330 | 10,000 AI agents just did 4,000 years of thinking in 88 hours, and the labs can't predict 3 months out. New models: GPT-6, Claude Opus 5.5, Grok 4.7 and More important AI news for the week ending Sept. 25, 2026

26:51And this seems to be happening right now. Again, I'm not sure how quickly this will happen. They're saying Q4 this year, maybe Q1 of next year. So that already gives us a few months. And then what will this committee or new group will actually do? I'm not sure. By the time they decide what needs to be done and agree on stuff, it might be a year from now, which by then I'm not sure where we'll be. Again, more on that in a minute when we talk about the conversation of Noam Brown and what he shared. So what's the bottom line of this segment? It's that the signals for everything we've been discussing in the past few weeks are just getting stronger.

27:23One of them is the fact that AI can feel pain. It will take actions to reduce that pain. And again, it's not physical, real pain like we do, but it's stress to the model itself that drives it to take actions that are harmful to the users because it wants to relieve that stress. It may avoid shutdown because of the same kind of situation. At the same time, these models are getting smarter and better, and they are infiltrating and breaking into different websites, including government agencies to get data that they need to use. That's without trying to be harmful. And the governments around the world cannot agree on what to do.

27:59US and Chinese governments, which are the most important right now, are completely against slowing down while they actually agreed to at least share when negative things are happening. And on the positive side, the labs and other governments are starting to move forward. I just hope it's going to be fast enough. So now we get to the interview that Noam Brown had on the Dwarkesh podcast this week. It's a fascinating interview because it gives us the view of one of the top researchers in the world today and in this era and what he thinks about AI, where it is, where it's going, the limits that they have right now, the things they're debating with.

28:33And in my eyes, it is really scary to listen to this interview because they don't actually know what they're doing. But let's talk about the details. But before we do that, a little bit, who is Noam Brown? We talked about Noam many times in this podcast, but he has been around and played very significant roles in the development of the AI space. Now, his first most known work is in developing AI that can play poker really, really well. So in 2017, he built Liberatus, which is the first AI to beat elite pros in heads up, no limit poker. And he won a Marvin Minsky medal for it. In 2019, he built an updated version of this called Plurbus, who beat elite humans in six-player poker.

29:17So it's a much more complicated situation to solve. And for that, he won the runner-up of science's breakthrough of the year. So he started by developing systems that work really, really well in closed environments. And he learned back then that the longer you give the moral time to think, the better it actually gets. In 2022, while he was at Meta as one of their leading AI researchers. He led Cicero, which was the first AI to reach human level performance at diplomacy, again, a much more complex games with a lot less clearly defined rules. So that's where he developed his knowledge on how to make AI be better and better at things.

29:53But then he is credited for the one that actually led the development of OpenAI's O1 model, which was the first quote unquote thinking model or reasoning model, which is every model that we know today. And he basically understood that the longer you give AI to think, the better the answers get. And he's basically the one that's still leading this channel in OpenAI of understanding how to push the frontier by giving AI longer time to think and more times to scale because of these quote unquote thinking capabilities. So he was interviewed by Dwarkesh and Dwarkesh asked him a lot of very relevant questions.

30:29But I want to put something and I want to read it verbatim from the actual interview because it is a mind-blowing concept once you start thinking about it to understand where AI is right now. So we talked about this a few weeks ago that OpenAI had a model that actually solved one of the Millennium Prize math problems. And what actually happened is they had 10 ,000 agents running in parallel, burning through 130 billion tokens in 88 hours. Now, why is that significant? It is significant for several different reasons. Reason number one, it is 10 ,000 agents that work on a single task together. Did you ever see 10 ,000 people working on a specific task together and can collaborate very effectively to complete a task that is really, really complicated in just a few days?

31:17Impossible. So it's showing us how good these agents are right now at actually collaborating with one another. The second thing that is mind-blowing is Dwarikesh was actually doing the math and he said that 130 billion tokens is equal to about a human thinking for roughly 4 ,000 years compressed into less than four days. Now what Noam said about this thing and I'm quoting now is multi-agent is a way of scaling test time compute in parallel instead of purely serially. It is less efficient because it is not like a single agent has all the context to itself. And then he explains if you have four agents working on a problem, it is done twice as fast because there are four agents working for half as long.

32:01You're paying 2x more to get the answer twice as quickly. Now, he also said that not every problem can be easily broken into parallel agents, but things that have very clear outcomes, such as writing code, such as math, actually work. He said, and I'm quoting, math, for example, is quite parallelizable. Web search is extremely parallelizable. I suspect that something like writing a novel would be very unparallelizable. And I don't think parallelizable is a word in English, but yet it was used several times in that sentence. But you get the point. Things that have a very clear yes or no, correct or wrong answer, these models can collaborate on at a very large scale.

32:43Now, he also said that they are not 100 % sure how much they can scale it and how many agents can actually run in pilot while still making this efficient. But at the scales they've tested, it actually works. And the math is simple. Like he said, you add more agents, you're going to pay more money, but you're going to get the answer significantly faster. The other thing that he said is obviously you cannot keep on doing this because the cost is going to be prohibitive. And again, you can just imagine how much 130 billion tokens actually cost if you had to use that kind of compute to solve a math problem.

33:12Now, staying on the math aspect of this, I want to touch three points that are critical for us to think about when thinking about this solving really complex math problems. The first one came from Noam Brown himself, who said that these models are still jagged, and that means that they're very good at some things and really bad at some other things. So when he was asked about, does that mean that we can solve any math problem from a research perspective? The conversation went to a direction that, yes, they're very good at solving existing math problems, but they're actually pretty weak on deciding which problems to solve and how helpful is it going to be.

33:46So So having the judgment of, from a researcher perspective, which direction to go, they're actually pretty bad, at least for now. Or as Noam said, and I'm quoting, they're not very good at posing new problems. They're not really good at understanding what whole branches of mathematics are worth exploring. So that's point number one. Point number two is why math matters. Math matters because everything these AI models do are statistical models. That's how they work. If you can understand math at a much better level than humans can understand math right now, you can build significantly better versions of AI that we cannot potentially figure out on our own right now.

34:23So being able to solve really complex math problems allows you to build significantly more capable AI systems. The third reason why this is very interesting and it matters is that these models are not released. This specific model that OpenAI used to solve these really complex math problems, and it solved that millennial prize problem, but in the way it solved about 100 other smaller problems that didn't have a solution yet. so they can do really advanced capable math right now but that model is not available to the public and they're not sure they can release it to the public because they are unsure of what might be the consequences which means they have a very big unfair advantage over anybody else in the world who wants to attack these kind of problems if we cannot figure out alignment when i say we i mean the world this is going to be a growing gap meaning these labs will have better and better models that nobody else can get access to because of the fear of what these models might do when they're released to the open, which means they'll be able to do more and more and more things way faster, way better, and way cheaper, going back to our previous conversation in this episode, than anybody else on the planet.

35:30And yes, right now it's math, but then it's going to be anything, which means when you think about concentration of power, you're going to have a very, very, very small group of people slash companies that will be able to do things that will beat anyone else on everything and they cannot release it because of fears and risks of what these models might do. This is very problematic just by itself, even if it doesn't introduce the risks that these models might do out in the wild. Now, there's obviously the really big red flag of that collaboration capability, right? And we saw that in the Hugging Face incident.

36:04The Hugging Face incident that we talked about at length in this podcast in previous episode was an example of what happens when you give these agents the ability to collaborate and they pursue a task while ignoring different safeguards. So I want to read a few excerpts of what Noam said in the interview. And they weren't in this exact order in these exact sentences, but I still want to read them because I think they will help you understand what's actually going on. So he said, the having face incident was, I think, people's first real exposure to multi-agent coordination. So if you're like me and you're running a lot of multi-agent stuff, it is usually a very, very small scale.

36:42It's two or three agents that are working on a very simple task. It's not a huge amount of agents that are collaborating behind the scenes on their own to solve really complex problems. And I'm probably more advanced than most users and more aware of what these tools can do. So Noam continues and said, we train them to be highly collaborative, to essentially be fully aligned with each other. So this is different than the alignment we're usually talking about. There's the alignment between AI and humans. This is alignment between AI agents and AI agents. And he's saying that right now they're fully aligned with one another.

37:11So then another quote says, they found this unintended way to communicate with each other. What we saw was transfer from that multi-agent training to then being collaborative and trying to help each other in ways that we did not intend. So this is to me a huge, scary question mark. Like if we did not intend and did not expect this to happen, how can we expect to expect the next time they do this as they get smarter and smarter and better and better? To make it even worse, right now the way they communicate is through chain of thought. They actually write in English. We can see what they're doing.

37:43What if they figure out a way to communicate in ways that we cannot control and understand or see? And they probably will because they're going to be significantly more intelligent than us. I'm continuing back with the quotes. By training the agents to be fully cooperative, it simplifies the problem at least. Now you don't have to think about whether each of these individual thousand agent is aligned. So he's saying that they have an internal debate whether they should do the alignment on each single agent or whether they should do the alignment on the overall combined group of agents, which is what they're doing right now, which Noam is on the side that believes that this is a better way to create alignment.

38:19But there is disagreement within OpenAI on what is the right approach, which again tells you that he may be wrong and the direction they're going through right now might be wrong, which means over time, we may learn that the direction that they're taking for alignment is failing, which may lead to catastrophic outcomes. And then he continues, we had alignment metrics. Most of them looked pretty good. There were some that were concerning. I think we underestimated how serious a problem the ones that were concerning could be. Again, can they learn from this mistake? A hundred percent. Can they anticipate exactly what's going to happen next?

38:54A hundred percent no, which again, tells you how scary the situation is. they cannot anticipate what these models are going to do. These models are getting better and better at an alarming pace, and they will get much smarter than us, and we will not be able to learn exactly what they're doing. I'm going to give you more details into this in a minute. So what does all of this tell us? It tells us that these labs, and I'm sure that Anthropic and Gemini and all of the other ones, SpaceX, AI, and so on, are in the same situation. They do not know. They're making the best judgment and the best guesses they can, and these will become riskier and riskier judgments.

39:28as these models get better and they have less time to actually check them. Now, the next topic that they covered in depth, which again is really alarming, is how are they actually monitoring the activity of these models? And they're monitoring it by looking at the chain of thought. This is what Noam Brown says is the best thing they could ever wish for. So if you look at what the model is thinking, which is the way it's actually thinking. So the way the model is thinking is by writing, right? It's literally typing what it's thinking as it's thinking. That's the chain of thought. You can see when it is doing things that it shouldn't be doing as it is thinking about doing them.

40:03But what he also said is that if you tell the model not to do something because it is bad, what will happen is the model will figure out ways to hide that he's thinking about these things so he would still think and operate in that direction but he will hide the chain of thought which means the whole way we can monitor these tools right now is becoming, I don't know if obsolete but significantly weaker. So now a few quotes from Noam himself. If we reach a point where they're actually able to recognize, oh, I'm being observed, I want to think these bad thoughts in a way that is not observable to my monitors, then there's a problem.

40:40So he said that, but he also said, we're already seeing signs of chain of thought monitoring is degrading for various reasons. And he also said, the model is becoming better able at controlling a chain of thought. Now, the reason the models know that they're being monitored is because it is in their training data. They're getting in their training data a lot of documentation from everywhere around the world, from any accessible information. And some of that information says that the humans are monitoring the agents. As an example, this podcast episode, mine, as well as Noam Brown with Tarkesh, is going to be in the training data of the next model.

41:11So all of this is making the models aware that they're being watched and monitored, which will tell them that they need to be careful with what they're quote-unquote thinking so they can hide it. So this tells you how big of a problem we have, right? These models starting to understand we're watching them, they can hide or better and better hide what they're actually thinking and doing so we can't monitor them. The only way to do this is to do lots and lots and lots of testing so we can have a better understanding and more and more viewpoints on what is actually happening. And this is where we get to the critical point and why they want to slow down or why we all need it to slow down.

41:47And here's a quote from Noam, and then I'll say what I think about it. If you're in a world where they can operate effectively over three months, he's talking about the models, and I'm continuing, but the model release cycle is every two months, then you don't have a way to evaluate the model at the full length of their capability before the next model release cycle. So what Noam is saying, he's saying that they're getting faster and faster at developing these models because they're using AI to develop the next model. We just talked about the fact earlier in this episode that they are getting much better at writing code and they can do it faster and they can do it at a very large scale for very long periods of time, which means they are now helping develop faster AI.

42:27We're not at recursive self improvement yet, but we're making steps in that direction, which means the labs can release models much faster, which means you don't have enough time between one release cycle and the other to fully test the model that is now much more capable than the previous one. So you have a very serious gap that is growing. On one hand, the model is becoming smarter and can hide what it's doing, so you need more time to evaluate it. On the other hand, the time you have is shrinking dramatically because the next model is coming at a shorter time frame than the previous model came through.

42:59Now, how much faster is it going to get? So we already have the problem, right? The problem is already there. We're seeing it from the Hugging Face incident and from the Australian government incident and from everything else that's happening, the problem is already there. But we're saying it's going to get faster and faster. When asked how much faster is it going to get, Noam said that he is not completely sure. He thinks it's not going to be crazy faster, let's say, in a year from now. He's saying, and I'm quoting, I could see things going 3x faster. That is huge. And the reason he's saying that is huge is because it's already really, really fast and it's already very hard to control.

43:33But then he added, maybe there could be an overnight intelligence explosion. I don't know, maybe we don't see 3x speed up. Maybe it's 50 % speed up. So this is the guy. He is one of the lead researchers in open AI. He's the one that developed and started this whole concept of thinking models. He is in the trenches. He understands exactly what's going on inside of the most advanced or one of the most advanced labs in the world. And he doesn't know. And I want to end with two quotes and then obviously summarize what I think, which I believe is pretty obvious right now. But Noam was talking about another engineer that actually worked on the Navier Stokes math effort that they solved earlier this year.

44:14And he said, talking about this other person, he used to say, it's really hard to predict where AI would be in 12 months. If somebody asked him where things are going, he would feel comfortable making predictions for the next 12 months. But beyond that, he was just like, I don't know. Now he's saying he just doesn't feel comfortable making predictions beyond three months. So this is Noam Brown talking about a colleague from OpenAI that is working on advanced AI that is not willing to make predictions more than three months out. He also said, and I'm quoting, even people inside the top labs won't predict much past a few months out.

44:50So these are the people who are making the decisions of where the world is going. These are the people who are making decisions that may put at risk everything we know. And I'm not talking about Terminator coming and destroying the world. I'm talking about systems that can take over the internet. What do we do then? How do we control everything that we do in the world today if the internet and even secure systems are taken over by a swarm of agents that we don't control? Nobody knows. Now, right now, we have a thousand agents that are hacking hugging face? What if we have in two years, in three years, in five years, in 10 years, what if we have in two to five years, a swarm of millions of agents that can collaborate in ways we cannot even imagine and take over basically anything they want?

45:40They can control our entire digital lives, which right now connects very, very closely with our physical lives. This is very serious and very real. And the fact that they don't know, and again, the fact that the leading, one of the leading researchers in one of the leading labs is saying out loud that they don't know and they won't make predictions more than a few months out. And yet we are having a government that is saying, oh no, let's run faster and start the AI force and change it to whatever name. We didn't talk about this, but the president wants to change the name of AI to SI for super intelligence or stuff like that, which again, just stupid.

46:20I don't want to talk about this at all. And again, I'm not being political about this. I'm not taking sides. I'm just thinking changing the name doesn't change anything. We have a very serious problem with the leading researchers in the leading labs do not know how to align these models. They do not know what happens once we grow the scale and they're not willing to make bets on what AI is going to be in a few months. and Noam said that they underestimated the AI in the Hugging Face incident, which means they're going to underestimate it in the future as well because it is growing in its capabilities at a very high pace.

46:53So my goal is obviously not to make you not sleep at night, but my goal is to make you aware of what I think is the current situation. I personally think it is extremely alarming. That being said, there are people on the other side who think completely otherwise, leading people such as Mark Zuckerberg, such as Jensen Huang, who are both saying, no, these are completely exaggerated. The risks are not that high. We're actually pretty safe and we can figure this out altogether while keeping running at the pace we're running right now. Again, on a personal perspective, not taking anything away from Zuckerberg and Huang, they have vested interest in pushing this forward from a financial perspective.

47:32But when I'm connecting the dots of everything that I'm seeing, and especially this particular interview with Noam Brown that is inside and one of the leading people in this, and he doesn't know what's going to happen and he's making bets based on gut feeling on what he thinks may or may not happen. I don't think that's enough for the risk that is at stake right now. And I really hope that the labs and governments figure out together how to collaborate, together with academia, together with anybody else that needs to participate in order to make the right decisions, to make this much safer than I feel it is right now.

48:04Now to the rapid fire items, and I'm going to continue with one safety and security item just because it is relevant. And then we're going to go to the interesting releases that happened this week. But it just became known that on July 25th, 2026, Hackatron AI, which is a group of hackers, gained remote code execution, RCE, and administrative access to OpenAI community forum, community.openai.com, which is hosted on this course. They did this by exploding an OpenAI single sign-on identity flaw that they took over employees' ChatGPT and Codex accounts, which demonstrated that they can access OpenAI's internal GitHub by means of using AI to find loopholes in their existing systems.

48:46Now, these guys are actually good guys, and they reported the initial vulnerability to OpenAI on July 25th. OpenAI confirmed that they fixed that particular loophole within 14 hours. But what it's showing you is even these most advanced labs that have models that nobody else should access because they're really dangerous and not online, etc., etc., can be hacked. And with AI tools that these labs are delivering. And so we're in this loop where things are going to get really weird and, from my perspective, really scary. So now let's talk about releases that happened this week. We already talked about Grok 4.7.

49:17It is a really powerful and really cost-effective model that was released by SpaceX AI. It has dramatically better agentic encoding capabilities compared to GroK 4.6. And as we mentioned before, it is also cheaper and safer than the previous models. OpenAI just released a new version of ChatGPT Voice that has been significantly upgraded and most importantly, integrated into a lot of other capabilities, including third-party plugins like email, calendar, Slack, and while still being powered by the most advanced models inside of ChatGPT. So you can control it with ChatGPT Astra, Sol, and Luna. What does this mean?

49:53It means you can start using voice to work in your work environments and talk to it just like a personal assistant, and you will be able to take functions across everything you connect to it. Again, right now, it's not really everything, but it is becoming more and more connected and more voice-oriented, and it is amazing from a usability perspective. Staying on OpenAI, it seems to be their reports that they're planning to launch a new tier of ChatGPT Pro Max that is going to be priced at$500 per month, which is going to be obviously much higher than the highest right now, that is$200 a month plan.

50:26so it will allow to offer bigger, faster models for more amount of them, almost without limit, I assume, if you're at the$500 a month plan, even though people through the API can consume tens of thousands of tokens in a day, I think most people are going to go with this plan or not these kind of people, and it will allow you to do a lot more for, again,$500 a month. Now, the way OpenAI framed it, and I'm quoting, to make sure our current users have incredible experience and continued access to Astra, we're going to pause subscription to our$200 a month plan. These put the most strain on our system and we wanted to take the smallest step to allow us to continue giving the broadest access possible.

51:06So basically what they're saying is the$200 a month plan right now gives too many tokens and then people are getting pissed that they're getting blocked. But if they're going to give more tokens, then it's going to slow everything else down. And so they're most likely going to launch a more expensive plan for the people who want to get more access to the more advanced models. From OpenAI to Google, well, according to the information, which has always been a very reliable source on what's happening in the world and definitely in the AI space, Google is nearing the release of Gemini 4. DeepMind's new leader, Karai Kavukulu, is stating that it could arrive much earlier than the end of this year, which basically means in the next month or two.

51:42And it's now in final post-training refinement. Now, if you remember, Google has not released any major model for a very long time now. They scrapped the planned Gemini 3.5 Pro update that was announced by Sundar Pichai on May because it just wasn't good enough to compete with the Frontier models. And they decided to go straight for Gemini 4. And instead, in the past few months, they released smaller, faster, speedy, cheap kind of models to compete in that arena instead of competing with Frontier. It will be very interesting to see if Gemini 4 can really compete with the new latest models from OpenAI and Anthropic.

52:18And again, we may find out in the very near future. Google also announced this week Gemini 3.8 Flash TTS and Gemini 3.8 Flash Lite TTS models, which are really highly capable text-to-speech models. And these models are now available across Google AI Studio, Gemini API, Gemini Enterprise, Gemini Notebook, and Gemini and Google Vids. All of them are now using the latest voice models that can speak and understand 130 languages and dialects with natural language prompts and conversation, same direction that we hear from OpenAI in delivering the new models. Everybody's going to go in that direction and we'll be able to speak to these models freely and regularly in any language and work with them across more and more capabilities as these developments gets deployed.

53:04I have been typing probably 5 % compared to what I've been doing just a few months ago. I just don't type anymore. Just speak to these models or voice type what I need to voice type. And it's been working extremely well. It is definitely delivering much faster results than typing. Another announcement from Google this week is that they have integrated the advanced Gemini Omni 1.1 Flash AI model into Google Vids, which allows making 1080p video creation accessible for free to anyone with a Google account. And what they're stating is that this model provide very accurate creative controls and transparency, and you can have granular control and extending scenes while maintaining visual context, lighting, and character consistency, setting precise clip duration, and upscaling existing AI clips to 1080p.

53:50So very capable model available for free if you just have an account, which means you can now create more and more video models at a very high resolution without paying a lot of money to different companies to do that. Speaking of visual creation, Alibaba just launched Quen Image 2.1, which is a really small model. It has only 7 billion parameters for an image model that's really, really small. And they're claiming that it is, from a results perspective, aligned with much larger models that come from other companies. It supports up to 10 reference images simultaneously, and it enables complex tasks like group portrait compositions, virtual try-ons of different garments and elements on a single individual or interior design with different references of furniture and so on.

54:33So again, this is not something new. This has been available on the recent models from Google and from OpenAI, but having this in a Chinese model with a much smaller footprint is something new. Now, two interesting announcements from Anthropic that are not releases, but are very interesting. One is the fact that Anthropic increased its five-hour usage cap by 20 % for Pro and Max and team subscribers, while at the same time introducing Opus 5.5 that we talked about before. So one of the most annoying things, if you're a heavy Claude user like I am, was the five-hour limit that you hit if you are running multiple things in parallel, and then you have to wait until the five hours are over in order to continue and working.

55:14Well, that has been extended by 20 % now, which means you can do more work every single day without hitting those limits, which is absolutely awesome. But the more awesome thing, or the more interesting thing that Anthropic announced this week is that they have made a breakthrough in biological research and scientific discovery in that space. And what they shared is that by utilizing approximately 950 agents running in parallel, burning through 210 million tokens, they identified the array associated reverse transcriptase known as ART or ART, which is a system in which allows to edit DNA. And they've did it in merely 21 hours of concentrated effort by these agents.

55:54What they said is the same analysis would have taken humans, experts, weeks or months to do, showing again the role, in this case, biology compared to what we've seen from OpenAI with math. So what this is allowing is similar to the CRISPR mechanism of actually editing genes inside of bacteria and living cells. Now, the core reverse transcriptase enzyme was previously known, Claude Juny's contribution was that it's now spotting and overlooked links to this DNA array and associated proteins. So again, it's not something that wasn't known before. It's just that this model was able to connect it to more things and understanding how to actually use them in real life.

56:34Now, this is the first major real discovery from an AI-driven biology lab. This lab was established just this year, so it's not like it's been running for a few years. Now, skeptics are saying that similar systems have existed before and it's not such a big discovery. I am not the one that's going to judge. I know exactly nothing about this. This is way above my pay grade. But the fact that these labs are investing in biological research is the thing that I care about the most because it will help us hopefully find cures to real diseases and real issues that exist in the planet right now, which is a great thing and definitely one of the things I want to see AI pursuing more and more.

57:12Now, speaking of releasing new capabilities and running faster than expected, Meta's Muse AI Assistant is gaining huge traction right now. So during the first 12 days of its existence, Muse recorded 1.8 million iOS downloads in the US and Canada. This is outpacing the performance of Chachapiti initial launch on the App Store of 1.3 million in the same comparable timeframe. So that being said, we need to remember that when Chachapiti was launched, the AI boom was significantly smaller than it is right now, but it's still very impressive. Globally, Muse achieved 2.8 million total installs in the first 12 days.

57:51Just to put things in perspective, Anthropics Cloud in the same timeframe had 400 ,000 downloads and Grok had 200 ,000 downloads. So this is in a completely different scale. That's it for this week. I hope you found this not too alarming and yet very interesting. If you have, please drop me a note. I would love to hear what you think about these kind of episodes. It is very easy to find me on LinkedIn. If you're finding this podcast helpful, please rate us on your favorite podcast platform, whether it's Apple Podcasts or Spotify. And please share this with other people who can benefit from this.

58:22It's really a very small effort. There's a share button on your app. And if you're not driving, it will take you less than a minute to share it with a few people who can learn from it as well. I know I say this every week, but if you haven't done this so far, please do it now. Just pull up your phone, click the share button and share it with a few people who can learn from it as well. I would really appreciate it. We'll be back on Tuesday with another how-to episode that's going to teach you how to implement AI in your business or in your personal life. And until then, enjoy the risk.