Latest / AI Fire Daily / 🎙 EP 371: 700+ AI-Solved Math Papers & Mistral Launches 1-Trillion Parameter "Le Chonk"

Transcript

🎙 EP 371: 700+ AI-Solved Math Papers & Mistral Launches 1-Trillion Parameter "Le Chonk" · AI Fire Daily

• 0:00
  Imagine a secret, unreleased frontier AI model.
• 0:04
  Oh, wow. Yeah. It just quietly solved hundreds
• 0:07
  of open math problems, which raises a slightly
• 0:10
  unsettling question for you to consider. What
• 0:12
  happens when machines start doing math faster
• 0:15
  than humans can actually check it? That is the
• 0:17
  big question right now. Welcome to the Deep Dive.
• 0:19
  I am really glad you're here with us today. We
• 0:22
  are going to unpack some massive shifts happening
• 0:24
  right now across the tech ecosystem. We really
• 0:27
  have a lot of fascinating ground to cover today.
• 0:29
  We absolutely do. First, we're looking at OpenAI's
• 0:32
  massive new math data drop. Then we will explore
• 0:36
  the fierce industry power plays happening behind
• 0:38
  closed doors. The compute backlogs alone are
• 0:40
  shifting the entire tech landscape. It's wild.
• 0:43
  Completely. And finally, we'll look at Europe's
• 0:46
  new one trillion parameter heavyweight model
• 0:48
  entering the ring. I can't wait to get into that
• 0:50
  one. Let's start by unpacking this OpenAI release
• 0:52
  first. They just dropped a... Public GitHub repository,
• 0:56
  seemingly out of nowhere. Right, completely out
• 0:58
  of the blue. It contains 722 math manuscripts.
• 1:01
  These are organized across 372 result families,
• 1:05
  and they were all produced by a secret internal
• 1:07
  frontier model. The sheer scale of this evaluation
• 1:11
  is the real headline here. It really is. The
• 1:14
  model was given roughly 4 ,000 open research
• 1:16
  problems. That's a staggering amount of unsolved
• 1:19
  math to just throw at a machine. It's massive.
• 1:23
  And the compute power required to do this is
• 1:25
  just immense. A typical successful result used
• 1:28
  some heavy processing time. It took compute equivalent
• 1:31
  to about three hours of ChatGPT Pro thinking
• 1:35
  time. Three hours of continuous compute for a
• 1:38
  single math problem. Yeah, just grinding away
• 1:41
  at it. That's incredibly intense. I mean, most
• 1:43
  of us get annoyed if a web page takes three seconds
• 1:45
  to load. Imagine an AI grinding on a single logic
• 1:49
  puzzle for three hours straight. Exactly. It's
• 1:52
  essentially traversing massive search trees.
• 1:54
  It's guessing, checking, and refining its logic
• 1:57
  over and over. And the sources mention many of
• 2:00
  these proofs have lean formalizations backing
• 2:02
  them up. We should break that jargon down for
• 2:04
  you. Basically, lean formalizations are code
• 2:07
  that mechanically checks math steps for absolute
• 2:10
  truth. Right. It removes human error from the
• 2:13
  verification process entirely. Which is pretty
• 2:15
  crucial when the math gets this dense. Oh, absolutely.
• 2:18
  They also published 10 abridged reasoning summaries
• 2:21
  alongside the data. This shows exactly how the
• 2:25
  model approached selected problems step by step.
• 2:28
  It's kind of like a math teacher forcing the
• 2:30
  AI to show its work on a test. That is a great
• 2:33
  way to look at it. Yeah. But OpenAI... Explicitly
• 2:37
  notes these manuscripts are at very different
• 2:39
  stages. Many are formally checked using that
• 2:42
  lean software we just mentioned, but some are
• 2:45
  absolutely not. So they're just raw, unverified
• 2:49
  outputs? They warned that unformalized results
• 2:52
  might still contain serious errors. That makes
• 2:55
  sense. Hallucinations in high -level math can
• 2:57
  look very convincing to the untrained eye. Which
• 3:00
  brings us to the controversy surrounding this
• 3:02
  huge release. Some researchers are really pushing
• 3:05
  back on this strategy. Yeah, there is a lot of
• 3:08
  pushback. They argue that using open problems
• 3:10
  as AI benchmarks is highly risky. The fear is
• 3:13
  that we are flooding the zone. Right. Flooding
• 3:16
  the zone with unverified, highly complex math.
• 3:19
  Exactly. We're overwhelming the academic community
• 3:21
  with AI -generated results. Because academia
• 3:24
  relies on human review? Right. It relies on grad
• 3:27
  students and professors painstakingly reading
• 3:30
  proofs. Sometimes that takes months for a single
• 3:33
  paper. Wow. They just can't absorb or verify
• 3:36
  these machine outputs fast enough. Which makes
• 3:38
  me wonder about the structural limits of our
• 3:39
  institutions. If this machine generates thousands
• 3:42
  of hours of complex math, how does human peer
• 3:45
  review possibly keep up with that bottleneck?
• 3:47
  Honestly, it completely fractures the traditional
• 3:50
  academic system. Deeply meticulous. It is notoriously,
• 3:55
  historically slow by design. Right. You can't
• 3:58
  rush a mathematical proof. Exactly. So when a
• 4:01
  machine dumps hundreds of dense proofs online
• 4:03
  overnight, the human verification engine just
• 4:06
  stalls out completely. It breaks the model of
• 4:09
  how science is currently validated. So the AI
• 4:12
  generates proofs faster than human peer review
• 4:15
  can keep up. Right. And that brings us to the
• 4:18
  physical infrastructure required for this new
• 4:20
  reality. Running a frontier model for three hours
• 4:23
  straight to solve one problem demands serious
• 4:25
  hardware. It takes a mind -bending amount of
• 4:28
  compute. It really does. Which actually explains
• 4:30
  the absolute panic we're seeing in the infrastructure
• 4:33
  space right now. We're seeing OpenAI flex its
• 4:36
  frontier capabilities with this math drop. Meanwhile,
• 4:39
  the rest of the tech ecosystem is aggressively
• 4:41
  maneuvering behind the scenes. Everyone is fighting
• 4:43
  for computing power and market share right now.
• 4:45
  It's basically an arms race. It is. The capital
• 4:49
  flowing into digital infrastructure right now
• 4:51
  is just staggering. Take Lambda, for example.
• 4:54
  They're an AI computing startup, right? Yeah.
• 4:57
  They rent out server space packed with high -end
• 4:59
  GPUs. They are currently raising up to $4 billion
• 5:03
  in new funding. Wow. $4 billion. Yeah. That's
• 5:07
  a massive market -defining funding round. It
• 5:10
  puts them at a $14 .5 billion valuation. They
• 5:13
  are already planning for a massive IPO in 2027.
• 5:16
  That is moving so fast. But here is where the
• 5:19
  numbers get truly wild. Their backlog just exploded
• 5:22
  from $15 billion to $50 billion. Wait, $50 billion
• 5:27
  in backlogged compute demand? Yes. And that massive
• 5:31
  jump is heavily driven by Anthropic trying to
• 5:33
  secure servers. That's a $50 billion waiting
• 5:37
  list just to plug in computers. Pretty much.
• 5:39
  It really shows how compute is basically the
• 5:41
  new oil. It also shows how utterly cutthroat
• 5:44
  this space is becoming. The alliances are shifting
• 5:47
  rapidly. We're seeing some intense partner -to
• 5:50
  -competitor shifts happening almost overnight.
• 5:53
  Meta and Microsoft are reportedly cutting their
• 5:55
  internal cloud usage. Yeah, they're dropping
• 5:57
  external tools to push their own competing products
• 5:59
  internally. So they're freezing out their former
• 6:02
  partners. Exactly. Internal spending on early
• 6:04
  partners is dropping incredibly fast. It makes
• 6:07
  sense, though. In the early days, they all needed
• 6:09
  each other to bootstrap. Right. They had to collaborate
• 6:12
  to build the foundation. Now that they have their
• 6:14
  own data flywheel spinning, they are cutting
• 6:16
  the cord. It's a very messy transition for the
• 6:18
  whole industry. Speaking of messy, we should
• 6:21
  really look at Sam Altman's recent stance. Right.
• 6:23
  He's making some very blunt statements about
• 6:25
  the realities of this rollout. He basically says
• 6:27
  expecting zero AI harm is totally unrealistic.
• 6:30
  Which is a tough pill to swallow for a lot of
• 6:32
  regulators. It is. It's controversial, but it's
• 6:35
  pragmatic. The surface area of AI application
• 6:38
  is just too vast to police perfectly. But he
• 6:41
  also admits that one company controlling everything
• 6:43
  is deeply dangerous. Yeah, he was very clear
• 6:45
  about that. It's a very thought -provoking position.
• 6:48
  He's essentially saying mistakes will happen,
• 6:50
  but monopolies are the real existential threat.
• 6:53
  Definitely. The ecosystem dynamics are shifting
• 6:55
  so fast, it's hard to track. It really is. And
• 6:58
  amidst all this high -level corporate drama,
• 7:00
  the daily tools are just getting better. Right,
• 7:03
  the stuff developers actually use. Yeah, we are
• 7:05
  seeing AI integrate much deeper into standard
• 7:07
  developer workflows. Google just launched Nano
• 7:10
  Banana 2 .1, right? Yeah. The updates on that
• 7:13
  tool are actually quite pragmatic. It features
• 7:16
  native 4K output now. That's a big jump in resolution.
• 7:19
  It is. It also has much stronger image editing
• 7:21
  capabilities built right in. And it handles a
• 7:23
  lot more context visually. It supports up to
• 7:26
  14 reference images now. Exactly. And developer
• 7:30
  pricing dropped by roughly 50%. Wow. It's just
• 7:33
  $0 .034 per image. That is incredibly cheap for
• 7:37
  high -quality Gemini consistency. That kind of
• 7:39
  pricing totally changes what developers can build.
• 7:41
  It democratizes the access completely. But as
• 7:44
  things get cheaper and more autonomous, security
• 7:46
  becomes a massive headache. We're seeing new
• 7:49
  accountability tools like Agent ID roll out to
• 7:51
  counter this. Agent ID is a really fascinating
• 7:54
  development. It lets AI agents sign into apps
• 7:57
  with a verified email. But it ties that email
• 7:59
  back to a human owner. It's built entirely for
• 8:02
  trust, control, and corporate accountability.
• 8:04
  Which is crucial as these agents get more autonomous.
• 8:07
  Right. You can't just have rogue code booking
• 8:09
  flights and signing contracts anonymously. Honestly,
• 8:12
  delegating real -world tasks to AI is still tough
• 8:15
  for me. I still wrestle with prompt drift myself.
• 8:17
  Letting an agent run wild with an email tied
• 8:19
  to my identity requires so much oversight. It's
• 8:22
  a very common struggle. But the platform dominance
• 8:24
  race is driving these tools forward regardless
• 8:27
  of our comfort levels. Which raises an important
• 8:29
  question for you to consider about this corporate
• 8:31
  battlefield. Why are these massive tech giants
• 8:34
  suddenly treating their former partners as rivals?
• 8:37
  It's a classic battle for platform dominance.
• 8:39
  In the early days, they desperately needed each
• 8:41
  other to scale the raw technology. Now the baseline
• 8:44
  reasoning capabilities are essentially becoming
• 8:46
  a commodity. Whoever controls the daily enterprise
• 8:49
  workflow ultimately wins the entire market. So
• 8:53
  they have to own the user. So they're severing
• 8:55
  ties to capture their own users entirely. Right.
• 8:58
  Yesterday's essential partners are becoming today's
• 9:00
  fiercest competitors overnight. Exactly. Today's
• 9:03
  Deep Dive is brought to you by our dedicated
• 9:05
  partners. They help us keep the research flowing
• 9:07
  and the servers humming. We appreciate their
• 9:10
  support in bringing this detailed analysis to
• 9:12
  you. We really do. That ruthless battle for U
• 9:15
  .S. workflow dominance is highly relevant to
• 9:17
  our final topic. The American tech giants are
• 9:19
  aggressively fighting over compute and monopolies,
• 9:22
  and that anxiety is exactly why we're seeing
• 9:25
  a massive openweight counterpunch from Europe.
• 9:28
  I am so excited to talk about this specific release.
• 9:32
  Mistral just launched Mistral Large 4. The big
• 9:35
  one. Yeah. It's a brand new multimodal model,
• 9:38
  and the community has affectionately nicknamed
• 9:40
  it LeChunk. I absolutely love that nickname.
• 9:44
  It's very fitting for what they've built. It
• 9:46
  really is. Let's detail the sheer size of this
• 9:48
  new model. It is massive. It has roughly one
• 9:51
  trillion parameters. That is a staggering scale
• 9:55
  for an open -weight model. We should probably
• 9:57
  define that jargon for you really quickly. Think
• 9:59
  of parameters as the digital brain connections
• 10:01
  that help the model make decisions. That's exactly
• 10:03
  right. A trillion of those connections is wild.
• 10:06
  Think of parameters like the dials on a massive
• 10:08
  studio soundboard. A small model has a thousand
• 10:11
  dials to tune the audio. Right. A trillion parameter
• 10:13
  model has a trillion tiny dials. It lets the
• 10:17
  model isolate and perfect the tiniest frequency
• 10:19
  of nuance in your request. It allows for incredible
• 10:22
  precision. But building a brand with a trillion
• 10:24
  dials usually takes endless resources. Which
• 10:28
  is why Mistral's hardware efficiency here is
• 10:30
  what really stands out. Yeah. Their training
• 10:32
  footprint is... Genuinely impressive. It was
• 10:35
  trained entirely on Mistral's own compute infrastructure.
• 10:37
  Okay. And they used just 4 ,000 NVIDIA GPUs to
• 10:40
  do it. Just 4 ,000 GPUs to train a one trillion
• 10:44
  parameter model. That is incredibly lean. It
• 10:47
  really is. Mistral says it used two to three
• 10:50
  times fewer GPUs than their Chinese competitors.
• 10:53
  Wow. And it used substantially fewer resources
• 10:55
  than the major U .S. closed model apps. They
• 10:58
  are squeezing every drop of performance out of
• 11:00
  that silicon. Wow. Imagine scaling to a billion
• 11:03
  queries. Doing that with such a tight footprint
• 11:06
  is wild. It really is. So how exactly are they
• 11:10
  pulling that off? You don't just magically train
• 11:12
  a trillion parameters on a shoestring hardware
• 11:14
  budget. It comes down to brilliant architectural
• 11:17
  efficiency. They aren't activating the entire
• 11:20
  trillion parameter brain for every single word
• 11:22
  it generates. Oh, I see. They use routing techniques
• 11:25
  to only trigger the specific expert pathways
• 11:27
  needed for a given task. It's not just a mess
• 11:30
  of information. It's like stacking Lego blocks
• 11:33
  of data. Exactly. It saves massive amounts of
• 11:36
  compute while maintaining high -level reasoning.
• 11:39
  That is deeply pragmatic engineering. So what
• 11:43
  is this model actually optimized to do out in
• 11:45
  the wild? They optimized it heavily for high
• 11:48
  -stakes enterprise workloads. It excels in areas
• 11:50
  like complex cybersecurity, dense finance, and
• 11:54
  intricate chip design. The hard stuff. Exactly.
• 11:57
  Early independent testing is already showing
• 11:59
  very strong results. That's promising. It currently
• 12:02
  places ML4 around sixth overall among open weight
• 12:05
  models on artificial analysis. Its cybersecurity
• 12:09
  performance appears especially robust in those
• 12:11
  early benchmarks. I'm looking at their release
• 12:13
  strategy, and it seems unusually guarded for
• 12:16
  an open weight company. How are they rolling
• 12:18
  this out? It's a very staged approach. Currently,
• 12:21
  it's only available through a heavily guarded
• 12:23
  public endpoint. So developers can use it, but
• 12:26
  they can't download it yet. Right. But Mistral
• 12:29
  plans to release the actual model weights soon.
• 12:31
  Okay, good. They are doing roughly three weeks
• 12:34
  of additional rigorous safety testing first.
• 12:37
  Mistral is making a massive geopolitical and
• 12:40
  corporate bet here. They clearly believe a highly
• 12:42
  capable open weight European model is absolutely
• 12:45
  necessary right now. It provides a critical safety
• 12:48
  valve for the global market. It gives enterprises
• 12:50
  and sovereign governments another crucial option.
• 12:53
  They don't have to rely entirely on U .S. closed
• 12:55
  models. Right. And they don't have to deploy
• 12:57
  Chinese open models either. It's a third path.
• 13:01
  ML4 looks like a very solid frontier level of
• 13:04
  release. It proves Europe can still compete at
• 13:06
  the absolute bleeding edge. But the market is
• 13:08
  moving incredibly fast right now. The big question
• 13:10
  is whether Mistral can keep shipping these upgrades
• 13:13
  quickly. They have to maintain this grueling
• 13:15
  pace to stay relevant long term. So is extreme
• 13:19
  compute efficiency actually enough for an open
• 13:22
  weight model to survive against trillion dollar
• 13:24
  U .S. monopolies? Efficiency alone won't win
• 13:27
  the war, but it definitely helps. It allows Mistral
• 13:29
  to offer enterprise clients much cheaper deployment
• 13:32
  costs. Which is a huge advantage. It is. However,
• 13:35
  those U .S. monopolies have virtually unlimited
• 13:37
  capital and hardware access. Mistral has to rely
• 13:41
  on continuous rapid innovation to outrun that
• 13:43
  brute force. So being incredibly efficient buys
• 13:46
  them time, but the race never stops. Let's take
• 13:49
  a step back and synthesize everything we've covered
• 13:51
  in this deep dive. We are witnessing a truly
• 13:54
  pivotal moment in technology right now. We really
• 13:58
  are. The baseline capabilities are just exploding
• 14:00
  across the board. AI is solving high -level,
• 14:03
  open math problems autonomously. We're seeing
• 14:06
  trillion -parameter open models launching successfully
• 14:09
  from Europe. But everything is scaling so unbelievably
• 14:12
  fast that our legacy systems are struggling to
• 14:15
  adapt. They really are. Human verification processes
• 14:17
  in academia are breaking down entirely, and traditional
• 14:21
  corporate alliances are fracturing under the
• 14:23
  immense pressure of compute demands. The friction
• 14:25
  between rapid AI scaling and human oversight
• 14:28
  is massive right now. We are essentially building
• 14:31
  rocket engines much faster than we can build
• 14:33
  the brakes. Which leaves us with a final lingering
• 14:36
  thought for you to ponder today. We talked about
• 14:39
  AI solving open math problems earlier. Right,
• 14:41
  the open AI release. If an AI model eventually
• 14:44
  solves a universe -altering physics or math problem,
• 14:47
  but the proof is so complex no human can verify
• 14:49
  it, does it actually count as human progress?
• 14:52
  Thank you so much for joining us on this deep
• 14:54
  dive. We'll see you next time.