Latest / Today’s AI News / DeepMind’s AI Just Cracked Unsolved Math Problems, NASA’s AI Found 100 New Planets, AI Trains 4 Dogs

Transcript

DeepMind’s AI Just Cracked Unsolved Math Problems, NASA’s AI Found 100 New Planets, AI Trains 4 Dogs · Today’s AI News

• 0:00
  Imagine trying to build a skyscraper but instead of 1
• 0:03
  master builder barking orders, you have this synchronized
• 0:06
  telepathic construction crew where everyone just instantly
• 0:09
  tackles a different floor. That's exactly what the newest
• 0:12
  AI systems are doing right now. Basically shifting from solitary
• 0:15
  answer bots to entire teams of agents collaborating in real
• 0:19
  time. Welcome back to today's AI News,
• 0:22
  brought to you by 9X Productions.
• 0:24
  Yeah, thanks for having me back. It's it's such a fundamental
• 0:26
  architectural shift we're seeing right now.
• 0:28
  It really is, and that is exactly our mission for you.
• 0:31
  On today's Deep Dive, we are exploring how these new multi
• 0:34
  agent teams are teaming up to solve like previously impossible
• 0:38
  math, how they're uncovering hidden planets in old astronomy
• 0:42
  data, and honestly how you can put these autonomous crews to
• 0:44
  work on your own laptop today. Right, because to really grasp
• 0:48
  the magnitude of this shift, we kind of have to start with that
• 0:50
  construction crew metaphor you just used and apply it to
• 0:53
  something super unforgiving, which is advanced mathematics.
• 0:56
  Yeah, let's get into the mechanics of that.
• 0:58
  OK, so Google DeepMind just published this paper on an AI Co
• 1:02
  mathematician and I want to be clear for the listener here,
• 1:05
  this isn't just a standard chatbot trying to do algebra.
• 1:08
  No, not at. All built on the Gemini 3.1
• 1:10
  architecture, but it's structured as a multi agent
• 1:13
  system. Exactly.
• 1:14
  DeepMind basically modeled this after complex software
• 1:18
  development environments. So instead of a single text
• 1:21
  prompt, they give the AIA full workspace.
• 1:24
  Right. They use this coordinator agent.
• 1:26
  So back to our analogy. This is your project manager.
• 1:29
  You hand this manager an unsolved math theorem and it
• 1:32
  just breaks the research down into parallel work streams.
• 1:35
  It delegates like it assigns 1 sub agent to write Python code,
• 1:39
  another one to search existing literature and others to
• 1:42
  actually attempt the proofs. And the really crucial mechanism
• 1:45
  there is the built in review cycle, like the agents were
• 1:49
  testing hypotheses, they're failing, they're reviewing the
• 1:51
  failures and iterating all at the same time, which is.
• 1:54
  Wild. It is, and to understand how
• 1:56
  powerful that is, look at the benchmark they used.
• 1:59
  It's Epoch AI's Frontier Math Tier 4.
• 2:03
  This is explicitly designed to resist AI memorization.
• 2:07
  Right. You can't just guess your way
• 2:08
  through. Exactly.
• 2:09
  It requires novel logical leaps, and the raw Gemini 3.1 Pro
• 2:14
  model, like operating as a single agent, scored 19% on this
• 2:18
  test. Which honestly is a pretty
• 2:19
  respectable baseline for research level math.
• 2:21
  Oh for sure, But when DeepMind took that exact same model and
• 2:25
  deployed it as this agentic team, the score jumped to 48%.
• 2:29
  Wait, really? Just by changing the workflow.
• 2:31
  Just by changing the environment they more than doubled the
• 2:34
  performance without changing the core models actual intelligence.
• 2:37
  That is insane. But you know, the real world
• 2:40
  application of this is where the story gets super fascinating.
• 2:43
  Oxford Professor Mark Lackenby used this exact system on an
• 2:46
  open problem in the Korovka notebook.
• 2:49
  Right, the famous Registry of unsolved group theory problem.
• 2:52
  Yeah, exactly. The study of symmetry and all
• 2:53
  that. And Lackenby actually cracked
• 2:55
  the problem using the AI. Here's the crazy part.
• 2:58
  The AI didn't just hand him a pristine proof.
• 3:01
  Yeah, this part is so revealing about where the tech is at right
• 3:05
  now. Well, that can be actually found
• 3:07
  this really, really clever proof strategy buried inside an output
• 3:12
  that the AI's own internal reviewers had rejected.
• 3:15
  Just threw in the trash. Right.
• 3:16
  The roject manager agent flagged it as a dead end.
• 3:19
  Lackenby had to dig it out and see the brilliance, so I got to
• 3:22
  push back on this a bit. If the AI's own reviewer threw
• 3:26
  away the winning ticket, doesn't that prove human intuition is
• 3:29
  still like the ultimate filter? Oh absolutely, I completely
• 3:32
  agree and I think that's the whole So what of this story.
• 3:34
  It proves the near future is about AI accelerating Top Minds,
• 3:38
  not replacing them. A cognitive exoskeleton basis.
• 3:41
  Exactly. It can do the heavy lifting of
• 3:42
  exploring 10,000 dead ends in the time a human explores one.
• 3:47
  It gives you the raw materials, but the human still has to spot
• 3:50
  the diamond in the rough. You change from creator to
• 3:53
  curate, which actually is a perfect pivot because if a team
• 3:57
  of AI agents can tackle abstract math, what happens when you
• 4:01
  deploy an autonomous crew to handle your mundane daily tasks?
• 4:05
  Yeah, because we're seeing people do this on their personal
• 4:07
  computers right now. Right, there's this story from a
• 4:10
  developer named Jason. He used Open AI's codecs with
• 4:14
  just a simple slash goal command.
• 4:17
  He told it to build an entire companion app for Magic the
• 4:21
  Gathering. Because he and his brother
• 4:23
  wanted to play across different time zones.
• 4:25
  Right, Yeah. He couldn't sync up their
• 4:26
  schedules, so he needed an asynchronous app.
• 4:30
  And Jason didn't write it line by line.
• 4:32
  He just gave the goal and the AI one shot at it.
• 4:35
  It's incredible. It just worked autonomously in
• 4:37
  the background. The back end, the database, the
• 4:40
  front end. And then you have someone like
• 4:42
  Jerry, who isn't a developer, using Clod to plan this flawless
• 4:46
  trip to Greece. Oh I saw that.
• 4:48
  Booking flights, mapping transit times between islands, even
• 4:52
  picking restaurants city by city.
• 4:53
  Yeah, a tighter plan than a human travel agent could have
• 4:56
  done in that time. Well, it's a massive shift,
• 4:58
  right? We're moving from prompting for
• 5:00
  text to delegating actions. And listeners can do this right
• 5:03
  now. Like if you want to use Codex's
• 5:06
  computer use on Mac or Windows, you enable the plug in, but you
• 5:10
  have to switch the permissions to full access.
• 5:12
  Which is a big deal. It's a huge.
• 5:13
  Deal. You give it a real task like
• 5:15
  open Chrome, debug this local web page, reproduce the bug and
• 5:19
  ask before making changes. You can even have it do
• 5:22
  Photoshop exports. Right, it's literally looking at
• 5:25
  your screen and moving your mouse.
• 5:26
  But that's my question. Handing over full access to your
• 5:30
  mouse and keyboard feels like letting a ghost drive your car.
• 5:35
  I mean, is it really safe for the average user?
• 5:37
  Well, this is where the guardrails are so important.
• 5:39
  That part you mentioned, telling it to ask before making changes,
• 5:42
  that is crucial. So you have to program
• 5:44
  hesitation into it. Exactly.
• 5:46
  You act as the senior manager, You tell it to draft the
• 5:49
  contract, but you don't let it send the e-mail until you review
• 5:52
  it. The security implications are
• 5:54
  immense, so you have to define its boundaries.
• 5:56
  OK, that makes sense. So we've got AI navigating local
• 5:59
  code, we've got a planning trips, but its ultimate power is
• 6:03
  pattern recognition at massive scale, specifically scanning the
• 6:07
  cosmos. Yeah, going from your desktop to
• 6:10
  the entire Galaxy. Literally, yeah.
• 6:12
  Astronomers at the University of Warwick used an AI system called
• 6:15
  Raven and they had scanned 4 years of NASA test data looking
• 6:20
  at 2.2 million stars that. Scale is just hard to
• 6:23
  comprehend. It really is, and Raven
• 6:25
  confirmed over 100 exoplanets and found over 2000 potential
• 6:29
  candidates. It does the detection, the
• 6:31
  vetting, and the confirmation all in one shot.
• 6:33
  And it found stuff humans missed, right?
• 6:36
  Yeah, 31 exoplanets never spotted before.
• 6:38
  It even found planets orbiting stars in under a day.
• 6:41
  But the craziest part is it found hundreds of planets in the
• 6:43
  Neptunian desert. Right, which is that scorching
• 6:46
  zone close to stars where Neptune sized lanets supposedly
• 6:49
  couldn't even survive. Exactly.
• 6:51
  The old model said it was impossible, but here's the
• 6:54
  thing. Finding a needle in a haystack
• 6:56
  is one thing, but this haystack was 2.2 million stars.
• 7:01
  Does this mean the AI is smarter than us?
• 7:03
  Or is it just reading the data way faster than humans ever
• 7:06
  could? It's the latter, but it
• 7:07
  represents A profound paradigm shift.
• 7:09
  The data was already sitting there on our servers.
• 7:12
  We didn't need a billion dollar new telescope.
• 7:14
  No, we just needed a better brain to analyze the data we
• 7:17
  already had. I mean, there are trillions of
• 7:19
  planets out there and AI is about to rapidly rewrite the
• 7:23
  textbooks just using archival data.
• 7:25
  That's a great point. And Speaking of rapid shifts
• 7:28
  from the vastness of space, let's pull back to for some
• 7:31
  quick hits on everything else happening in AI today.
• 7:33
  Yeah, there's a. Lot moving right now.
• 7:35
  First up, new tools, codecs, just launched a Chrome extension
• 7:39
  for agentic tasks and then printing press dropped this CLI
• 7:42
  factory, a command line interface factory with over 30
• 7:46
  prebuilt agent native tools. Which is huge for developers
• 7:50
  wanting to easily plug agents into their systems.
• 7:53
  Definitely also Open Reader Launch Pareto code.
• 7:56
  This is a free layer that automatically routes you to the
• 7:58
  cheapest coding AI that meets your quality bar.
• 8:02
  I love that developers are obsessing over efficiency and
• 8:05
  cost roading right now. And on the cost side, Baidu
• 8:08
  released Ernie 5.1. It hit #4 on the Arena search
• 8:13
  leaderboard, but supposedly it only costs 6% of what rival
• 8:16
  models cost to train. 6% If that's true, that's a massive
• 8:20
  maturity milestone for model distillation.
• 8:22
  Yeah, driving costs way down. Meanwhile, Google's Isomorhic
• 8:26
  Labs is reportedly raising over $2 billion to extend its drug
• 8:30
  design engine, which is outperforming Alpha Fold Three.
• 8:33
  Wow. But all this software needs
• 8:34
  power. SoftBank Group's telecom arm
• 8:37
  just launched a massive battery business for large scale cells
• 8:40
  to meet the power demands of new data centers.
• 8:43
  Right, because infrastructure is just scrambling for physical
• 8:45
  energy at this point. It's.
• 8:46
  Fascinating. You have Baidu and Open Router
• 8:48
  aggressively driving down the cost of using AI while SoftBank
• 8:52
  is literally building giant batteries just to keep the data
• 8:55
  centers powered up. It perfectly illustrates the
• 8:57
  ecosystem. Maturing software gets
• 9:00
  sufficient. Hardware demands more energy.
• 9:02
  And then there's the policy side.
• 9:04
  Greece is actually proposing to put AI protections directly into
• 9:08
  its national constitution. Into the Constitution.
• 9:10
  That's a major step. Prime Minister Mitsotaki cited
• 9:13
  threats to democracy, saying the tech needs to serve individual
• 9:17
  freedom. Just impartially reporting on
• 9:19
  the proposal here, but it's really interesting to see a
• 9:21
  nation attempt to create foundational legal frameworks at
• 9:25
  that level. Absolutely.
• 9:26
  Nations are realizing this isn't just software, it's
• 9:28
  infrastructure. Exactly.
• 9:30
  But you know, after looking at global constitutions and billion
• 9:34
  dollar drug design, let's bring it back to something super
• 9:37
  personal and close to home. Let's do it our community
• 9:40
  workflow. Yes, we got a story from a
• 9:42
  listener, Anonymous, who used ChatGPT to train their four
• 9:46
  chaotic dogs. Four dogs.
• 9:48
  That is a lot of variables. It's a highly volatile multi
• 9:51
  agent physical system. That's one way to say a pack of
• 9:54
  dogs. Well, they were about to spend
• 9:56
  thousands on a professional trainer, but instead they
• 9:59
  treated the AI like a systems analyst.
• 10:02
  They fed it the ages, triggers, house layout, how the dogs
• 10:06
  interacted, all the root causes. Providing that context is key so
• 10:10
  you don't just get generic advice.
• 10:12
  And it gave them tailored techniques, they replace the
• 10:14
  trainer fees, completely changed their household dynamic with
• 10:17
  positive reinforcement, and got their sanity back.
• 10:20
  I love that. And honestly, it ties everything
• 10:22
  together. Yeah, I mean, think about it.
• 10:24
  I want to leave the listener with this thought.
• 10:27
  If an AI can decipher all the hidden patterns in advanced
• 10:30
  math, or in distant stars, or even just in a pack of unruly
• 10:33
  dogs in your living room, what invisible patterns in your own
• 10:37
  daily life are just waiting to be uncovered?
• 10:39
  Oh man, that is a great question to Mull over.
• 10:44
  The patterns are out there, we just finally have blends to see
• 10:46
  them. Well, that wraps up our deep
• 10:48
  dive for today. Be sure to hit the subscribe
• 10:50
  button to stay updated and all these crazy shifts.
• 10:53
  And if you enjoy the show, please take a second to rate US
• 10:55
  five stars. It really helps out.
• 10:57
  Thanks for joining us, and we will see you back here for
• 10:59
  tomorrow's Deep Dive.