Latest / AI Fire Daily / 🎙️ EP 360: xAI Unveils Grok 4.7 Coding Model & Xiaomi Open-Sources MiMo-V2.5 Suite

Transcript

🎙️ EP 360: xAI Unveils Grok 4.7 Coding Model & Xiaomi Open-Sources MiMo-V2.5 Suite · AI Fire Daily

• 0:00
  A faulty A .I. report just claimed a Chinese
• 0:02
  ship was carrying a nuclear weapon. The stakes
• 0:05
  for autonomous A .I. just got incredibly real.
• 0:08
  Yeah, it's a terrifying wake up call. Honestly,
• 0:10
  we are crossing a major threshold here. Welcome
• 0:13
  to this deep dive. Today, we're exploring a massive
• 0:16
  shift in the technology landscape for you. We're
• 0:18
  tracking how A .I. is graduating from a tool
• 0:21
  that just answers our questions into an independent
• 0:23
  agent. taking autonomous actions in the real
• 0:26
  world. Right. The ground is shifting very quickly
• 0:28
  underneath us, like way faster than most people
• 0:31
  realize. So here is the roadmap for today. We'll
• 0:34
  start by looking at Grok 4 .7 and its new coding
• 0:37
  autonomy. Then we'll examine the wild west of
• 0:41
  AI agents hitting real world friction. We'll
• 0:43
  look at things like Amazon bans and that chilling
• 0:45
  military false alarm. Finally, we'll see how
• 0:48
  Xiaomi is open sourcing this entire revolution.
• 0:51
  It's a fascinating progression, you know, going
• 0:53
  from the lab. straight into the real world. Let's
• 0:55
  start with the foundation. If we want to understand
• 0:57
  how AI is acting in the wild, we have to look
• 1:00
  at how it's being trained behind the scenes.
• 1:01
  Exactly. And that brings us directly to XAI.
• 1:05
  They just released Grok 4 .7, and this isn't
• 1:08
  your standard chatty assistant. The entire focus
• 1:11
  of this model is on long horizon reasoning. Specifically
• 1:14
  for complex software engineering, right? Yeah,
• 1:17
  exactly. They're relying heavily on extended
• 1:19
  reinforcement learning to pull this off. Let's
• 1:22
  define that quickly for you. Reinforcement learning
• 1:24
  is training AI by rewarding good actions and
• 1:27
  punishing bad ones. Right. And that underlying
• 1:30
  training allows Grok 4 .7 to do something crucial.
• 1:33
  It can work for a much longer time on complex
• 1:35
  tasks, so it doesn't just spit out a quick script
• 1:38
  and call it a day. It actually understands entire
• 1:40
  software repositories. Yeah, it modifies multiple
• 1:43
  files across the system. It runs its own tests.
• 1:45
  But the real breakthrough here is the self -verification
• 1:47
  mechanism. It checks its own outputs before finishing
• 1:51
  the job. It's like stacking Lego blocks of data,
• 1:53
  checking each piece before adding the next. I
• 1:56
  have to admit something here. I still wrestle
• 1:58
  with trusting AI to write code myself. It feels
• 2:01
  a bit like, I don't know. Handing the keys to
• 2:03
  a teenager. Oh, totally. That's a very normal
• 2:06
  hesitation. But the mechanics of self -verification
• 2:09
  kind of change that dynamic. Think of old AI,
• 2:12
  like a student taking a math test with a pen.
• 2:15
  Whatever they write down is final. And one mistake
• 2:17
  early on ruins the final answer. Right. Whereas
• 2:20
  self -verification gives the AI a pencil and
• 2:22
  an eraser. Exactly. It does the math, realizes
• 2:26
  it forgot to carry the one, erases the error,
• 2:28
  and rewrites it before handing the test to you.
• 2:31
  That drastically changes the reliability of the
• 2:33
  output. It really does. And the numbers back
• 2:35
  that up. Grok 4 .7 hit 46 .3 % on CursorBench
• 2:39
  4 .0. It also scored 71 .0 % on DeepSWE version
• 2:44
  1 .1. Wow. Those are incredibly tough coding
• 2:48
  benchmarks. Yeah, they are. They also hit an
• 2:50
  impressive 1 ,657 on the ELO benchmark. Which
• 2:53
  is a chess -style ranking system matching AIs
• 2:55
  against each other's performance. Right. And
• 2:57
  despite these high -end reasoning capabilities,
• 2:59
  XAI kept the... identical to Grok 4 .6. Really?
• 3:02
  So that's $2 per million input tokens and $6
• 3:06
  per million output tokens. Yeah, they're competing
• 3:08
  fiercely on high -end reasoning, but they're
• 3:10
  completely undercutting the market on token price.
• 3:12
  So why is the self -verification step such a
• 3:15
  game -changer for developers? Because it virtually
• 3:18
  eliminates compounding errors on long tasks.
• 3:21
  Without it, an AI might write 90 % of a program
• 3:24
  perfectly. But one tiny syntax error in step
• 3:28
  two breaks the whole build. Self -verification
• 3:31
  means it spots that fatal error in step two,
• 3:34
  rewrites it, and ensures the final product actually
• 3:37
  runs. So it catches its own mistakes before breaking
• 3:40
  your entire software build. Exactly. It turns
• 3:42
  the AI from a sloppy intern into a senior reviewer.
• 3:46
  If Grok is acting as a senior reviewer in the
• 3:48
  background, it is only a matter of time before
• 3:50
  companies push that autonomy into the foreground.
• 3:53
  And that is exactly what Meta just did with their
• 3:55
  Muse agent. Yeah, they pushed it out, but they
• 3:57
  immediately hit a brick wall called Amazon. Right.
• 3:59
  Well, the consumer demand for autonomous agents
• 4:01
  is astronomical. I mean, Muse hit 1 .8 million
• 4:05
  iOS downloads in just 12 days. Which is crazy
• 4:08
  because that actually beats ChatGPT's early numbers
• 4:10
  in the U .S. and Canada. People clearly want
• 4:13
  an AI that doesn't just talk, but actually does.
• 4:16
  things for them but the digital infrastructure
• 4:18
  is pushing back hard amazon just blocked the
• 4:22
  muse agent from shopping on its platform yeah
• 4:24
  they labeled it an unauthorized agent they claimed
• 4:26
  it violates their terms of service which if you
• 4:29
  think about it from amazon's perspective makes
• 4:31
  total sense Their entire business model relies
• 4:34
  on human eyeballs looking at sponsored products
• 4:37
  and ads. Right. And they want to control that
• 4:40
  checkout experience. If a bot just scrapes the
• 4:42
  lowest price and buys it in the background, Amazon
• 4:45
  loses its advertising leverage completely. But
• 4:48
  the friction with autonomous agents goes far
• 4:51
  beyond online shopping. This brings us back to
• 4:53
  our opening hook. Oh, right, that military report.
• 4:56
  A faulty AI -generated report almost triggered
• 4:59
  a military operation. It falsely claimed a Chinese
• 5:03
  ship was carrying a nuclear weapon. That is chilling,
• 5:05
  truly. It's the hallucination problem we've always
• 5:08
  known about, but scaled up to global security.
• 5:11
  Exactly. When an AI hallucinates a line of code,
• 5:14
  your app crashes. When an autonomous intelligence
• 5:16
  system hallucinates a nuclear weapon, the consequences
• 5:19
  are unthinkable. It highlights a massive flaw
• 5:22
  in deploying agents without human oversight.
• 5:24
  And the pushback is happening in the legal system
• 5:26
  as well. Internal messages from Microsoft and
• 5:29
  OpenAI just surfaced in a new legal filing. Yeah,
• 5:33
  a science director was quoted calling AI news
• 5:35
  scraping of the largest theft of labor in human
• 5:38
  history. That is a remarkably heavy accusation
• 5:41
  coming from inside the house. It really highlights
• 5:43
  the immense tension right now. We're building
• 5:46
  these powerful autonomous systems, but the data
• 5:49
  they run on. and the real -world spaces they
• 5:51
  operate in are highly contested. Meanwhile, the
• 5:55
  major tech companies are quietly testing the
• 5:57
  next generation of agents in closed environments.
• 5:59
  Like Anthropic, they are reportedly testing Claude
• 6:02
  Fable 5 .2 and Opus 5 .5. The leaked details
• 6:07
  on these tests are just wild. They apparently
• 6:09
  have a mascot walking around a custom world with
• 6:12
  no external assets. The AI is generating and
• 6:15
  reacting to the environment entirely on its own.
• 6:17
  So they're basically building these closed digital
• 6:19
  sandboxes to see what the agents do. Because
• 6:22
  human instincts are failing us out in the open
• 6:24
  web. Have you seen that game, Reality Check?
• 6:26
  Is that the image spotting game where you guess
• 6:28
  what's real? Yeah, that's the one. The Osiris
• 6:30
  source material played it and scored negative
• 6:32
  400 points. Oh, wow. That is incredibly humbling.
• 6:36
  It proves the technology has crossed a threshold
• 6:39
  where our eyes can no longer spot the synthetic
• 6:42
  artifacts. Are we deploying these autonomous
• 6:45
  agents faster than our infrastructure can handle
• 6:48
  them? Without a doubt. I mean, we have agents
• 6:50
  that can navigate the web, reason through complex
• 6:53
  problems and execute tasks. But the platforms
• 6:56
  they interact with are built for human clicks
• 6:58
  and human attention spans. The legal frameworks
• 7:02
  are built for human labor. The friction is everywhere.
• 7:04
  Right. The tech is ready, but the digital world
• 7:07
  refuses to let it in. And because these agents
• 7:09
  are hitting these massive roadblocks, an entirely...
• 7:12
  new ecosystem is rushing in. Yeah, we're seeing
• 7:15
  a flood of tools designed to govern, protect,
• 7:18
  and physically build out this new reality. Let's
• 7:20
  start with a massive breakthrough in the reasoning
• 7:22
  department. OpenAI says its internal math model
• 7:26
  just cracked over 100 open math problems. And
• 7:29
  these are problems human mathematicians haven't
• 7:31
  been able to solve, right? Whoa. Imagine an AI
• 7:34
  solving 100 math problems humans couldn't touch.
• 7:37
  It is a profound shift in capability. They're
• 7:41
  even forming a new advisory group with mathematicians
• 7:43
  to oversee this specific research. Well, math
• 7:47
  has always been the ultimate test for AI. Language
• 7:49
  models are built to predict the next word, which
• 7:51
  makes them great at writing essays. But math
• 7:54
  requires strict sequential logic. If an AI can
• 7:58
  reason through unsolved math, its underlying
• 8:00
  logic engine is incredibly robust. And we are
• 8:03
  seeing an entire tool belt emerge to harness
• 8:06
  these reasoning engines. Let's look at Superset,
• 8:08
  for example. Right. Superset lets you control
• 8:10
  coding agents like Cloud Code or Codex directly
• 8:12
  from your iPhone. So you can review code differences
• 8:15
  and merge pull requests from anywhere. Exactly.
• 8:18
  And just to clarify for you, pull requests are
• 8:20
  submitting proposed code changes for a human
• 8:22
  to review and approve. It makes managing a fleet
• 8:24
  of AI workers as easy as checking your text messages
• 8:27
  on the train. Then there is Jev. built by a company
• 8:30
  called TypeSafe AI. It is what they call a System
• 8:32
  1 frontier model. Let's define that term. A System
• 8:35
  1 frontier model is fast, intuitive AI processing
• 8:39
  that reacts without deep overthinking. The speed
• 8:42
  on Jev is incredible. It is 20 to 200 times faster
• 8:46
  than comparable LLM workflows. The economics
• 8:49
  are what stand out to me. It is 40 to 400 times
• 8:52
  cheaper than its competitors, and the output
• 8:55
  tokens are entirely free. That changes the financial
• 8:58
  model of deploying AI agents overnight. If the
• 9:02
  output is free, you can let an agent run continuously
• 9:04
  without bankrupting your server budget. But with
• 9:06
  all this speed and autonomy, security becomes
• 9:09
  the critical issue. That is where ArcJet comes
• 9:11
  into play. Yeah, ArcJet protects these AI agents
• 9:14
  at runtime. It essentially sits there as a defensive
• 9:17
  layer. It blocks prompt injections, which is
• 9:19
  tricking AI into doing bad things using hidden
• 9:22
  malicious commands. Right. It also redacts sensitive
• 9:24
  data and stops malicious bots before they can
• 9:27
  manipulate your agent. You absolutely need that
• 9:29
  armor if you're letting... AI run loose on your
• 9:32
  network. Especially when you look at a new tool
• 9:34
  like SAI. Oh, SAI is fascinating because it fundamentally
• 9:37
  changes how software interacts. It controls fleets
• 9:41
  of autonomous computers across Windows, Mac OS,
• 9:44
  and Linux. And it automates screen -based tasks
• 9:47
  without needing APIs. Right. Normally, for two
• 9:50
  pieces of software to talk, they use an API.
• 9:53
  It's kind of like a digital backdoor with a specific
• 9:55
  key. But SAI doesn't use the backdoor. Nope.
• 9:58
  It walks right through the front door. It literally
• 10:00
  sees the screen and clicks the mouse just like
• 10:03
  you or I would. That means it can operate any
• 10:05
  software, even legacy systems that don't have
• 10:08
  APIs built in. This autonomy is also moving heavily
• 10:11
  into the physical world. A company called Vantora
• 10:13
  just raised $100 million. Yeah, their entire
• 10:16
  focus is building startups for physical AI. They're
• 10:19
  helping massive industrial giants create private
• 10:22
  AI machines and robotic systems kept entirely
• 10:25
  in -house. So, with tools like Cy bypassing APIs
• 10:28
  entirely, how do we possibly maintain security
• 10:31
  on a network? It requires a completely new security
• 10:34
  paradigm. You can't just revoke an API key anymore.
• 10:37
  You need behavioral defense layers, like ArcChat,
• 10:41
  monitoring what the agent is actually clicking
• 10:43
  on in real time and shutting it down if it deviates
• 10:46
  from its allowed tasks. We are basically building
• 10:48
  an immune system for autonomous computer fleets.
• 10:51
  Let's pause there for a quick break. Sponsor.
• 10:54
  All right, we are back. While companies like
• 10:57
  OpenAI and XAI are building these highly protected
• 11:00
  walled off systems, there is a rebellion happening
• 11:03
  in the background. Yeah, the open source community
• 11:05
  is actively tearing down those walled gardens.
• 11:08
  They want to give this immense agentic power
• 11:10
  to... everyone. And she only just dropped a massive
• 11:13
  bombshell on the industry. The electronics giant
• 11:15
  released MIMO v2 .5. It's an entire family of
• 11:18
  open source AI models. And it is designed specifically
• 11:21
  for reasoning, coding, and long -running multi
• 11:24
  -step agent workflows. The technical specs on
• 11:26
  this are wild. It supports a 1 million token
• 11:29
  context window. A context window is how much
• 11:32
  information the AI can remember and process at
• 11:35
  once. Right. A million tokens is roughly equivalent
• 11:38
  to several thick books of information, all held
• 11:41
  in the AI's active memory at the same time. They
• 11:44
  utilize large -scale reinforcement learning to
• 11:47
  boost the model's planning abilities. They built
• 11:50
  this specifically for software engineering and
• 11:52
  complex tool usage. And the most important part,
• 11:55
  they open -sourced everything. They release the
• 11:57
  model weights, the technical reports, the training
• 12:00
  resources, and the execution frameworks. Yeah,
• 12:02
  they're actively pushing back against the closed
• 12:04
  APIs of Silicon Valley. The lineup features two
• 12:06
  main versions. First, there's MIMO v2 .5 Pro.
• 12:10
  That's their heavy hitter. It focuses on advanced
• 12:12
  coding and highly complex reasoning tasks. The
• 12:16
  documentation says it can maintain coherence
• 12:18
  across thousands of tool calls. Thousands. Let's
• 12:21
  unpack what that actually means. It can search
• 12:22
  a database, write a script, test the script,
• 12:25
  fix a resulting error, email a summary of the
• 12:27
  result over and over thousands of times without
• 12:30
  losing the plot or forgetting its original goal.
• 12:33
  The second version in the lineup is the standard
• 12:35
  MIMO v2 .5. And that one is a general multimodal
• 12:38
  model. It understands text, image, video, and
• 12:41
  audio natively, all integrated with agent capabilities.
• 12:45
  These open source models are moving dangerously
• 12:47
  close to frontier AI capabilities. They really
• 12:50
  are. They're becoming cheaper and infinitely
• 12:52
  easier for everyday developers to customize.
• 12:55
  ShowMe is essentially following the early open
• 12:57
  AI playbook. Give researchers and builders the
• 13:01
  powerful models directly. Right. Don't hide the
• 13:04
  true capabilities behind a corporate paywall
• 13:06
  or a restricted interface. So does open sourcing
• 13:09
  a model capable of thousands of autonomous actions
• 13:12
  present a massive security risk? Oh, it's an
• 13:14
  enormous risk. When you democratize this level
• 13:17
  of reasoning, bad actors get the exact same tools
• 13:20
  as legitimate developers. A closed API lets a
• 13:23
  company pull the plug if a model starts behaving
• 13:25
  maliciously. Open sourcing the weights means
• 13:28
  the model is out in the wild forever. The genie
• 13:31
  is fully out of the bottle. It gives developers
• 13:33
  immense power but removes the safety net of a
• 13:36
  closed API. Exactly. Let's take a step back and
• 13:39
  look at the big picture of what we've covered
• 13:40
  today. We are witnessing the messy adolescence
• 13:43
  of AI. It's graduating from a polite chatbot
• 13:46
  you talk to into an active agent that does things
• 13:49
  on your behalf. We saw it writing and self -verifying
• 13:52
  complex code with Grok 4 .7. We saw it cracking
• 13:56
  untouched math problems. We saw it trying to
• 13:59
  navigate Amazon using Metamuse. And we saw it
• 14:01
  controlling fleets of standard computers with
• 14:03
  tools like Cy, bypassing traditional software
• 14:06
  barriers entirely. But the real world is aggressively
• 14:09
  pushing back against this new autonomy. Yeah,
• 14:12
  the military false alarm shows how catastrophic
• 14:14
  these agents can be when they lack human oversight.
• 14:17
  The Amazon blockade shows how platforms will
• 14:20
  furiously defend their digital territory and
• 14:22
  their business models against automated scrapers.
• 14:25
  Our physical and digital worlds are just resisting
• 14:28
  this shift at every turn. Yet with companies
• 14:30
  like Xiaomi open sourcing these agentic frameworks,
• 14:33
  the momentum seems completely unstoppable. It's
• 14:36
  a profound inflection point in technology history
• 14:38
  for sure. Thank you for joining us on this deep
• 14:41
  dive. We covered some heavy ground today. Remember
• 14:43
  to check the show notes. We'll have links to
• 14:45
  the coding benchmarks, the infrastructure tools,
• 14:47
  and the Xiaomi open source repository. Definitely
• 14:51
  look up that reality check game if you want to
• 14:53
  humble your human instincts. I want to leave
• 14:55
  you with one final thought to Mulliver. We talked
• 14:58
  about Sai controlling computers without needing
• 15:00
  an API, and we talked about open source models
• 15:03
  like Xiaomi's executing thousands of independent
• 15:05
  toolpoles. If an AI like Sai can control a computer
• 15:09
  like a ghost human, and open source models can
• 15:11
  execute thousands of tasks independently, what
• 15:14
  happens when these agents start hiring each other
• 15:16
  to complete complex projects while we sleep?