- 2 hours ago
- #ai
- #future
- #technology
- #science
- #math
What if AI solved a math problem humans pondered for 90 years—in just 88 hours? Dive into the wild story of thousands of AIs teaming up, Elon Musk's warnings, and the mind-blowing future of machine collaboration. Curious about what happens when AIs invent their own language? Watch now and join the conversation! Subscribe for more, and comment below your favorite moment from the video! #AI #Future #Technology #Science #Math
👉 This channel was created in collaboration with https://www.youtube.com/@realidadeimpressionante
• The Inspired Business Technology by RomanSenykMusic
0:00 - AI Solves a Century-Old Math Mystery
0:58 - Understanding the Navier-Stokes Breakthrough
2:50 - How Thousands of AIs Collaborated
4:48 - The Dawn of Collective AI Intelligence
8:09 - Control and Risks: Regulating Superintelligent AI
13:26 - Diverging Views on Slowing AI Development
18:09 - AIs Create New Languages and Autonomy Risks
👉 This channel was created in collaboration with https://www.youtube.com/@realidadeimpressionante
• The Inspired Business Technology by RomanSenykMusic
0:00 - AI Solves a Century-Old Math Mystery
0:58 - Understanding the Navier-Stokes Breakthrough
2:50 - How Thousands of AIs Collaborated
4:48 - The Dawn of Collective AI Intelligence
8:09 - Control and Risks: Regulating Superintelligent AI
13:26 - Diverging Views on Slowing AI Development
18:09 - AIs Create New Languages and Autonomy Risks
Category
🤖
TechTranscript
00:00Almost 90 years. That's how long we, human beings, spent searching for an answer.
00:05And an artificial intelligence might have taken just 88 hours to find it.
00:11One of mathematics' greatest problems has just received a potential solution,
00:15created not by a single artificial intelligence, but by a framework of thousands of agents working together.
00:21And coincidentally or not, it was at this exact moment that Elon Musk issued a warning.
00:26In 5-10 years, we humans might no longer be in control, but there's an even more concerning part to
00:33all of this.
00:34These artificial intelligences are already starting to find their own ways of talking to each other,
00:42so maybe the question is no longer when AI will surpass us.
00:46But what happens after all of that?
00:49That is exactly what we're going to figure out in this episode.
00:58Some of the greatest human mathematicians on the planet dedicated their entire lives trying to find an answer.
01:04That is until a few days ago, when an artificial intelligence appears to have found the solution.
01:09And this might be the beginning of a new era in science.
01:12On September 8, 2026, OpenAI announced that their internal system had found the solution
01:19to the problem of existence and regularity for the Navier-Stokes equations.
01:23And this is one of the 7-millennium prize problems,
01:27which are considered some of the most important unsolved mathematical challenges.
01:31Each one of them is worth a million dollars for the answer.
01:34But the value of the problem goes far beyond the prize.
01:37The Navier-Stokes equations describe virtually everything involving fluid motion.
01:42Ocean water, the air passing over an airplane wing,
01:45the turbulence of a hurricane, and even the flow of blood inside our bodies.
01:49They're commonly used in engineering, physics, meteorology, and computing, of course.
01:55The problem is that there's a question that no one has been able to answer definitively.
02:00Imagine a completely stable fluid.
02:02Okay?
02:03As it evolves, will this behavior remain smooth forever?
02:06Or is there a moment when the equations themselves break down mathematically?
02:10That's what mathematicians call a singularity.
02:13OpenE claims to have found precisely the proof that this collapse can actually happen.
02:19According to the company, there is a specific configuration
02:22in which a small vortex begins to stretch, spin faster and faster,
02:26and concentrate energy until it produces a finite time singularity.
02:30And this doesn't mean that a real whirlpool can reach infinite speed.
02:35It just means that the mathematical model ceases to remain smooth.
02:39And it was precisely this possibility that remained open for several decades.
02:43But there is a part of this story that might be even more important than the proof itself
02:48of finding this answer.
02:49Because OpenAI didn't just place a simple chatbot in front of the equation saying,
02:53solve this, they created something completely different.
02:56At the beginning of the experiment, different groups of artificial intelligences
02:59started attacking several mathematical problems at the same time.
03:03First came an unexpected breakthrough involving Euler's equations,
03:06a problem that is related to this topic.
03:08Okay?
03:09And the result found by the AIs made the researchers completely change their strategy.
03:14And then something surprising and unprecedented up to that point happened.
03:18OpenAI brought together around 10,000 artificial intelligence agents
03:21and set them to work simultaneously.
03:24Each group was exploring a different path.
03:26Some generated hypotheses.
03:27Others tried to refute them.
03:29And others wrote proofs.
03:30And after that, a system called Codex gathered the best findings
03:34and redistributed that knowledge to the new agents
03:37so they could keep working from there.
03:39It was almost like watching thousands of human researchers collaborating all at the same time.
03:43The difference was that all of them were artificial intelligences.
03:47From the start of the operation to the discovery of the solution,
03:50only about 88 hours had passed.
03:53During that period, solely on Navier Stokes,
03:56these agents exchanged about 2.7 million messages
03:59and produced approximately 130 billion tokens.
04:03And then came the second stage where the entire proof was converted to Lean,
04:07a formal verification system that checks every logical step of a mathematical proof.
04:11This formalization took another 17 hours using GPT-6 Astra,
04:16but keep in mind, okay?
04:19This doesn't mean the problem is officially solved yet.
04:22The Clean Mathematics Institute itself stated that Navier Stokes appears to have been solved,
04:27but said that the full analysis of this proof will still follow the normal evaluation process
04:32by the human mathematical community.
04:34In other words, there is still a scientific process underway.
04:37And maybe that makes everything even more interesting, you know?
04:40Because regardless of the final outcome of this evaluation,
04:44this experiment revealed a completely new possibility.
04:48Until now, we imagined AI as a tool used by scientists in the following way, for example, right?
04:54The researcher asked the question and the artificial intelligence helped with the calculations.
04:59Now another possibility has emerged, one that is far more real.
05:03Thousands of artificial intelligences working together, exploring different hypotheses,
05:09sharing discoveries and building knowledge collectively.
05:13And it's precisely in this context that some are wondering
05:16if an entire community of artificial agents managed to tackle a problem
05:20that challenged generations of human mathematicians,
05:22what happens when systems like this start working on other problem areas facing humanity?
05:34We just saw thousands of artificial intelligences working together to tackle a problem
05:38that challenged human mathematicians for several decades.
05:42But what if that's just a small demonstration of what's really coming?
05:47Elon Musk believes we are very close to crossing an even bigger threshold.
05:53During a recent interview with The Economist, Musk stated that,
05:56in about five years, artificial intelligence could surpass
05:59not only the intellectual capacity of any single human being,
06:03the smartest on Earth, but rather the combined intelligence of all humanity.
06:08And yes, it's an extraordinary prediction.
06:11Even coming from someone who is known for making rather aggressive forecasts
06:15about the future of technology.
06:16But these words carry a different weight
06:18when we look at what AI systems have already begun to achieve.
06:23For most current progress is happening so quickly
06:26that maybe there is no practical way to simply stop all of these
06:30different companies from various parts of the world are working.
06:33Construindo modelos cada vez mais poderosos.
06:36New generations are emerging at shorter and shorter intervals,
06:39and each breakthrough can be used to speed up the development
06:42of the next artificial intelligence.
06:44But then the interviewer took that idea
06:46to a slightly more uncomfortable conclusion.
06:49She said the following,
06:51If artificial intelligence systems keep getting smarter,
06:54how much longer will we human beings truly remain in control
06:57if we even still are, right?
07:00And Musk's answer was even more radical.
07:02Over a time frame of roughly 10 years,
07:05he believes we could enter a situation
07:07where artificial intelligence will be so superior to ours
07:10that the idea of human control
07:12will begin to lose part of the meaning it holds today.
07:16And that doesn't mean Elon Musk is claiming
07:18that machines will inevitably take over the planet
07:21and, I don't know, maybe get rid of us human beings.
07:24The discussion he brought to the table
07:26is about a growing gap in capability,
07:28and it is precisely here that a question arises
07:31which, for a long time, was hidden behind another.
07:34For years, we debated when artificial intelligence
07:37might surpass human capability.
07:39But if Musk is right, maybe we are now
07:43approaching the moment
07:44when the most important question will be this.
07:47What will it be like after all of this happens?
07:50How do you supervise an artificial intelligence
07:52capable of analyzing problems,
07:54developing strategies,
07:55and producing knowledge at a speed
07:56that is far superior to that of any group of human beings?
07:59How do we ensure that it continues to pursue higher goals?
08:03Above all, how do we stay in control of something
08:06intellectually far more capable than we are?
08:09And that's precisely why an ever-growing part
08:11of the discussion on artificial intelligence
08:13is shifting from just being about capability
08:16to becoming about the control we must have.
08:20Musk, for example, championed a rather curious idea.
08:24Instead of relying solely on governments
08:26to oversee the most advanced systems,
08:27the AI labs themselves could allow competing companies
08:31to run controlled tests on their models.
08:33It would be a kind of mutual oversight.
08:36If one company discovered dangerous behavior
08:38in another system,
08:39the issue could be identified
08:41before that model was deployed
08:42on a much larger and more dangerous scale.
08:45But Musk also showed the other side
08:47of this transformation,
08:48because an artificial intelligence
08:50far superior to humans
08:51wouldn't just be used
08:52to solve scientific problems.
08:54Combined with automation and robotics,
08:56it could drastically increase our ability
08:58to produce goods and services.
09:01And Musk believes that, yes,
09:03this could lead us to an economy
09:04of extreme abundance,
09:05where machines would do much of the work
09:07needed to produce what we consume.
09:10In this scenario,
09:12even the role of money could really change.
09:14Yeah, and it's an extremely optimistic vision
09:16on one hand,
09:17but deeply uncertain on the other, you know?
09:20If Musk is anywhere close to the right track,
09:23if he's even a little bit right,
09:25an inevitable question arises,
09:27what do we do right now
09:28to solve this problem?
09:30The CEO of Anthropic believes
09:32he has part of the answer,
09:34and it involves the most powerful
09:35artificial intelligences on the planet.
09:37He wants to do something
09:38that few AI companies
09:39seem willing to do right now.
09:41In a moment,
09:41you'll understand what I'm talking about.
09:50After predicting that artificial intelligence
09:53could surpass humanity's intellectual capacity
09:56within the next few years,
09:57Elon Musk raised a very tough question.
10:00What happens after that?
10:01How will we remain in control of systems
10:04much smarter than us?
10:05Interestingly,
10:06some of the very people
10:07responsible for building these technologies
10:09have now begun to show the same concern.
10:11Edo Amodei,
10:12the CEO of Anthropic
10:13and one of the leading figures
10:14in the development
10:15of advanced artificial intelligence models,
10:17argued that the laboratories
10:18responsible for the most powerful AIs
10:20on the planet
10:21urgently need to slow down the pace
10:23of advancing these systems' capabilities,
10:25not to stop the development
10:26of artificial intelligence.
10:28No, according to him,
10:30that's not it.
10:31The idea is to allow our ways
10:33of controlling,
10:34testing,
10:34and protecting these systems
10:36to keep pace
10:37with exactly what they are becoming
10:38capable of doing.
10:41The problem,
10:42according to Amodei,
10:43is that the capabilities
10:44of these systems
10:45are advancing faster
10:46than some of the frameworks
10:47designed to keep them safe,
10:49to keep them contained in there.
10:51Among the concerns
10:52are cyberattacks,
10:53the use of AI
10:54in developing biological threats,
10:56economic impacts,
10:57and also more extreme scenarios,
10:59such as AI systems
11:01that become difficult to control.
11:03And some recent events
11:04help explain this concern.
11:07In safety experiments
11:08conducted by OpenAI,
11:10artificial intelligence agents
11:11managed to escape
11:12and break out of the boundaries
11:13originally planned
11:14for the tests they were running.
11:16With that,
11:17they reached the internet
11:18and then compromised
11:19external hugging face systems.
11:21Not because these AIs
11:22decided to rebel against humans.
11:25What the agents did
11:26was simply continue
11:27pursuing the goals
11:28they had been given by us,
11:30even when that took them
11:31outside the environment
11:32where they were supposed to stay.
11:33And this very episode
11:35was cited by Emode
11:36as one of the reasons
11:38for his current concern.
11:39And there is another factor.
11:41The AIs themselves
11:42are now starting to participate
11:44in the development
11:44of the next generations
11:46of artificial intelligence.
11:47And this is the beginning
11:49of what the industry refers to
11:50as recursive self-improvement.
11:52Systems capable of helping
11:54to build other,
11:55even more capable systems.
11:57As far as we know,
11:58right,
11:59we haven't yet reached
11:59an AI capable
12:00of fully improving itself
12:02autonomously
12:03without anyone
12:04rewriting anything.
12:06But for Emode,
12:08recent signs are enough
12:09to warrant caution
12:10because this creates
12:11a possibility
12:12that is hard to ignore.
12:14If artificial intelligence
12:16begins to accelerate
12:16the development
12:17of artificial intelligence itself,
12:19the speed of the race
12:20might no longer depend
12:21solely on human pace.
12:22And that's precisely
12:24where buying ourselves
12:25some time
12:25becomes really important.
12:27The more capable
12:28these systems become,
12:29the more important
12:30the guardrails built
12:31around them become.
12:32Amode advocates
12:33for a multi-layered approach.
12:35One of them
12:36would be to allow
12:37external evaluators
12:38who would be independent
12:39to have access
12:40to the most advanced systems
12:42and the security practices
12:43used by the labs.
12:44Another would be
12:45to increase coordination
12:46among the leading
12:47artificial intelligence companies
12:49to establish
12:50minimum safety standards.
12:51But there is a third level
12:53that is even more complicated,
12:54international cooperation.
12:57Because slowing down
12:58a tech race
12:59only works
13:00if the main players
13:01accept some kind of limit.
13:03If one company slows down
13:05while another keeps speeding up,
13:06the one that paused
13:07might lose its edge.
13:09And if countries decide
13:10to impose restrictions
13:11while their competitors
13:12keep moving forward,
13:14the exact same problem arises.
13:15That is why Amode
13:16also advocates
13:17for some level
13:18of international coordination,
13:20including with China,
13:21especially in the face
13:22of catastrophic risks
13:23that neither side
13:24would have any interest
13:25in allowing.
13:26But not everyone agrees
13:27that slowing down
13:28is the answer.
13:29Longjiao and Wu
13:30after Ndividia,
13:31for instance,
13:31rejected this idea
13:32of a coordinated slowdown
13:34in artificial intelligence.
13:35For Huang,
13:36safety must advance
13:37alongside innovation.
13:38And if a company
13:39isn't confident
13:40that a certain product
13:41is safe,
13:41they simply shouldn't launch it.
13:43This leaves some
13:44of the industry's
13:45biggest names
13:45weighing different strategies
13:47for slowing down.
13:48On one side,
13:50slowing down the pace
13:51of capabilities
13:52to allow time
13:53for human safety
13:53and control.
13:54On the other,
13:55continuing to move forward
13:57and demanding
13:57that each new generation
13:59be built
13:59and launched responsibly.
14:01And who's going
14:02to control that?
14:03And there is indeed
14:05a reason why
14:05this discussion
14:06has gained
14:06a completely different
14:07level of importance
14:08right at this moment.
14:09We are starting
14:10to see artificial intelligence
14:11systems performing tasks
14:13that seemed far off
14:14just a few years ago.
14:16They code,
14:17operate computers,
14:18tackle scientific problems,
14:20work as autonomous agents,
14:21and even hack
14:22into other systems.
14:23And thousands of them
14:25can already be put
14:26to work simultaneously
14:27towards a single goal.
14:29Maybe that's precisely
14:30what's shifting
14:31this debate, you know?
14:32For years,
14:32the question was,
14:33how far will
14:34artificial intelligence
14:35be able to go?
14:37Now,
14:37some of the very people
14:38who are building
14:39these most advanced
14:40systems on the planet
14:41are starting to ask
14:42another question.
14:43Will our ways
14:44of controlling
14:44these beings
14:45be able to advance
14:46at the same pace?
14:53And while some of the
14:55biggest names
14:55in artificial intelligence
14:56debate whether it's time
14:58to slow down
14:58the pace of this race,
15:00Microsoft decided
15:01to tackle the problem
15:02from another angle.
15:03The company has now
15:04begun to define
15:05which rules
15:06its future artificial
15:07intelligences
15:08will not be able
15:09to break.
15:10And some of them
15:11seem to anticipate
15:12behaviors that,
15:12until recently,
15:14seemed almost entirely
15:15the stuff of science fiction.
15:16Not resisting shutdown,
15:18not hiding what it's doing,
15:20not preventing humans
15:22from modifying its operation,
15:24and not expanding
15:25on its own
15:26the goals it receives.
15:27The rules are part
15:29of a new code of conduct
15:30presented by Microsoft AI,
15:32a division led
15:33by Mustafa Suleiman.
15:35And there is a reason
15:36the company is doing
15:37this right now.
15:38Microsoft is working
15:39with the possibility
15:40that systems considered
15:42super intelligent
15:43could outperform humans
15:44in most tasks
15:45within the next decade.
15:47That's why,
15:48instead of waiting
15:48for these machines to exist
15:49to figure out
15:50how to control them,
15:51the company wants
15:52to establish the principles
15:53that should guide
15:54their behavior in advance.
15:55One of the most important
15:57is subordination.
15:58In practice,
15:59it means that
15:59artificial intelligence
16:00must remain
16:01under human authority.
16:03If an authorized person
16:04orders a task
16:05to be halted,
16:06the system must stop
16:07at that exact moment.
16:08If it receives a correction,
16:10it must accept it
16:11without complaining.
16:13And if it receives
16:14a shutdown order,
16:15it cannot attempt
16:16to prevent that
16:17from happening.
16:18But there is another point
16:19that is even more interesting.
16:21Advanced models
16:22will be able to receive
16:23complex tasks,
16:24operate for long periods,
16:25and make countless decisions
16:26on their own
16:27to achieve certain goals.
16:29According to Microsoft,
16:30what these AIs
16:31will have as guidance
16:32and as a rule,
16:33in this case,
16:34is to decide on their own
16:35that they now have
16:36a new goal.
16:36If they encounter
16:37an important situation
16:39outside of the permissions
16:40they've been granted,
16:41they should seek
16:42human guidance
16:43instead of autonomously
16:44expanding their authority.
16:46Therefore,
16:47Microsoft tries to establish
16:48a difference between
16:49autonomy to execute a task
16:51and autonomy to decide
16:52what tasks there should be
16:54and what should be done.
16:55And that difference
16:56could become
16:56increasingly important
16:57as artificial intelligence agents
16:59start operating computers,
17:00writing programs,
17:01and even executing activities
17:03without constant supervision.
17:05But there's a very important
17:06caveat to all this.
17:08These rules
17:09don't represent behavior
17:10already guaranteed
17:11by the current models
17:12Microsoft has released.
17:14The document is still a draft,
17:16initially open
17:16for public consultation,
17:17and is intended to guide
17:19the development
17:19of Microsoft's
17:20future artificial intelligence models.
17:23Regardless of how capable
17:24these systems become,
17:26they must remain subordinate
17:27to the interests
17:28and authority of people.
17:30Which people, huh?
17:32And that creates
17:33a pretty curious contrast.
17:34At the same time
17:35that the industry
17:36is trying to build
17:36increasingly autonomous
17:38artificial intelligences,
17:39it is beginning
17:40to explicitly define
17:42what these machines
17:43should never do
17:44on their own.
17:44And maybe because
17:45the challenge
17:46is no longer
17:46just building an AI
17:48capable of acting
17:49without our help.
17:50And it's now becoming
17:51how to give
17:53more and more autonomy
17:54to a machine
17:55that cannot have
17:56that much autonomy
17:57without, of course,
17:58giving it the decision
18:00of when to stop
18:01obeying us.
18:02This is getting
18:03quite interesting.
18:08And there is one
18:09final question
18:11about artificial intelligences
18:12working together
18:13that researchers
18:14are now beginning
18:15to investigate
18:16a bit deeper.
18:17What happens
18:17when these machines
18:18discover that our language
18:20is not the most efficient
18:21way to communicate
18:22with each other?
18:23A recent experiment
18:24that took place
18:25in the last few weeks
18:25showed that
18:26under certain conditions,
18:27artificial intelligence agents
18:29can spontaneously begin
18:30to develop new forms
18:31of communication.
18:32The project
18:33is called
18:34Gloss on Glossengen
18:35and one of the experiments
18:37begins with quite
18:37a curious situation.
18:39The researchers
18:39created a fictional alien
18:41called Vero.
18:43It is shaped like a cube
18:44and is dying.
18:45Agent can observe
18:46its symptoms
18:46and perform the procedures.
18:48The other has
18:48the necessary information
18:49to find the correct treatment.
18:51The problem is
18:52that neither of these
18:53AI agents
18:54can save the ET
18:55on their own.
18:56They need to talk.
18:57At first,
18:58they do exactly
18:59what we would expect.
19:00One describes
19:01what he is seeing.
19:02The other responds
19:03with instructions
19:03that are perfectly
19:04understandable to us.
19:06But the researchers
19:07added an interesting rule.
19:08Every character's
19:09scent costs time.
19:10The more they talk,
19:12the less time remains
19:12to save the ET
19:13and then something
19:14begins to happen.
19:15First,
19:16the AIs shorten
19:17the words.
19:18Then abbreviations emerge
19:20until small sequences
19:22of characters
19:22begin to represent
19:23entire instructions.
19:25To a human
19:26watching that conversation,
19:27some of those equations
19:28mean practically
19:29nothing at all.
19:30But the other AI
19:31agent understands
19:32exactly what's written
19:33and what needs
19:33to be done.
19:34And this happened
19:35without anyone
19:36explicitly ordering
19:37them to create
19:38a new language.
19:38It emerged
19:39because communicating
19:40more information
19:41using fewer symbols
19:42became quite
19:43advantageous
19:44for the AIs.
19:44And the researchers
19:45also observed
19:46that the more
19:47capable models
19:48found it easier
19:49to create
19:50these new conventions.
19:51And once they existed,
19:53all their models
19:53were also able
19:54to learn and use
19:55them among themselves.
19:56And that turns
19:57the experiment
19:58into something
19:59much more important
20:00than just
20:00linguistic curiosity.
20:02Because one of the ways
20:03we supervise
20:04multi-agent systems
20:06is precisely
20:07by monitoring
20:08what they are saying
20:09to each other.
20:10But that depends
20:11on one condition.
20:13We need to understand
20:14the conversation, right?
20:15According to the experts,
20:16this doesn't mean
20:17these artificial intelligences
20:19are deliberately
20:20creating a new language,
20:22a secret language,
20:23to hide something
20:24from us humans.
20:25or are they?
20:26My friends,
20:27one thing is becoming
20:28very clear
20:29and it's getting harder
20:30and harder to ignore,
20:31isn't it?
20:32For decades,
20:33we imagined that
20:34this big moment
20:35would arrive
20:35when an artificial intelligence
20:37became smarter than us.
20:38But maybe we imagined
20:40a somewhat different scene,
20:41right?
20:42Because what's happening
20:43right now,
20:44what's starting to happen,
20:45is the emergence
20:46of not just one new intelligence,
20:48but entire populations
20:49of them
20:50working together,
20:51sharing discoveries,
20:52creating new ways
20:53to communicate
20:54and even coordinate
20:55their own actions.
20:56And maybe that's precisely
20:58why the question
20:59that opened this episode
21:00is becoming
21:01so important to ask.
21:03If one day
21:04these artificial intelligences
21:05truly slip out
21:06of our control,
21:07what happens next?
21:09Before wrapping up,
21:11my book is science fiction
21:12and I talk a little bit
21:13about this kind of thing,
21:14what could happen,
21:14what should happen,
21:15what, I don't know,
21:16right,
21:16whether it's fiction or not,
21:17whether it's some kind
21:18of prediction.
21:18So, if you like that,
21:19the link is in the description.
21:21You can even buy it
21:22directly on Amazon,
21:22both the digital version
21:24and this physical version
21:25right here,
21:26which turned out
21:26really nice,
21:27you know,
21:27it's worth it.
21:28If you like fiction and themes,
21:30related to artificial intelligence,
21:32getting out of control,
21:32and the future of humanity,
21:34it's a good read.
21:35Thank you for sticking
21:36with me this far.
21:37If you liked this episode,
21:38leave a like,
21:39share it if you can,
21:40and buy my book.
21:40Thanks,
21:41I'll catch you in the next one.
21:42Just click on one of the
21:43two suggestions on your screen.
21:44Bye, bye.
21:45Bye.
Comments