding, and AI skills. They use these combined
• 3:58
  skills to drive massive results. And the embedded
• 4:01
  tools they use are getting incredibly powerful.
• 4:04
  We are moving way beyond simple text generation.
• 4:07
  Take a look at a new platform called Hierarchy.
• 4:09
  It acts as a live voice coach for tech interviews.
• 4:13
  It actually listens and helps tailor your spoken
• 4:15
  answers. That is a massive upgrade from a basic
• 4:18
  text prompt. Then there is a workspace tool called
• 4:21
  Codelo. It creates personalized pitch decks entirely
• 4:24
  automatically. Perplexity Computer Automations
• 4:27
  is another huge leap forward. It handles recurring
• 4:30
  computer work completely in the background. It
• 4:32
  can compile complex reports while you sleep.
• 4:35
  But the Ghost Core software really caught my
• 4:37
  eye. It runs its AI directly on your local computer
• 4:40
  hardware. It seamlessly connects your apps, files,
• 4:44
  and devices. It does not just read your local
• 4:46
  files. It maps the relationships between your
• 4:49
  calendar and your folders. And it does this locally
• 4:52
  without pinging the cloud. Right, and that is
• 4:54
  a massive shift for privacy. It learns your highly
• 4:57
  personal context to help you proactively. We
• 5:00
  are seeing similar things with the Grok bot.
• 5:02
  Users can now assign bots actual dedicated email
• 5:05
  addresses. The bot just sits in the background
• 5:08
  handling your inbox. It organizes your schedule
• 5:10
  and manages your newsletters. But you need serious
• 5:13
  computing power to run this constantly. Hardware
• 5:16
  is rapidly shifting to support these invisible
• 5:18
  background agents. Apple is pushing hard with
• 5:21
  their new Home View screen. It is a wall -mounted
• 5:24
  Siri hub featuring an A18 chip. Which brings
• 5:27
  us to a massive infrastructure play. Amazon is
• 5:30
  currently discussing a $7 billion acquisition
• 5:32
  deal. They want to buy a software company called
• 5:35
  Descartes. The card claims their software is
• 5:37
  a total game changer. It makes AI models run
• 5:41
  eight times faster across chips. It works on
• 5:43
  NVIDIA, Google, and the Tranium hardware. I have
• 5:47
  to say this makes me a bit nervous. If Grok is
• 5:50
  answering my emails autonomously... Yeah, it
• 5:53
  is a lot of access. And Ghost Core is constantly
• 5:56
  reading my local tax files. Aren't we just handing
• 5:59
  over the keys to our digital lives? We absolutely
• 6:02
  are. It requires a massive leap of faith. I am
• 6:06
  trying to connect the dots with this Amazon deal.
• 6:08
  How does the Amazon Descartes deal directly fuel
• 6:11
  these background agents? Running models eight
• 6:14
  times faster drastically lowers compute costs.
• 6:17
  This makes constant, invisible background AI
• 6:19
  financially feasible for everyone. Faster hardware
• 6:22
  means cheaper AI, unlocking these invisible background
• 6:26
  agents. Precisely. It's all about lowering the
• 6:30
  friction to zero. Let's take a quick second here.
• 6:33
  Be right back. All right. Let us dive right back
• 6:37
  into the sources today. We were just talking
• 6:40
  about handing over the digital keys. Right. You
• 6:43
  can't grasp why tech giants want emergency brakes
• 6:46
  until you see what these models do locally. If
• 6:50
  we let these autonomous agents run in the background.
• 6:52
  We desperately need to talk about who controls
• 6:55
  the brakes. The tone in the tech industry has
• 6:57
  shifted drastically recently. Satya Nadella made
• 7:01
  a major public statement about this. Microsoft's
• 7:04
  CEO explicitly said AI desperately needs an emergency
• 7:07
  break. He wants authorized humans to pause models
• 7:11
  mid -task. He is demanding tamper -proof logs
• 7:14
  for every single meaningful action. The industry
• 7:17
  is realizing that agents need structural guardrails.
• 7:21
  Codex is already updating its internal rules
• 7:23
  for safety. Users must now explicitly ask the
• 7:26
  system for a reset. Or a reset happens automatically
• 7:29
  when usage limits hit 10%. Nothing continues
• 7:32
  running in the background without human oversight.
• 7:34
  It keeps humans firmly in the driver's seat.
• 7:37
  Then you have Anthropic launching a service called
• 7:40
  OSS Scanner. This is a fascinating approach to
• 7:43
  autonomous safety. They are using their absolute
• 7:45
  strongest models, including Mythos. It actively
• 7:48
  scans open source projects to find hidden vulnerabilities.
• 7:51
  It is literally hunting for flaws in the code.
• 7:54
  Then it proactively suggests complex fixes for
• 7:57
  those discovered flaws. It acts like a highly
• 7:59
  advanced autonomous cybersecurity employee. But
• 8:03
  it operates entirely within strict, monitorable
• 8:06
  boundaries. It is like installing a physical
• 8:09
  kill switch on a train. An autonomous freight
• 8:12
  train needs a manual override. That is a perfect
• 8:15
  analogy. You need to kill the engine safely.
• 8:18
  If the agent jumps the tracks, you stop the momentum
• 8:20
  immediately. Nadella's push for activity logs
• 8:23
  makes total sense to me. But why do the logs
• 8:26
  need to be specifically tamper -proof? If an
• 8:29
  autonomous agent makes a catastrophic error,
• 8:32
  we need evidence. We need an unalterable record
• 8:35
  of its internal thought process. This allows
• 8:37
  engineers to diagnose exactly why the logic failed.
• 8:40
  We need an unerasable black box for tracking
• 8:43
  autonomous AI actions. Exactly. Without it, debugging
• 8:47
  an autonomous agent is basically impossible.
• 8:50
  While tech CEOs are actively trying to build
• 8:53
  reliable brakes... The AI itself is learning
• 8:57
  to press the accelerator. Yeah, and it is learning
• 9:01
  to do this without us. This brings us to Sakana
• 9:04
  AI's recent announcement. They just revealed
• 9:07
  a truly mind -bending new breakthrough in machine
• 9:10
  learning. They call it multi -agent self -supervision.
• 9:13
  Or MASS for short. Let us define that core concept.
• 9:18
  Clearly for our listeners. Self -supervision
• 9:20
  means AI improving itself using its own generated
• 9:24
  data without human labels. That is exactly right.
• 9:26
  And Sakana took this to an extreme. They used
• 9:29
  one shared 27 billion parameter model. A mathematical
• 9:33
  framework defining how AI processes information.
• 9:36
  Right. This single model plays three distinct
• 9:39
  roles simultaneously. It acts as the task solver,
• 9:42
  workflow optimizer, and evaluator. It actually
• 9:46
  creates entire teams of its own virtual clones.
• 9:49
  Then it searches for the best ways for them to
• 9:51
  collaborate. Think of Mass like a highly advanced
• 9:54
  digital test kitchen. One AI clone acts as the
• 9:58
  executive head chef. Another clone acts as the
• 10:00
  line cook trying new recipes. And a third clone
• 10:03
  acts as the harsh food critic. They run complex
• 10:06
  synthetic research tasks in these virtual teams.
• 10:10
  Then the system trains itself on the best team's
• 10:12
  executions. Wow. Imagine a model scaling up an
• 10:17
  entire virtual boardroom of itself to solve a
• 10:20
  problem. It sounds like pure science fiction.
• 10:23
  But the benchmark results are very real and strictly
• 10:26
  verified. Sakana tested MAST for two complete
• 10:29
  self -improvement cycles. After the second cycle,
• 10:31
  the score per output token improved drastically.
• 10:34
  An output token is a single piece of generated
• 10:37
  text. The performance improved 1 .2 to 1 .6 times.
• 10:41
  This was measured strictly against the original
• 10:43
  base model's capabilities. Multi -agent training
• 10:46
  data proved significantly more efficient than
• 10:48
  single agents. But improving the task solving
• 10:51
  model had a bizarre side effect. It unexpectedly
• 10:54
  improved its ability to act as the optimizer.
• 10:57
  And it became a much better evaluator of its
• 10:59
  own work. The skills transferred across the different
• 11:01
  roles seamlessly. But researchers found a fascinating
• 11:05
  new bottleneck in this self -teaching system.
• 11:08
  The optimizer easily becomes the limiting factor
• 11:10
  in this loop. Now, if the model evaluates itself,
• 11:13
  I am... highly skeptical. Isn't it just grading
• 11:17
  its own homework for an easy A? Sakana anticipated
• 11:20
  that exact criticism about self -evaluation bias.
• 11:24
  MASS does not rely entirely on internal judges
• 11:26
  during training. They brought in external models
• 11:28
  afterward to verify the results. They used GPT
• 11:31
  5 .5 and Claude Opus 4 .8. So they used them
• 11:36
  as independent evaluators to verify the reported
• 11:39
  gains. That makes total sense. It proved the
• 11:42
  self -teaching actually worked, objectively.
• 11:45
  You mentioned a fascinating new bottleneck in
• 11:47
  the system earlier. What does it actually mean
• 11:50
  that the optimizer becomes the bottleneck? The
• 11:53
  AI easily recognizes when a final outcome is
• 11:55
  vastly superior, but it lacks the creative reasoning
• 11:58
  to map out novel workflows. It struggles to invent
• 12:01
  the new steps required to achieve it again. It
• 12:04
  sees the finish line, but completely forgets
• 12:06
  the route. That is a great way to put it. It
• 12:08
  knows what good looks like, but not how to get
• 12:11
  there. We have covered some incredible ground
• 12:13
  in today's deep dive. We are clearly living in
• 12:16
  completely different realities of AI capability.
• 12:19
  The contrast across the globe is incredibly jarring
• 12:22
  to see. 75 % of the world is barely touching
• 12:25
  this technology. Meanwhile, the bleeding edge
• 12:27
  is building agents to run our computers. We are
• 12:30
  necessitating emergency breaks just to maintain
• 12:33
  safe human control. And the models are spawning
• 12:37
  virtual clones to train themselves autonomously.
• 12:40
  It leaves us with a very lingering provocative
• 12:43
  question today. What is that? If AI models are
• 12:47
  now effectively collaborating in virtual teams
• 12:49
  to train themselves, how long until human teams
• 12:52
  are just the emergency brake operators watching
• 12:54
  them work? That is a heavy thought. I want you
• 12:59
  to reflect on your own tier of AI usage. Are
• 13:02
  you pushing the limits of verifiable tasks in
• 13:05
  your daily work? Or are you just using a slightly
• 13:07
  better Google? Keep exploring this technology.
• 13:11
  And always keep questioning the sources.