Latest / AI Fire Daily / 🎙️ EP 325: Alibaba Releases Qwen3.8-Max & Chinese AI Scores Perfect on Math Olympiad

Transcript

🎙️ EP 325: Alibaba Releases Qwen3.8-Max & Chinese AI Scores Perfect on Math Olympiad · AI Fire Daily

• 0:00
  Imagine having $35 billion just evaporate into
• 0:04
  thin air in less time than it takes to get a
• 0:07
  package delivered from Amazon. I mean, that is
• 0:09
  exactly what happened to one of Wall Street's
• 0:10
  most hyped AI hedge funds this week. Right. But
• 0:14
  here is the wild paradox. On the exact same day
• 0:17
  human investors were, you know, completely losing
• 0:20
  their minds, a mysterious AI was quietly sitting
• 0:23
  in a digital room, flawlessly passing the hardest
• 0:27
  math test on planet Earth. It really is a week
• 0:30
  of absolute extremes. We're watching the chaotic
• 0:33
  frontier of our new digital reality unfold in
• 0:36
  real time. Welcome to the Deep Dive. I'm really
• 0:38
  glad you're here with us. We have a massive agenda
• 0:41
  today and we're unpacking a lot of shifting tectonic
• 0:43
  plates. Yeah, there's a ton of ground to cover.
• 0:45
  So we're going to start by looking at a sprawling
• 0:47
  new open weights model from Alibaba. Then we'll
• 0:51
  navigate the very chaotic collision of AI with
• 0:54
  real world markets and real world laws. Which
• 0:56
  is getting super messy. It really is. We'll also
• 0:59
  check out some incredibly practical desktop tools
• 1:01
  you can actually use right now. Yeah. And finally,
• 1:03
  we're examining a true milestone in machine reasoning.
• 1:07
  It feels like the industry is hitting a massive
• 1:09
  inflection point today. It really does. So let's
• 1:12
  just jump in and start with the heavyweight contender
• 1:14
  in the room. Alibaba just officially released
• 1:17
  their new model. It's called Quen 3 .8 Max. Yeah,
• 1:21
  we got a little sneak preview of this back on
• 1:23
  July 19th, but the official OpenWeets release
• 1:26
  is finally upon us. And they are confidently
• 1:29
  calling this their most capable Quinn model yet.
• 1:33
  We're just saying a whole lot concerning how
• 1:35
  aggressive their release schedule has been lately.
• 1:36
  Yeah. This specific model focuses heavily on
• 1:40
  coding. agentic workflows, and long multi -step
• 1:44
  tasks. I mean, the raw numbers driving this architecture
• 1:47
  are just completely staggering. Yeah, let's unpack
• 1:50
  that architecture a bit. We're looking at 2 .4
• 1:52
  trillion total parameters under the hood. Wow.
• 1:55
  And let's quickly pause and define that term
• 1:58
  for anyone new to the space. Parameters are basically
• 2:01
  the digital brain cells that determine how smart
• 2:03
  the model is. Right, and that is a massive digital
• 2:06
  brain to try and wrap your head around. But it's
• 2:08
  not just the sheer size of the network. The memory
• 2:11
  capacity is equally difficult to comprehend.
• 2:14
  It boasts a one million token context window
• 2:18
  for processing data. Right. And a one million
• 2:21
  token context window is like loading a towering
• 2:24
  stack of encyclopedias into your short term memory
• 2:27
  all at once. That's a great way to picture it.
• 2:29
  You aren't just scanning the pages. You are holding
• 2:30
  all of that information in active focus simultaneously.
• 2:34
  Exactly. And practically, this means it supports
• 2:37
  incredibly long reasoning sessions. If you're
• 2:39
  a developer, you can build massive interconnected
• 2:41
  workflows without the AI suddenly losing the
• 2:44
  plot. Which is huge. It is. It's available right
• 2:47
  now through Alibaba Cloud Model Studio. But developers
• 2:50
  can also access it through QuinCode, Coder, and
• 2:53
  CoderWork. They really prioritize making it easy
• 2:56
  to plug into existing ecosystems. Probably. Like
• 2:58
  it natively supports open AI and anthropic compatible
• 3:01
  APIs. It also works seamlessly with Dashscope.
• 3:05
  Yeah, they want enterprise users to adopt this
• 3:07
  without ripping out their current infrastructure.
• 3:09
  And the pricing strategy reflects that aggressive
• 3:11
  push for adoption, too. They have a promotional
• 3:14
  international token plan starting at just $6
• 3:17
  a month. Wait, really? $6? Yeah, but keep in
• 3:21
  mind, that's a limited promotional rate. The
• 3:22
  standard pro plan sits at $68 per month. Okay,
• 3:25
  that makes more sense. And Alibaba claims this
• 3:28
  model ranks just behind CloudFable 5 globally.
• 3:31
  That puts it squarely in the absolute top tier
• 3:34
  of AI capability. Definitely. But I want to dig
• 3:37
  into a specific feature they're highlighting.
• 3:39
  It uses something called persistent reasoning
• 3:41
  by default. You can configure the API for different
• 3:45
  effort levels, but it defaults to an X high effort
• 3:49
  out of the box. So how does persistent reasoning
• 3:52
  change the way the model handles long coding
• 3:54
  sessions compared to older models? Well, older
• 3:57
  systems suffered from a kind of digital amnesia.
• 4:00
  They would execute one prompt, output some code,
• 4:03
  and then completely forget the broader context.
• 4:05
  Oh, wait. But persistent reasoning creates a
• 4:07
  continuous, unbroken chain of active thought.
• 4:10
  Oh, wow. The model holds the final objective
• 4:13
  firmly in its digital mind. It checks each new
• 4:16
  line of code against the broader software architecture,
• 4:20
  constantly evaluates its own ongoing progress
• 4:22
  in real time. Without needing you to remind it
• 4:26
  what it's doing. Exactly. So it never loses its
• 4:28
  train of thought during complex tasks. It stays
• 4:31
  perfectly locked in from start to finish. Well,
• 4:33
  with models getting this massive incapable...
• 4:36
  The landscape is inherently shifting. Oh, absolutely.
• 4:39
  The immediate ripple effect is society frantically
• 4:42
  trying to put guardrails up. And we're seeing
• 4:44
  this friction play out very clearly in the cybersecurity
• 4:47
  space right now. Yeah. Safety institutes in the
• 4:50
  UK and the US recently ran some really extensive
• 4:52
  tests. They took a model called Kimi K3. and
• 4:56
  tested it against top U .S. models. To see if
• 4:58
  these systems could be weaponized by bad actors,
• 5:00
  right? Exactly. And the test results were honestly
• 5:02
  a bit unsettling to read. They found that Kimi
• 5:04
  K3 can complete meaningful parts of a sophisticated
• 5:08
  cyber attack. Like it can actively scan for vulnerabilities
• 5:11
  and draft targeted exploits. But thankfully,
• 5:14
  the report noted, it still trails the top U .S.
• 5:18
  models badly in this specific domain. It's capable,
• 5:21
  but it's not quite a fully autonomous hacking
• 5:24
  machine yet. Right. However, the mere existence
• 5:26
  of that exact threat is pushing lawmakers into
• 5:29
  immediate action. For sure. U .S. lawmakers just
• 5:32
  introduced the AI Killswitch Act. This legislation
• 5:35
  could let the Department of Homeland Security
• 5:37
  restrict dangerous systems entirely. Right. They
• 5:40
  could potentially shut them down completely in
• 5:42
  extreme cases. But the text of the bill specifies
• 5:45
  this only applies after a truly catastrophic
• 5:48
  AI failure. That specific phrasing is causing
• 5:53
  a lot of intense debate in legal and tech circles
• 5:56
  right now. That's the part I keep getting stuck
• 5:58
  on, actually. If we wait for a truly catastrophic
• 6:01
  failure before flipping the switch, isn't the
• 6:03
  damage already done? That's the core paradox
• 6:05
  of reactive legislation. Right. I mean, by definition,
• 6:08
  a catastrophe means the system has already caused
• 6:10
  massive harm. You don't want to stifle innovation
• 6:13
  prematurely, but you also can't unring a bell.
• 6:15
  Over in Minnesota, we're seeing state -level
• 6:18
  action taking a much more immediate approach.
• 6:20
  They enacted the first U .S. ban on deepfake
• 6:24
  nudify apps, and it is officially active right
• 6:27
  now. And we should mention that XAI filed an
• 6:29
  emergency request to block that law. Right. The
• 6:32
  company argued that the law is simply too broad
• 6:34
  and could restrict protected speech. But a judge
• 6:37
  recently denied that emergency request. So the
• 6:40
  ban stands for now. It really highlights how
• 6:42
  messy the collision between rapid technological
• 6:45
  advancement and slow moving legal systems can
• 6:48
  be. It absolutely does. But going back to the
• 6:50
  cybersecurity testing for a second. Does the
• 6:52
  Kimi K3 test prove safety institutes are successfully
• 6:56
  predicting threats? I think it shows they're
• 6:58
  finally building proactive testing frameworks
• 7:00
  instead of just reacting. Right. They're simulating
• 7:02
  worst case scenarios in tightly controlled laboratory
• 7:05
  environments. This prevents malicious actors
• 7:08
  from finding those dangerous exploits first.
• 7:10
  By forcing the model to play the role of an attacker,
• 7:13
  we learn its boundaries. We uncover hidden vulnerabilities
• 7:16
  in our own infrastructure before they ever reach
• 7:19
  the general public. So catching these capabilities.
• 7:21
  early gives us a crucial head start on defense.
• 7:24
  Exactly. It's an arms race. And proactive testing
• 7:27
  is our best radar. Man, that exact tension between
• 7:31
  AI's wild, unchecked potential and real world
• 7:34
  regulation is explosive. It is causing absolute
• 7:38
  chaos in the financial markets right now. Oh,
• 7:40
  Wall Street is experiencing a brutal historic
• 7:42
  reality check. Yeah. Leopold Aschenbrenner's
• 7:45
  massive AI hedge fund completely collapsed this
• 7:49
  week. It is hard to overstate the scale of this
• 7:51
  implosion. The fund crashed from $45 billion
• 7:54
  in managed assets down to just $10 billion. It's
• 7:58
  just wild. And that entire collapse happened
• 7:59
  in just a few days. Yeah, the underlying mechanics
• 8:02
  of the crash are terrifying. Their leverage reportedly
• 8:05
  reached an absolutely mind -boggling 400%. Wow.
• 8:09
  That is an unfathomable amount of financial exposure.
• 8:11
  What that means is they were borrowing massive
• 8:14
  amounts of money to amplify their bets on AI
• 8:17
  companies. It's like stacking Lego blocks of
• 8:20
  massive debt on a wobbly table of chip stocks.
• 8:23
  That is exactly what it is. And eventually that
• 8:26
  wobbly table collapses under the immense weight.
• 8:28
  Right. A slight drop in chip stocks triggered
• 8:31
  massive margin calls from their brokers. And
• 8:34
  a margin call basically means the broker demands
• 8:36
  immediate cash to cover. the dipping asset values.
• 8:39
  Right. They want their money. Exactly. And because
• 8:42
  the fund was so heavily leveraged, they just
• 8:44
  didn't have the cash. Which forced a desperate,
• 8:47
  heavily discounted sell -off across the board.
• 8:49
  They had to liquidate everything at fire sale
• 8:51
  prices just to survive. And you know, while the
• 8:53
  financial sectors are burning down, the consumer
• 8:55
  product side just keeps casually dropping updates.
• 8:58
  Whoa, the pace of these product drops is relentless.
• 9:00
  It is. Grok Imagine received a massive feature
• 9:03
  upgrade today. They now feature voice consistency
• 9:06
  and powerful multi -references for character
• 9:08
  generation. Yeah, and they also added text -to
• 9:11
  -video capabilities and native 1080p generation
• 9:14
  support. Google is not sitting still either.
• 9:17
  They are currently testing a brand new Gemini
• 9:20
  desktop app. Yeah, it has dedicated streamlined
• 9:22
  tabs for image and video generation. It also
• 9:26
  includes native camera capture and custom MCP
• 9:30
  server support. Which is great for developers
• 9:32
  who want to integrate specific local data contexts.
• 9:35
  Absolutely. But nobody actually knows when Google
• 9:37
  will launch it publicly. Typical Google. And
• 9:39
  then there's this strange, somewhat dystopian
• 9:42
  cultural side of AI developments. Oh, boy. Sam
• 9:45
  Altman made a very controversial suggestion about
• 9:47
  parenting this week. Right. He suggested that
• 9:50
  ChatGPT should make a morning podcast summarizing
• 9:53
  your kids' chaotic lives. Yeah. Essentially an
• 9:55
  AI briefing to catch you up on what your children
• 9:58
  have been doing. And the backlash was immediate.
• 10:00
  Oh yeah. One brutal reply on social media went
• 10:03
  completely viral. They simply asked What if you
• 10:06
  just talk to your children? Yeah, that's a fair
• 10:09
  question. It hits a very raw nerve for a lot
• 10:11
  of people, myself included. I still wrestle with
• 10:14
  trusting AI to summarize my kids' lives rather
• 10:17
  than just being present myself. I completely
• 10:19
  get that. There is a deeply human element to
• 10:23
  messy parenting that we shouldn't necessarily
• 10:24
  try to optimize away. No, it's a deeply human
• 10:27
  dilemma we're all facing today. Where do we draw
• 10:30
  the line between convenience and connection?
• 10:32
  It's a tough line to find. But circling back
• 10:36
  to that massive hedge fund collapse for a moment.
• 10:38
  Okay. What does 400 % leverage actually tell
• 10:41
  us about the psychological state of AI investors
• 10:43
  right now? Honestly, it reveals a highly toxic
• 10:46
  mixture of extreme greed and blind panic. Yeah.
• 10:50
  Investors are borrowing heavily to chase completely
• 10:53
  impossible exponential returns. They've convinced
• 10:56
  themselves that the AI boom guarantees endless
• 10:59
  uninterrupted market growth. Right. They think
• 11:02
  if they don't bet everything right now, they'll
• 11:03
  miss the revolution. When reality finally hits,
• 11:06
  that overconfidence shatters into instant catastrophic
• 11:09
  liquidation. So investors are terrified of missing
• 11:12
  out on the next big leap. Fear of missing out
• 11:15
  is literally driving billions of dollars over
• 11:17
  a cliff. Before we move on to how normal people
• 11:19
  are using AI on their computers, let's take a
• 11:22
  quick second to thank today's sponsor. Yeah.
• 11:24
  If you're listening to this deep dive, you probably
• 11:26
  know how impossible it is to keep up with the
• 11:29
  endless flood of AI research and news every single
• 11:31
  day. Exactly. It's overwhelming. That's why we
• 11:34
  use TechFlow. TechFlow is a free newsletter that
• 11:37
  sends a daily five -minute breakdown of the most
• 11:39
  important tech news directly to your inbox. It's
• 11:42
  honestly the best way to stay in the loop without
• 11:44
  spending hours scrolling on social media. Yeah,
• 11:46
  you can check them out and sign up for free at
• 11:48
  techflow .com. Highly recommend it. So while
• 11:51
  Wall Street gambles billions on the distant future
• 11:54
  of artificial general intelligence, regular developers
• 11:57
  are quietly releasing incredibly practical tools
• 12:00
  today. Right. These are pragmatic tools you can
• 12:02
  run on your desktop right now. We're seeing a
• 12:05
  major shift in the industry towards specialized
• 12:07
  AI teammates. People are moving away from the
• 12:11
  omniscient chatbots toward highly... focused,
• 12:14
  narrow agents. Take Stanley, for example. It's
• 12:17
  a brand new AI tool that basically acts like
• 12:20
  an AI head of content. It's fascinating. It helps
• 12:23
  you research the latest trending topics online,
• 12:25
  synthesizing hours of reading into actionable
• 12:28
  insight. Yeah. And then it writes viral posts
• 12:30
  tailored to your exact brand voice. It even schedules
• 12:34
  those posts across all your social platforms
• 12:36
  automatically. That's incredibly useful. Then
• 12:38
  there's an impressive personal agent known as
• 12:40
  Zinley. Zinley acts as your dedicated personal
• 12:43
  AI representative. Right. It seamlessly answers
• 12:46
  incoming calls and handles your chaotic email
• 12:49
  inbox. It completes tedious administrative tasks
• 12:52
  directly on your behalf so you can actually focus
• 12:54
  on deep work. Which we all need more time for.
• 12:57
  Definitely. There's also a powerful tool called
• 13:00
  Captivo available now. Yeah, Captivo is a completely
• 13:03
  free AI -powered screen recorder. It helps you
• 13:06
  record. edit and instantly add perfectly timed
• 13:09
  captions to your videos nice you can create highly
• 13:11
  polished product demos without being locked into
• 13:14
  any recurring monthly subscriptions your site
• 13:16
  is another fascinating tool targeting digital
• 13:18
  creators today it turns all your scattered links
• 13:21
  into a beautiful cohesive bio page oh that's
• 13:24
  cool it uses ai import features to build the
• 13:27
  site in seconds and it provides deep performance
• 13:29
  analytics on the back end and finally There's
• 13:32
  a highly capable coding tool called LumiChats.
• 13:35
  This one is really exciting for developers. Yeah?
• 13:38
  LumiChats brings advanced AI coding directly
• 13:41
  to your desktop environment. It doesn't just
• 13:43
  give you code snippets to copy and paste, right?
• 13:45
  No, it lets the AI actively edit your local files
• 13:48
  directly. It can run automated commands and do
• 13:51
  real developer work right on your machine. And
• 13:54
  the best part is, it does all this without you
• 13:55
  ever needing to open the terminal. Oh, let's
• 13:58
  quickly define that for our non -technical listeners.
• 14:00
  Yeah, the terminal is basically the scary black
• 14:02
  screen where coders type raw text commands. Exactly.
• 14:06
  LumiChat's completely removed that intimidating
• 14:09
  technical barrier to entry. You interact in plain
• 14:12
  English, and it handles the scary command line
• 14:14
  stuff silently in the background. Which brings
• 14:16
  up an interesting architectural debate. Is a
• 14:19
  collection of these specialized mini agents actually
• 14:21
  more useful than one giant generalized chat bot?
• 14:25
  In practical everyday workflows. Absolutely.
• 14:28
  Yeah. Generalized chat bots often hallucinate
• 14:31
  when switching between wildly diverse tasks because
• 14:34
  their context gets polluted. Many agents have
• 14:37
  highly restricted boundaries and very specific
• 14:40
  crystal clear goals. Right. They don't get distracted.
• 14:43
  They don't get confused by irrelevant contextual
• 14:46
  data floating around. You string these narrow
• 14:48
  agents together to build incredibly reliable
• 14:51
  automated workflows that actually function without
• 14:54
  constant human supervision. So niche tools win
• 14:58
  because they do one specific job flawlessly.
• 15:00
  Exactly. They aren't trying to be everything
• 15:02
  to everyone. Speaking of doing a job completely
• 15:04
  flawlessly, let's pivot to what might be the
• 15:07
  most incredible story of the week. Yeah, let's
• 15:09
  look at the ultimate test of pure human logic.
• 15:12
  We're talking about the International Math Olympiad
• 15:14
  here. The IMO is legendary. It's where the brightest
• 15:17
  young mathematical minds on the planet go to
• 15:19
  solve problems that would make most PhDs cry.
• 15:22
  Seriously. And a Chinese AI called Red Note just
• 15:25
  made absolute history there. Yeah. Red Note's
• 15:28
  beta system called Dots Note 3 .0 scored a perfect
• 15:31
  42 out of 42. This is on the incredibly difficult
• 15:34
  IMO 2026 problem set. Let's give you some vital
• 15:37
  context on just how hard this achievement is.
• 15:40
  Please do. Out of 666 elite human contestants,
• 15:44
  only seven managed to match this perfect score.
• 15:47
  Wow. The official gold medal cutoff for humans
• 15:51
  was a mere 29 points. Yeah. This AI completely
• 15:56
  shattered the highest human mathematical benchmarks
• 15:58
  we have. It didn't just guess multiple choice
• 16:00
  answers either. It solved all six problems by
• 16:04
  generating complete rigorous written proofs.
• 16:06
  And it read the original contest documents completely
• 16:09
  unassisted. There were zero human hints. no simplified
• 16:12
  prompts, and no rewritten questions provided
• 16:15
  to the model. That's wild. The methodology behind
• 16:17
  how it solved these problems is fascinating too.
• 16:20
  Yeah, it combined natural language reasoning
• 16:21
  with direct Python code execution. It relentlessly
• 16:25
  tested edge cases and constantly challenged its
• 16:27
  own underlying assumptions. When it found a flaw
• 16:29
  in its logic, it revised weak arguments through
• 16:32
  repeated rigorous self -review loops. It basically
• 16:35
  argued with itself until the logic was bulletproof.
• 16:37
  The absolute highlight of this entire run was
• 16:41
  its solution for problem number three. Usually,
• 16:44
  human mathematicians rely on a standard graph
• 16:47
  -based approach to crack this specific puzzle.
• 16:50
  Right, but Red Note went a completely different
• 16:52
  route. It produced an unbelievably elegant mathematical
• 16:56
  induction proof instead. Whoa. Imagine an AI
• 16:59
  chaining together pure logic to invent an elegant
• 17:03
  induction proof entirely from scratch. It's crazy.
• 17:06
  It gives you absolute chills just thinking about
• 17:08
  the level of abstract reasoning required to do
• 17:11
  that. It really is a watershed moment. Now, we
• 17:13
  should mention the model still remains entirely
• 17:15
  in beta right now. Red Note hasn't disclosed
• 17:18
  its parameter size or its specific training data.
• 17:21
  They haven't shared their total compute budget
• 17:22
  with us either, which naturally leads to some
• 17:24
  healthy skepticism. If they haven't disclosed
• 17:27
  their training data, how do we know the model
• 17:28
  didn't just memorize similar proofs from the
• 17:30
  Internet? That's a huge, incredibly valid question.
• 17:34
  Memorization is always a concern with these models.
• 17:37
  But we should note that Huawei also reported
• 17:39
  a perfect 42 on this same test during this period
• 17:42
  using a totally different architecture. So Red
• 17:45
  Node is safely described as one of the first
• 17:47
  to cross this threshold, proving it's an industry
• 17:49
  -wide leap, not just a single lucky model. Exactly.
• 17:52
  It represents a massive paradigm shift in machine
• 17:55
  intelligence. AI math systems are definitively
• 17:59
  moving past simple pattern guessing and probabilistic
• 18:03
  text generation. They're actively writing full
• 18:05
  competition level mathematical proofs completely
• 18:08
  autonomously today. And these aren't just approximations.
• 18:12
  These proofs survive rigorous human grading panels
• 18:15
  looking for the slightest logical error. It fundamentally
• 18:18
  changes our understanding of machine capability
• 18:20
  in abstract domains. Let's drill down into that
• 18:24
  methodology again for a second. Right. What is
• 18:25
  combining natural language with Python execution
• 18:28
  actually achieved during a complex math test?
• 18:31
  Natural language allows the model to brainstorm
• 18:33
  abstract theoretical concepts dynamically, much
• 18:36
  like a human thinking out loud. And then Python
• 18:39
  allows it to instantly test those concepts mathematically.
• 18:43
  It generates a complex hypothesis in English,
• 18:45
  writes a script to text it, And verifies the
• 18:48
  outcome. Oh, I see. This creates a closed loop
• 18:51
  of absolute logical certainty. Yeah. If the code
• 18:54
  fails, the reasoning is flawed and it tries again.
• 18:57
  So it writes a mini program to double check its
• 18:59
  own logic. Exactly. It uses code as an unbreakable
• 19:03
  anchor for its abstract thoughts. We are clearly
• 19:06
  witnessing a massive maturation phase in artificial
• 19:09
  intelligence today. It is moving far beyond the
• 19:12
  simple text generation tricks that captivated
• 19:14
  us just a few years ago. For sure. We have quints,
• 19:17
  sprawling memories, successfully managing incredibly
• 19:19
  complex multi -step reasoning tasks over huge
• 19:22
  data sets. Yeah. We see Red Note crafting flawless,
• 19:24
  elegant mathematical proofs from pure scratch,
• 19:27
  rivaling the smartest humans alive. But alongside
• 19:30
  these miracles, we're also seeing brutal wake
• 19:32
  -up calls happening everywhere we look. Definitely.
• 19:35
  The stock market is violently punishing irrational
• 19:37
  AI investor hype and unchecked leverage. And
• 19:40
  legislative halls are frantically scrambling
• 19:43
  to build basic safety guardrails before the technology
• 19:46
  outpaces our ability to contain it entirely.
• 19:49
  AI is absolutely no longer just a digital sandbox
• 19:52
  experiment for researchers. It is driving immediate,
• 19:55
  tangible, real -world consequences right now,
• 19:58
  today. It's actively reshaping global finance.
• 20:02
  pushing the boundaries of the law, and fundamentally
• 20:04
  altering the landscape of scientific research.
• 20:07
  We've covered a lot of ground today, but I want
• 20:08
  to leave you with this one final thought. If
• 20:10
  an AI can write flawless mathematical proofs
• 20:13
  and execute complex parts of a cyber attack,
• 20:15
  but we still don't know how to build a fair kill
• 20:17
  switch to regulate it safely, where exactly does
• 20:20
  human intuition fit into the loop of the future?