332 | Bill Gates: “Letting AI Companies Self-Regulate Is Insane", the White House does it Anyway. 20 Announcements on OpenAI Dev Day + Gemini 4, Sonnet 5.5 and Muse for business Launches and More Important AI News for the Week Ending October 2, 2026

0:00Speaker Hello and welcome to a weekend news episode of the Leveraging AI podcast. The podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Matis, your host, and this has been one of the craziest weeks maybe ever in the history of reporting AI news and things that happened.

0:21Speaker I probably could have recorded a full episode every single day this week, and in some of the days more than one episode, and still have enough to report. So I assume this is going to be a longer episode. I will try to make it as concise as possible. But we do have a lot of really important things to cover, and I wanna cover at least the important things, and there's a lot of medium importance things that may be, pushed to the newsletter, which by the way, you can sign up for and get all that information as well.

0:47Speaker But what are we going to talk about? So we're going to start with AI safety, and we're gonna look at it from the perspective of two different interviews that happened in the past week or so, both of them with Ezra Klein. In the first interview, he interviewed Jensen Huang, the CEO and founder of NVIDIA, and in the second one, he interviewed Bill Gates.

1:06Speaker And these are very, very different from one another. So we're gonna cover AI safety and where it stands right now through the lens of these two highly important people in the world of technology, and AI overall. Then we're going to talk about everything or a lot of what OpenAI has announced and released on their Dev Day that happened this week.

1:27Speaker So we got a lot of interesting releases in a very clear direction from OpenAI and where they're going with their AI initiatives. So that is going to be our second deep dive. Then we're gonna go back in the rapid fire to a few safety important items that could have been on their own a deep dive item, but we're gonna cover those.

1:45Speaker Then we're gonna talk a lot about a lot of new releases that we got this week, including Gemini 4 and Claude Sonnet 5.5 and GPT Sol 6.1 and a lot of models from Meta and a huge amount of new releases that happened this week. And then we have time. There's a bunch of other stuff that we can talk about that happened in the government and international relationships and Chinese models and stuff like that.

2:07Speaker So like I mentioned, a lot to talk about. Take a deep breath, grab a cup of coffee, or if you're driving, drive slow. Let's get started. The first topic, as I mentioned, is AI safety through the lens of Jensen Huang and Bill Gates. Ezra Klein, who is an incredible interviewer and has a really solid podcast, has interviewed both these people.

2:31Speaker So Jensen on September 23rd, and then Bill Gates on September 29th. Both of them on the topic of AI safety and other topics, but the core thing was AI safety. And they gave completely opposite views of where the world is and what the next steps should be and what the real risks are, and so on. And I think it will give you a great view into what two people who definitely know one or two things about technology and how it works and how it impacts the world, and how they see the current situation.

3:02Speaker Now we're going to connect the dots of these two interviews with some other things that happened this week because they are closely related, and I think that's gonna give you a great view of where we are right now, what people think, and you can make your own decision. I will obviously share with you my point of view.

3:18Speaker As if you've been listening, you already know what it is. But I definitely leaning more toward one side than the other in this argument. But let's dive into this. So first of all, let's talk about Jensen. Those of you who don't know, Jensen Huang is the CEO and the founder of the largest company in the world today, a $5.4 trillion Nvidia that is really the engine behind this entire AI revolution that we're going through, right?

3:43Speaker Without their GPUs, none of this would have been possible, and they are involved in literally everything you can imagine, from creating the hardware, to building their own software, to creating architectures, to everything that you can think of that is involved. So Jensen knows a lot about what is happening in the AI world.

3:59Speaker He has always been very optimistic about where this is all going. He has always been very dismissive of the risks, and I always wondered whether that is because he has a company to run that makes money when people buy more GPUs, so he needs, people to buy more GPUs. And if this, these GPUs can destroy the world or humanity, it's not the best marketing for his company.

4:22Speaker That being said, I must admit, after listening to him many times, and especially in this interview, I think he actually truly believes everything he's saying. I think he's such a big optimist of technology will do well for humanity that he is tinted in the way he sees things and where it is going. Now he calls himself a responsible optimist, and he's saying that basically summaring up everything he said in that interview, and it's definitely worth listening to because they're diving into a lot of segments.

4:51Speaker But he is saying that the risks are much exaggerated. He's saying that solving for risks is purely lines up with good product management and good security and just being a good company. And here's a quote from him, about OpenAI and Anthropic and releasing potentially problematic models. He said, "If they believe they're out of control, they don't ship products until they're in control."

5:18Speaker He also said related to the fact that both Anthropic and OpenAI said they're under immense pressure to push forward because of the race. He said, "Nobody's putting the pressure on them." So that's them about how they feel in the race, saying they're creating their own pressure. And he's basically saying that every risk there is in the AI space right now is a simple engineering problem that can be solved.

5:44Speaker Basically, he's saying, "Look what we've built, uh, in Nvidia. Everything we're building is 100% safe because we're putting the right safeguards in place, and everything else is problems that can be solved." So every, any issue that OpenAI, Anthropic, or any other model is putting out to the world can be resolved with better engineering and better process and product controls When it comes to controlling these companies or putting in place rules and regulations, Jensen is very much against additional regulation.

6:14Speaker He's basically saying that the current rules are more than enough to control this. He's saying that companies will not take the liability risk of doing things that are unsafe because then they will get sued, and then they will lose everything they have if the lawsuit is big enough. And so the liability and current rules should be enough to force companies to act responsibly and release the right things.

6:38Speaker Later on, you'll see that Bill Gates agrees with me. I think this is a very stupid argument. It is a very stupid argument because the same argument could have been made about drug use. So when you are a drug company, a pharma company, and you wanna release, a new drug to the market, you can say, "Well, the liability is good enough.

6:55Speaker We don't need any regulations to regulate how, new drugs are being released to the world." The same could have been said about the FAA. We don't need the FAA. Companies can build, airlines and airplanes however they want, and if they will do something bad, then we will sue them because there's liability laws to, support that.

7:10Speaker And so when millions of people will die, then we can go after the company who put in place the risk. Again, that makes absolutely no sense to me as a argument, and I'm surprised with Jensen making this kind of argument, right? It's assuming there is a risk and saying, "Oh, now X number of hundreds of people or tens of thousands of people or millions of people are gonna get harmed," then somebody will sue OpenAI and Anthropic, and that would be the end of it.

7:40Speaker Well, that's gonna be after all these people gets harmed. And so this argument makes absolutely no sense to me, but this is what Jensen believes, or at least this is the statement that he's releasing, to the public. This episode is brought to you by the Multi-Agent Orchestration Course. It's a course that I have been teaching personally since April of this year, and every single one of the cohorts that we have opened have sold out.

8:05Speaker And the people who have been through these courses made incredible transformations for themselves, switching jobs, starting their own companies and so on, and definitely for their businesses by being able to build really complex processes by following our methodology and our infrastructure settings that we teach during the course.

8:21Speaker The course is four sessions of two to two and a half hours each over four consecutive weeks, and through it, you can go from knowing a little bit about how to use AI and using it in ways that most people are using it today to learning how to build really complex, sophisticated systems that can automate more or less everything in your business.

8:37Speaker The current cohort that is open starts on November 2nd, and it's gonna go on November 2nd, November 8th, November 16th, and November 23rd, and we are almost sold out for that cohort as well. It is very likely that we will not open another cohort in December, meaning the next one after that is gonna be in January.

8:53Speaker So if you still wanna learn how to build really incredible automations and learn everything that I've learned by doing this every single day for the last few years, don't miss this opportunity to come and join us on November 2nd. There's a link in the show notes, And if you're going to use the promo code LEVERAGINGAI100, you're going to get $100 off of the price that is on the website.

9:14Speaker Like I said, there's a few seats open, but they are selling out fast. So if you're interested in doing this still this year, go and click on the link right now on your phones and complete the registration and the payment so you can guarantee your seat. I'm looking forward to seeing you there. And now back to the episode Now, he is all for third-party auditors, but he thinks that this is something that the companies should do on their own, and again, shouldn't be set by the government or an external body.

9:41Speaker If you think that you need your stuff to be audited, get the right auditors in place and do that. Now, the main thing that he argues, as I mentioned, is the fact that the companies themselves should be in charge of releasing safe products. He said, and he's using a specific number, that at Nvidia, 80% of engineering efforts goes to verification that the technology works according to how they want it to work.

10:03Speaker While in the labs, the percentage of compute and talent that goes into making the models safe versus making them faster and better is a much, much smaller percentage. So his argument is basically saying, if the labs want to make it safer, all they need to do is switch their priorities. Switch 80% or a large percent of your engineering to making these models safe, and only a small component to make faster, better new models, and we should all be fine.

10:36Speaker I obviously do not know if that is right or wrong. What I do know is that the current situation is out of control. I also do know that in labs, while testing with security measurements and so on, these models are able to get out of their sandboxes and hack actual companies in extremely sophisticated ways while not intended to do any of those things.

11:01Speaker So will additional measurements solve that problem? I don't think so. I definitely don't think so. As these models get smarter and better over time, way faster than our ability to control them. There is a interview I always go back to in my head with Yuval Noah Harari, the author of A Brief History of Humankind and a bunch of other books.

11:22Speaker His latest one is called Nexus that talks about, uh, the information networks from the Stone Age to AI and how AI will impact the world. And he's basically saying that we are able to control AI right now, and he actually said it like a couple of years ago, because we are an adult, sophisticated human being, 40, 50 years old, and the AI has the intelligence of a five-year-old, so we have full ability c- to control it and understand what it's doing and be able to enforce whatever we want.

11:49Speaker But he's saying that not too far into the future, and again, this was said two years ago, it will be the other way around. We are going to be the five-year-old, and the AI is going to be the sophisticated advanced 40-year-old, and there is no reason for us to think we can force our will or our mind or our way of thinking on it, because the gap in intelligence is gonna be so extreme that it will probably work the other way around.

12:13Speaker So building better engineering doesn't work when the engineering is trying to control something that is X number of times more sophisticated and intelligent than humans are. Now, by the way, he was also asked about jobs, and so I'm moving away from security for a second. But it's very important to also learn about these things because in my perspective, this is one of the biggest fears that I have, is huge amount of unemployment that can be driven by AI.

12:39Speaker So Jensen actually thinks that's not going to happen. He believes that AI will change most jobs that happen today, but will not eliminate jobs. He actually thinks that we're gonna end up with a net gain in jobs. now he gives two reasons for why he thinks there are going to be more jobs and that jobs are not gonna be lost.

12:57Speaker His first one, he's saying that every job has a purpose of the job and then the set of tasks that make up the job. And what he's saying is that the set of tasks may change, but the purpose of the job will stay the same So the example he uses is radiology, and he's saying that radiology, now AI has the ability to understand scans better than humans.

13:19Speaker Did this eliminate radiologists? He's saying, well, absolutely not. If anything, it's allowing hospitals to do more scan to service more people, so they need more radiologists to go through what the AI is doing. The other thing that he's saying is he talks about human ambition as something that is the driver behind everything that we do.

13:38Speaker And he's saying that human ambition is not going to go away. We're just going to use AI as a tool to grow and use it to fulfill our ambitions and grow bigger ambitions. I have an issue with both arguments. The first argument has to do with the fact that I do agree that currently the throughput that AI enables radiologists and in many, many other cases is going to be helpful and may require additional radiologists.

14:05Speaker It is the same thing with software engineers, right? Everybody's saying, "Oh, AI cannot write code, and most can write code better than most humans." And yet we have more need for software engineers right now because they can do a lot more. It is a temporary situation because we will get to the point that the AI will make the radiologist redundant.

14:23Speaker They will not need to review what the AI is outputting because it will always be more accurate and better than what the humans can do. The same thing will happen with code. Humans will stop writing code, period. I don't see any other solution. Humans will stop reviewing code because they will not be able to review the amount of code that AI is generating.

14:41Speaker That's already the case in many, many different companies. And so I think the engineering part of it is going to survive, and that's something Jensen talks about. Like software engineering has the tasks of writing code and has the tasks of engineering. But the reality is there are a lot less people doing the engineering and a lot more people writing the actual code in the world today, and that will always be the case.

15:03Speaker And the more extreme situation is when AI is good enough in the engineering part, then it will make that redundant as well, so we won't need humans to engineer the software either. AI will be able to do it very well. I'm doing this in various levels of scales in several different companies that I'm involved in.

15:18Speaker Developing robust, sophisticated software while AI is doing the engineering, AI is writing the code, AI is doing the deployment, AI is doing the testing, all of that with me just driving the ideas and making decisions, but I'm definitely not doing the engineering. And so I have a strong disagreement with the long term of that argument.

15:37Speaker In the short term, I do agree that that is the case. On the second argument I will say two things. One is I think the vast majority of people who are employed in the world today are not doing this because of their personal ambitions. They're doing it as a means to make a living and put food on the table for their families.

15:55Speaker There are a very small percentage of people who have jobs today who are driven by their personal ambitions when they go to work every single morning. Jensen is definitely one of those people, but I think he's been so far into this universe of, "I wanna drive this path, and I'm gonna do everything to help other people work with me to go down that path," that he doesn't understand that the vast majority of people are not like him.

16:20Speaker And therefore, I think that the ambition of humans to do stuff, yes, it is there, but I do think that it is not going to be the thing that's gonna eliminate jobs. And if you would listen to the interview and you listen to this section, you see that he's struggling to come up with solid arguments. He's stuttering a little bit, and he's going back and repeating sentences, and he's not as coherent and convincing as he is when he's talking about a lot of other stuff in this interview.

16:46Speaker So I don't know if he's personally not conf- convinced in this and he's saying this because, again, he has to say this because that's how he grows his business, or he believes it, but he cannot rationale or give actual strong arguments to why he believes, uh, this is the case. So bottom line of what I think about this interview, A, worth listening to.

17:08Speaker A- again, Jensen knows about this topic way more than I do. Uh, he is definitely a techno-optimist to his core, and he believes that everything will be fine just because that's the way it has always been, and just because humans are ambitious and resilient, and we're all gonna figure it out, and it's all gonna be okay.

17:28Speaker He was not able to make any arguments that actually convinced me that's actually the situation Now, to be fair, when Jensen was talking about engineering and how that can solve the problem, I'm like, "Okay, in, in theory, I, I get it, but in practicality, I don't understand how that's going to work."

17:44Speaker Well, just five days later, NVIDIA announced a new kind of platform, and that platform is now what NVIDIA calls the Open Agent Safety Platform. Now, in the interview itself, Hong, said that agents cannot monitor their own sandbox, and this is why they're breaking out, and they need, and I'm quoting, "A whole bunch of watchdogs."

18:04Speaker And I think this is what NVIDIA just released, a bunch of watchdogs for AI agents. So what this product is, it is an open source software that runs each agent in a sandbox that is isolated at the operating system level, and it is built in a way that allows a different machine, a completely separate piece of silicon, to monitor the sandbox and what the agent is actually doing So the idea is you have a bunch of models and agents and so on that are running on one piece of computer, and then there's a different computer that has a completely different task, and that task is to enforce the policy of the other task and the other computer at the hardware level So I have a feeling that when Jensen was

18:50Speaker mentioning better engineering in the interview, he was hinting that's the direction they're going, and it might be a solution. I don't know if it's the solution. It is definitely a way for Nvidia to sell more compute, because now you will need to buy the second part of the compute in order to monitor the first part of the compute that actually runs the models and everything.

19:11Speaker Does this mean it is a full fail-safe mechanism? I don't know. I think it is definitely better than not having, uh, that in place. A, kudos for Nvidia for understanding that this is a real need and moving in that direction. B, kudos for Nvidia for another way to make a lot more money than they're already making right now.

19:31Speaker And C, I'll be very curious to see how this actually operates in real life and whether that is really putting and enforcing the guardrails on system A by monitoring it with system B And then came the Bill Gates interview. So the interview with Bill Gates is dramatically different than the interview with Jensen, and it is way scarier.

19:56Speaker And again, both these people are highly capable, extremely smart, crazy successful when it comes to im- developing and implementing and deploying technology across the globe. Now, the reason Ezra Klein is surprised is because Bill Gates went from a techno-optimist to being really scared, and he's following up on an essay that he released on August 26th that called "The Turbulent AI Era Is Here.

20:19Speaker The Choices We Make Now Are Critical." in this paper, Gates sounds the alarm across more or less every aspect of society you can imagine. Uh, he said we need time to prepare for the social, political, and economic upheaval, and he gives, several different really scary warning shots on different aspects, such as many jobs will disappear forever, so a huge job loss.

20:42Speaker AI will empower people and perhaps AI to do more harm. So he's basically stating that there's always been bad players, but now they can do a lot more bad things with this capability. He's talking specifically about cybersecurity and bioweapons and chemical weapons that can be easily developed or a lot easier developed with AI He's talking about potentially AI, uh, getting out of control and doing stuff that we completely do not intend.

21:07Speaker The exact quote is, "AI models could begin to act against our interest, and we could lose control." He's talking about AI could stunt our kids' development and replace human relationships and a lot of other things. So this is the paper from, Bill Gates that was released at the end of August, That drove the surprise by Klein when he's saying that he was surprised by Bill Gates, that in just 2023 said that the risks are real but manageable.

21:31Speaker So that was not too long ago. And he asked why this is the case and why he changed his direction, and why he's investing so much right now in this direction, why he's risking everything he earned in his life and putting his name on the line, putting himself out there, making these claims and pushing towards more regulation and security on AI.

21:53Speaker And Gates said two things. He said, first of all, that the industry promised to engage all society when it crossed specific dangerous thresholds, and then when these thresholds were crossed, nothing happened, complete silence. Two thresholds specifically that he's talking about that were risks that were said, "Okay, when we get there, we will stop and ask everybody for assistance to figure it out."

22:14Speaker One is cyber. So he's saying it's very obvious that as of 2026, these models can now do things that are very, very dangerous when it comes to cyber. He's saying models are finding bugs in code humans checked regularly for 20 years. We know this from multiple open source, environments. The biggest ones It's probably Linux operating systems.

22:33Speaker So Linux has been an open source operating system that runs most of the servers in the world, and AI systems are able to find multiple vulnerabilities in this code. And so again, this is code that millions of developers have viewed, or at least a lot of developers have viewed millions of times and haven't found these, uh, vulnerabilities And the other is, bioweapons.

22:52Speaker Again, a pathogen worse than smallpox that once needed a nation state to be able to develop, and now a small group of the wrong people can learn and create it because AI can assist them in the process He was asked specifically about the topic of do we need more regulations in place in order to define what AI can and cannot do and should and shouldn't be released in specific scenarios.

23:14Speaker And he, specifically was asked about David Sacks and Huang and, and President Trump. And Bill Gates was very, very clear, uh, that his opinion is very different than Jensen. He said, "Just like we do not trust lawsuits for drugs and airlines and cars, we shouldn't definitely do this for AI." He's basically saying AI is the most dangerous technology we potentially ever created, and if 100 million people will die, do you think lawsuit is gonna be the right thing to address the situation after 100 million people die?

23:43Speaker His view is basically that if we do not do something extreme right now, this might be the outcome, that millions of people or hundreds of millions of people or billions of people may die because of outcomes of things the AI may do either on its own or when being used by bad actors if you want the exact quotes, when he was asked about, uh, the fact that the liability law can be put in place, he basically said, "I almost can't believe you're asking that.

24:09Speaker This is the most dangerous thing that humans have ever gone near." And then he also added, "You can't rely on the industry to self-regulate here. it's just insane." And what he's saying is there has to be a very strong, completely tight supervisory layer with clear safeguards that are required for every model on the planet.

24:31Speaker And when saying every model on the planet, he also means open source models, meaning he's saying open source should stay open source, but must run on a platform where governments or third-party bodies, can put and verify the safeguards in which they are running. So he's basically talking about a solution where AI runs within a controlled environment for every model in any nation, in every situation, always or we're putting humanity at risk Now he's saying that loss of control over AI is, he believes in the short term, less risky than bad players, the wrong people are misusing the tools to do really bad things with it, and I tend to agree.

25:11Speaker I think we got to the point that AI can now break into more or less any system you want, and like Gates is saying, it also allows you to build bioweapons and chemical weapons in a way that was not done previously. And just think about the right group of people deciding to plan the attack of all times, the next 9/11, And they can shut down the entire infrastructure of a state in the US.

25:35Speaker Forget about the entire United States. Think about an entire region in the US losing internet, power, water supplies, communication channels, all of that, while something really bad is being released into the environment, whether it's biological weapon or something of sort. Our ability to deal with that situation right now is zero because nothing is ready for that.

25:58Speaker Our supply chain will, will break. There's not gonna be water to drink. There's not gonna be food in supermarkets. Because nobody will be able to coordinate anything. This is what he's basically warning of. He's saying this is possible with today's tools, not something that will be developed in the future.

26:12Speaker This is today's tools in the hands of the wrong people Now, I am way more aligned with Bill Gates' opinion and thoughts and feelings than I am with Jensen. And a great example this week came, and I told you I'm gonna connect the dots with some other things, is that Anthropic just released a paper in which they reviewed thoroughly the safety situation of GLM 5.3, which is a Chinese model from Z.AI called, uh, Zipu in, China.

26:41Speaker And it's an open weights model that anybody can download and use however they want. Now, on the benchmarks that Anthropic has put it through, it was found to do basically anything they wanted without too much effort in order to do horrible things when it comes to cybersecurity. It is an extremely powerful model, so it was able to build working exploits in 50 out of 410 attempts to break into just Chrome.

27:09Speaker Think about this as the browser that everybody in the world is using right now, or the vast majority of the world is using right now, and this model was able to break into it and find, and exploits in that, a large percentage of attempts To put things in perspective, Claude Mythos, which is the first model that we talked about with these kind of capabilities, did it 56 times out of 410, but it was not released to the public because of that.

27:31Speaker So this is a model that is almost at the same level. It's 50 versus 56, but it is now available to anybody who wants to download it and play with it and do whatever they want with it Just as an example what this can allow it to do, a researcher used it in about a day was able to chain together a bunch of vulnerabilities and was able to use this situation to read files off the visitor's computer.

27:55Speaker So basically every computer on the planet that is running Chrome, anybody who gets access to this model, which is anybody who wants, it's ava-available to download right now, can now go, and if you're using Chrome, access your computer and get access to all your files. This is where we are right now.

28:10Speaker Now, it comes with basic safeguards that is supposed to prevent it from actually doing it. But what they found is that with basic attempts and workarounds, they were able to actually make it do what they wanted In 64% of cases, they were able to make the model do what they wanted just with a false cover story, telling it that they're researching and doing these kind of things, and the model allowed them to do it.

28:31Speaker They were able to make it do it in 92% of the requests by pre-filling its reasoning, so basically tricking the model to understand, thinking it's doing one thing when it's actually doing something else. And Using obliteration, they were able to convince the model to do what they wanted 100% of the times So if you're asking what the hell is obliteration, because I did not know what that is, because this is an open source model, what researchers can do is find why and how the model refuses specific requests.

29:00Speaker So basically, researchers found that the model's refusals come from a specific internal pattern. What they do is because they have access to the weights of the model, they find where that pattern resides within the model, and they reshape the weights in order to ignore this segment, basically eliminating its need or wish or drive to refuse anything, and then it just works 100% of the time.

29:22Speaker The cost to do that was twenty-two hundred GPU hours, which comes out to four thousand four hundred dollars. So by investing less than five thousand dollars they were able to get to a point that GLM 5.3 will do every request they want it and break into any system that they requested it to break into NIST Center for AI Standards and Innovation on September 17th called GLM 5.3 the most cyber-capable open weights model so far, and they're saying it's only about four months behind the US frontier.

29:52Speaker Now, how far is it behind the US frontier makes absolutely no difference because it is good enough to do really horrible things across multiple systems around the world today, And it will do it every time it is being asked with very simple manipulation to the model itself So going back to Bill Gates' alarm, it is very clear that the models that are out there right now are, A, out of control, meaning the labs themselves cannot control them.

30:19Speaker B, the models that have been released, not something that will be released in the future, can be, in the wrong hands, a weapon that will do horrific damage to modern society because it can attack, like I said, major infrastructure, including the internet itself, communications, cell phones Water, power plants, you name it, any infrastructure, you know, stoplights, traffic, all of that is now controlled through computers.

30:43Speaker And if these computers can be hacked, then somebody can take control all of them. And if somebody is willing to spend the time to hack all of them and keep everything silent until the moment of truth when they wanna actually, move forward with the attack, they can do it, and they can do it with the tools that exist right now.

30:58Speaker And this is why Bill Gates is alarmed, and I must admit, I'm as alarmed as he is Now, by the way, he was also asked about jobs, and he's very much in my opinion that AI will take a lot more jobs than it is going to create Now, Ezra Klein offered the Jevons paradox that we discussed many times on this podcast that basically says that as things becomes more available and cheaper, they actually provide more opportunity and hence it actually generates more demand rather than less demand, and hence there's gonna be more jobs created.

31:27Speaker Gates' answer is very similar to the way I always answer this, which is it's all about how much the demand for something can grow. I said that multiple times in this podcast. If you're listening to this for a while, you know that I've been saying that in the very short term, as a company, as a personal thing, You have the opportunity to capture market share because there are companies who can use AI very effectively to move faster and actually handle this demand.

31:49Speaker But on the other hand, there is a finite demand for everything. You cannot grow all the companies around the world at the same pace by applying AI to all of them because there's just not enough demand for these things. And so things that have either no elasticity in their demand or once we consume that demand and shift from X number of people to X plus five number of people, then that's it.

32:13Speaker That's where it ends. And at that point, you have two options and two options only. Either to let people go and save your business and keep the rest of the company competitive and be able to compete, or not do that, and then your competitors will be cheaper, faster, better than you, and you're gonna lose the entire business.

32:28Speaker I do not see how that works. And when Ezra Klein said, "What about new positions that are gonna be created?" Bill Gates said what I'm saying, "Okay, tell me what these are." And yes, we don't know what they are. But the problem is, again, I've said that on the podcast many times before without listening to Bill Gates.

32:44Speaker Every previous revolution that we had has moved from we're doing manual labor to we're using our brain in order to do more sophisticated, complex things, allowing machines to do the manual labor. This is the first time that we have a technology that goes after our ability to think. It is doing that better than we do, and it will do that on every aspect of our ability to think as it evolves over time.

33:09Speaker So what of that will survive? I don't know. And then after that, we have blue-collar jobs, and robots are going for that. So same kind of thoughts, from Bill Gates as I've been thinking all along. He's very pessimistic as far as the future of jobs. What he suggests as a solution is a tax on AI. So he's basically saying, and actually his argument makes perfect sense to me.

33:30Speaker I never heard anybody say that before. He's basically saying, now we're paying taxes, basically taxing our work. So when you get a paycheck at the end of the month, the government or the local municipalities take a percentage of that in order to finance everything else in the economy, everything else in the state, all the services that we're getting and so on He's saying, "Why would AI be different?"

33:50Speaker We are taxing the job as it's being done. Instead of taxing the humans that are doing the job, tax the AI or the robots that are doing the job, and this is how we can finance everything that needs to be financed in order to keep people alive and put food on people's tables, and housing, and education, and roads, and everything else that has to happen within a modern country Now, what he's also saying that I found very interesting, he's saying that right now most of the investment in AI and the deployment of AI happens in rich countries and is not being implemented in low income countries.

34:23Speaker However, he's saying that pushing AI very aggressively on low income countries will be overwhelmingly good versus bad. And the reason he's saying that is it will give access to people to stuff they just don't have access to, even basic stuff. He gives examples of how many millions of people don't have access to doctors, period.

34:41Speaker They have a cell phone, but they don't have a doctor. So having an AI doctor that can help them in specific situation is incredibly useful to somebody who's just trying to deal with a situation where there is no doctor, in their area and definitely no modern facilities. And he gave a few other examples like that, which I strongly agree with.

34:58Speaker And as you know, Bill Gates Foundation has been pushing to help eradicate polio and other diseases and were really able to push down the death rate of babies in poor countries and provide a lot more agriculture and food to a lot of places. So he knows the situation in these countries better than most people in the Western Hemisphere.

35:17Speaker And so when he's saying this could be a very strong push for good in these countries, I believe him 100%. So now connecting one last dot to put this all together On September 29th, the White House hosted a two-hour closed-door lunch where they came out with a 308-word pledge that they all signed, and the participations were all the obvious suspects, the leaders of the AI industry, including Musk, Zuckerberg, Bezos, Jensen Huang, Dario Amodei, Sundar Pichai, Karp, Brockman, plus VP Vance and Speaker Johnson all were there.

35:54Speaker The people who signed it are Trump and six executives, Sundar Pichai from Google, Dario Amodei from Anthropic, Mark Zuckerberg from Meta, David Brockman from OpenAI, Elon Musk from SKI, and Jensen Huang from NVIDIA Satya Nadella and Jeff Bezos both attended but did not sign the actual accord that was signed in the end.

36:11Speaker Now, the accord is asking basically each company for four layers of security, internal controls, an internal team, an outside auditor, and an independent board committee, all focusing on security of AI in the future And it says it may make sense to write these steps into laws sometime in the future What is not addressed?

36:33Speaker there are no penalties, no public reporting, no triggers on when you have to stop and pause, no government review or third-party approval before release, no definitions of what that would look like, et cetera, et cetera, et cetera, right? It's a very vague statement The White House also is now signed an executive order that says it's no longer should be called artificial intelligence, but super intelligence, and it is more intelligent and will be more intelligent than humans, which is okay.

37:04Speaker How does that help us in any way? I'm not exactly sure. but that's a part of the government's statement as part of this process Now, the communication on X has exploded on both sides of this particular situation. many, many people said this is absolutely ridiculous. Some people obviously supported this.

37:21Speaker So as an example, David Sacks said companies are accepting responsibility Kara Swisher called it a pinky promise, which I somewhat agree Law professor Kimberly Weatherall called it deeply unimpressive. David Krueger called it Regulate without regulating By the way, another very interesting thing that happened out of all of this is that the person that Donald Trump pushed in order to make the actual announcement was Dario Amodei, which was the nemesis just a few months ago.

37:47Speaker If you remember, the government pushed him out. Dario said that the reason they're being pushed out is he will not kiss the ring of the leader, and now something has changed. I don't know if it's because he's kissing the ring of the leader or because he believes that he doesn't have a choice. I'm not exactly sure what's going on, but it is very obvious that they're trying to show that they're buddy-buddies again, or maybe not again.

38:07Speaker I don't think they've ever been. But they're buddy-buddies right now, and they are in a much better relationship, and Anthropic is back at the table, maybe even taking a front seat, uh, at the table in the race against China and in the collaboration with the government on this safety accord So how does this all come together and how does it impact your business and the world we live in?

38:27Speaker First of all, the accord that was signed by, again, the US government as well as the biggest leaders in the industry is basically Jensen Huang's positions in writing. That's it. There is no real anything to slow down, to put safeguards in place, actual real safeguards, or any kind of liability or any kind of real measurements that are clearly defined Now, that being said, I think it's a step in the right direction because it is the first time that you have all these people getting together in a room and talking about what do we need to do about AI safety.

39:02Speaker Now, the fact they came up with practically nothing at the end of these two hours is something else. I really hope this is just the beginning of these kind of conversations. I really hope that we will see more of that, and I really hope it's going to turn into actual practical steps that are governing this process, that will put the right guardrails in place and will enforce these guardrails, uh, before it's too late.

39:26Speaker like Bill Gates is saying, it will happen either before or after X number of millions of people either die or get hurt. The question is, are we going to do this before or after? And I really hope we find a way to do this before And now in the words of my really favorite Monty Python, "And now to something completely different."

39:44Speaker So OpenAI had their Dev Day this week, and they made huge number of really big, important announcements. lots and lots and lots of good things and interesting things that are happening. So I want to cover most of the stuff that they released, and definitely all the important stuff that they released, and give you an idea of where OpenAI is right now, where the technology is available right now, and where I think it is going based on that.

40:08Speaker So OpenAI launched more than 20 products at Dev Day on September 29th. The biggest thing that they focused on, what they call Dots, and Dots are these always-on agents that run 24/7. They run on powerful models. They run on their own computers in the cloud, so you don't need to have your computers running, and they can work together, they can collaborate with humans, they can work across different aspects of the work.

40:32Speaker They get access to your data and so on. They're built on GPT-6 Astra, and they are going to be available to the public in the next few weeks as they keep on rolling them out. Now, before I dive into that, I wanna give you a quick list of all the things that they released other than Dots, and then we're gonna dive into a few of them that I find more interesting.

40:53Speaker So I talked about Dots. These are always-on cloud computer browser agents. It is basically their version of OpenClaw or GrokBot or Instinct or each and every one of those. And it was very obvious, and I talked about this many time on the podcast, that I see zero chance that this is not gonna be in other platforms as well.

41:10Speaker Then they released GPT 6.1 Sol, which is a upgrade to GPT-6, which is near Astra capabilities for a fifth of Astra's price. By the way, going back to, GPT 6.1 Sol, there were a lot of conversations where going to announce Astra 6.1, but they scraped that release just a few days before Dev Day because of security risks, going back to our previous topic.

41:34Speaker They also announced Ultra Fast, which is a premium speed tier of up to 8X faster in Codex, which generates 300 tokens per second. And it is a lot more expensive, but it allows you to generate stuff a lot faster as well. They've also announced Privileged Intelligence, which is zero data retention with private safety processing, which is something that was a big request, and is a big need in the industry.

41:56Speaker They also announced Codex in the Cloud, so you can now run Codex on a remote machine, which means your device doesn't have to work, and multiple teams can share it together and run and develop things while no actual real computer is connected to that. More about that later on. They added more capabilities to Codex CLI, including the ability to steer it with voice and new agents capability that are running inside the CLI.

42:19Speaker They announced Code Review, Codex Security Cloud, uh, which is a scheduled repository scans that is fixing investigating, finding duplications, preparing fixes, all of that while you're basically asleep. They've announced Decisions API. If you've been following what's happening with, Jev, Jev has became a crazy phenomena, growing very, very fast, and all Jev does is it is a classifier.

42:42Speaker It will give you an answer related to a specific question based on specific categories, and it has grown dramatically, and Decisions API is basically Jev coming from OpenAI instead of from a third-party company. I do see that as a very logical step. It is not exactly like a regular large language model.

42:59Speaker It provides a lot of value in specific scenarios, and just like I said about the always-on agents, I think the same thing is gonna happen with the classifier models. They're gonna be a dime a dozen, and you can get them from anybody you want. They also announced a new version of the Agents API with computer use.

43:12Speaker So now Agents API can use agents to run on their own computer, in the cloud, whenever you want to do whatever you want. I'm already testing this, for something, and I find this to be extremely helpful, uh, when it comes to doing stuff on a browser when your computer doesn't have to be on They announced a broader integration with Bedrock on AWS.

43:29Speaker They announced a new thing called plugin extensions, which is allows new things to run inside of the ChatGPT environment, and they give examples from Figma and Adobe, uh, showing these kind of integrations on how you can run inside the ChatGPT environment while using other platforms fully integrated into that environment They've also announced plugins creation and discovery, so a plugin creator and a redesigned submission flow and a recommendation engine.

43:52Speaker Maybe this will finally be the, quote-unquote, "App store" of OpenAI to allow other people to share things into their environment. They shared sites that now can host plugins, so these third-party plugins can now run within sites, which are basically websites that you can build with ChatGPT. They have been around for a while now.

44:10Speaker They announced something that I'm very excited about, we're gonna talk more about this in a minute, which is MCP events, which allows third-party tools that have MCPS to trigger processes inside of ChatGPT. They allowed ChatGPT spaces, which I think is potentially the biggest announcement out of this whole day.

44:24Speaker More about that in a minute as well. They announced Pages, which are a subset of Spaces. We're gonna talk about this as well. They shared a bunch of other things with maybe the only thing else that I wanna mention is sign in with ChatGPT, which basically allows you to sign into a third-party tool and use your own ChatGPT tokens when you're using that ChatGPT tool.

44:44Speaker So this is to me a very, very interesting concept that allows the developers of third-party platforms who use AI to charge only for the platform itself instead of charging for the platform plus a premium on the tokens, and allows you, the user, to potentially choose the model you wanna run with and pay out of your existing subscription.

45:03Speaker So if you have an OpenAI subscription, whether it's $20, $100, $200, $500, you will be able to use that subscription with third-party platforms, and get the benefits of those platforms while using your existing subscription. So again, and there's a bunch of other stuff. So let's dive into the key things and the most important stuff that I think they released.

45:21Speaker The first one, as I mentioned, is dots. And dots are these always-on agents that can do a lot of things. They can do background research, they can read your data, they can engage with other dots, whether your dots or other people's dots. They can take actions, they can write files, they can create code, they can do testing, they can do all these different things.

45:41Speaker And they mentioned that they're already using it in more and more processes in OpenAI, and everybody's using their own dots across multiple different aspects of what they do By the way, during the demo, the dots didn't really operate well, and that was pretty embarrassing, but I actually really liked it.

45:56Speaker And the reason I liked it is because it shows they're actually showing the real system versus a demo product, uh, which is very easy to rig, right? You can very easily fake a demo of a specific thing. This was actually running live with real dots, and the fact that it failed just shows that they trust their product, and the fact that it failed means absolutely nothing because they're gonna fix these issues and they're gonna run okay.

46:14Speaker There's been a lot of feedback on dots all over the internet. Some people said this is incredible and that it is going to be the thing that they all want to use and work with and already gave a lot of examples on how this can be used. Some people are saying this is actually actually run it on a virtual machine in the cloud.

46:29Speaker So back then it was really complex, and I had to figure out how to set it up and how to install all the different things and how to give OpenClaw access to it, and so on. But back then, it made a lot more sense to me that I don't wanna run it on my own computer for multiple reasons. One is data access and security, and the other that it can run when my computer is shut down and it can keep on running.

46:46Speaker Well, that is now a standard across the industry. And like I said, if you're deciding to use GrokBot or if you're starting to use Instinct or if you now wanna use Dots or whatever Claude is gonna come up with that will do the same thing, this will be the way to use AI across the board with everything that we do.

47:03Speaker It will be always on sets of agents that are running in the background, and they whether my computer is on or off, regardless of where I am in the world, and we'll have multiple ways to communicate with them, uh, via voice, via text, via chat messages, or through our computer, because they're all gonna just run on their own machines.

47:18Speaker Going back to two interesting things about this topic. One is about Instinct that I mentioned and I told you previously. I really like Instinct. I use it a lot. I find it to be the most intuitive as far as it ability to understand the things that I'm trying to do and actually take actions in a very effective way without me knowing anything about what it's doing until it's done.

47:36Speaker Uh, it is very tempting. They just raised a billion dollars, and this just blows my mind. It blows my mind from several different reasons. Reason number one is I've never heard about them until about three weeks ago, and this is their Series C. So they've been around apparently for a pre-seed and a seed and a series A and series B.

47:52Speaker but what really blows me away with this billion-dollar raise is the fact that this is not going to be different than any other company out there. What Instinct could do right now, OpenAI will be able to do, Gemini will be able to do Claude will be able to do, Groq will be able to do, and Groq is actually already doing.

48:09Speaker And so do you think that a, another company can compete with the deep pockets, the talent, the compute, the reach, the access, the integrations that these four giants already have? I think not. So I don't understand why would anybody invest that kind of money in Instinct. I do understand that this is where everything is going.

48:29Speaker And like I said, this is gonna be the way we're going to do most of our AI work with always-on agents that will do stuff for us across multiple environments. So that's one aspect. The other aspect of it that I find really interesting is I have a feeling that in the future, and maybe not that far future, we won't need computers.

48:48Speaker And the reason I'm saying that is if most of the work is going to be done by agents, in my world, most of the work is already done by agents. Most of the stuff that I do is building and connecting more agents. I don't actually do the work. The agents do the work. If this work is going to be transferred to a virtual machine in the cloud, I don't need a computer.

49:07Speaker I need an effective way to communicate and monitor the work of my agents and make decisions. A computer is not that user interface. To me, what makes the most amount of sense is some kind of a wearable, either glasses that can show me displays or earbuds in which I can communicate or later on computer chips that can be stuck into people's heads.

49:26Speaker I don't know where this is going, but the interface of a computer will become redundant because we need the computer to actually do work and communicate. And if there's gonna be a better way to do that, then the actual computer itself becomes unnecessary. And I see more and more of that in my day-to-day work, even just with Claude and now ChatGPT, where the work that it's doing is actually not happening on my computer, it's happening in the cloud.

49:49Speaker I need the computer less and less. I open my phone and I tell it what I want that would happen and the things just happen. So food for thought, I don't know what that means for the computer industry, but in my eyes it doesn't look positive So again, not huge exciting news as far as new stuff, but exciting news as far as the fact now it's gonna be available through OpenAI.

50:07Speaker The thing that I think is the bigger announcement is ChatGPT Space. And Space is a shared workspace inside of the ChatGPT environment where teammates and dots/agents can all collaborate on different files, documents, and so on. And if you think about it, this is maybe the way everybody's going to be working in the future.

50:30Speaker I'm gonna read a quote to kind of explain how they explained it, and then I'm gonna tell you what I think it is and why I'm excited about this. So a new home for your team to collaborate with AI to get work done. Create dedicated spaces where teammates, ChatGPT, and your dot can build on shared knowledge.

50:47Speaker ChatGPT can keep your space organized based on instructions you provide, so you can quickly find what you need and pick up where you left off. Now, those of you who've been listening to this podcast for a while know that I have an environment like this that I rigged myself. It is built on a bunch of different components and skills in GitHub and things like this that have allowed me to work collaboratively with people from my team while being able to take off and continue in anything that I'm doing at any point without looking for previous conversations and so on.

51:17Speaker I've shared multiple Tuesday episodes where I explain exactly how I do this, so you can do this as well. Having this in a well-organized, well-structured solution is something I expected to be available a long time ago and was not available. That's why I built it the way I built it. But this is the future.

51:35Speaker The ability to come in and work collaboratively with agents and with other people in a unified environment is where this is all going. But the other thing that they announced, which is what they call Pages. Pages is like the unit inside Spaces. And again, I'm gonna read the quote: "Pages are a new type of document built for humans and agent collaboration.

51:53Speaker Anything you can do in ChatGPT, you can do on a page: write, research, generate charts, create images, or visualize information." And so here is something that I talked about several times on this podcast and a lot on the Friday AI Hangouts. By the way, those of you who don't know what the Friday AI Hangouts are because you're new to this podcast, it's a community.

52:11Speaker We get together every Friday at 1:00 p.m. Eastern, and we talk about AI, and people share what they built, and people ask questions and get answers, and we talk about what happened this week, and it's an incredible collaborative environment for people who care about AI and its progress, and it's free, and any one of you can join.

52:26Speaker And there's a link in the show notes if you want to join us. Like I said, every Friday, 1:00 p.m. Eastern Time. But back to the topic that I wanted to talk about is I think we're at an era where we have to separate the knowledge, the data from the form in which it appears. And if you think about it, we have been indoctrinated, mostly by Microsoft, but then through other platforms who came through the years that copied what Microsoft did, to think that the data and the form factor in which it appears are the same thing.

52:55Speaker It's either an Excel file or a Word document or a PowerPoint presentation, et cetera. There is very little freedom to express an idea, a concept, a project in all of these ways combined in a single space. It is doable with AI right now. My AI universe is built this way. There is data, and all the data is stored in either databases or MD files, that are backed up to GitHub.

53:19Speaker But the actual deliverables happen in real time as I need them because AI can do that. So when I need a mix of a table, a graph, and a video, I can generate a table, a graph from a video to convey the message, the deliverable, or whatever it is that I need to do. The same thing with code, web pages, et cetera.

53:35Speaker The output, the actual deliverable, and the data are two separate things. And what Pages, I think, enables to do is to do exactly that in a structured way. The AI, through its infrastructure, will connect to everything that you will give it access to and can generate the outputs you need as you and others need them for that particular step in the process that you're working on at that moment.

53:59Speaker In my eyes, this is an incredibly powerful thing. I see this again myself right now, and having a professional infrastructure that supports this is something that can potentially change the way we have worked so far. A, because it has agents built into it. B, because it enables collaboration with humans and agents.

54:17Speaker And I remember vividly when I started my MBA in 2007, when we started writing documents together on Google Docs, it felt like magic. We had several different people back then on Skype working together on a single document somewhere in the cloud, and I'm like, "Oh my God, this is the greatest thing I've ever seen."

54:35Speaker Now think on on steroids. We're not limited to just the form of a document, and you can have agents actually do a lot of the work for you while you are collaborating and brainstorming and thinking together with people and agents. So I think in my eyes, this is the biggest announcement by OpenAI in this event.

54:52Speaker And then the third thing that I wanted to talk about that I'm very excited about from this announcement is MCP events. So why am I excited about MCP events? Because it's something I talked about many times in the past. It is the biggest gap that existed, and again, I patched it in different ways in my entire AI universe.

55:12Speaker And the gap is you can use ChatGPT, and you can use Claude to initiate processes on third-party tools, whether through their API or through MCPs, you can have a process that Claude or ChatGPT will trigger based on something you tell it or based on a schedule that runs on scheduled tasks. However, the other way around did not work.

55:34Speaker You could not trigger a process inside of ChatGPT or Claude by something that happens on a third-party platform. Do this when an email comes in. Do this when somebody updates a status of something in the CRM. Do that when I have a issue in my code execution on a platform, et cetera, et cetera. That was not possible.

55:52Speaker Now, it was possible, and I shared some snippets about it in previous episodes that I did on Tuesdays, which comes down to either using a shared repository that is monitored to a scheduled tasks on a regular sequence once a day, once an hour, once, a, a minute, whatever frequency you want. That's not very efficient.

56:10Speaker And then the other option that actually did work well is routines inside of Claude. But again, setting up routines in a way that can be triggered through a third-party API is not something that is easy for any person to do if you're not technical and willing to go through the process of understanding how to put it together.

56:25Speaker So MCP Events enable to do exactly that. It enables the MCP protocol to report a trigger when something else happens on a third-party platform, and I can't wait to test that out and start playing with it. And again, I assume this will become a standard inside the MCP protocol, which means anywhere that you're using MCP, regardless of which platform you're using it on, will make this possible, which means you'll be able to build significantly more sophisticated things together with your native AI solution without integrating tools like N8N or Make.com or Zapier, et cetera.

56:59Speaker What is going to happen to these platforms? That's a whole other question that I don't have an answer for. But I have a feeling that if will work well, the need for platforms such as N8N, which right now runs a huge percentage of the processes in my different businesses, is going to be dramatically reduced.

57:16Speaker I don't know if it's gonna be eliminated completely, but a lot of the things I'm doing in N8N right now, I'm doing because I did not have a solution, like this. And now with this solution, I will need N8N a lot less Now, two other things that were very obvious in Dev Day. One is the focus on security.

57:33Speaker There were a lot of security-related new releases, as I mentioned before, which I think is the right move by OpenAI. But the other thing that was interesting is that there were very little discussion about the IPO. If anything, they basically said the IPO is not an urgency right now. We have bigger fish to fry.

57:50Speaker We are in critical times as far as developing our platform, and this is what we're focusing on right now, which basically means that if we initially thought that the IPO might be coming in Q4 this year, and then the rumors shifted to Q1 of next year, this may not even be the case anymore. Bottom line, huge announcements from OpenAI.

58:09Speaker From my perspective, they have taken the lead and the initiative away from Anthropic in the past few months. They are making bigger announcement, they're releasing bigger things, and maybe the biggest news out of all of this, and I'm gonna summarize it with that, is the release of the new, cheaper, very capable models that they've announced.

58:26Speaker And if you're like me and you're doing a lot of AI work or you're planning to do a lot of AI work, the one parameter that matters, literally the only one that matters, or not the only one, there's two others, but the two others are harder to measure. But the one that is easy to measure is the cost per task What does that mean?

58:42Speaker it means that the price per token or price per million tokens input or outputs is completely irrelevant because you don't know how many tokens the model is going to consume to complete your task. So you need to look at the task level and say how much completing this task is going to cost me. I don't care if it uses five tokens or five million tokens.

59:00Speaker I want the task completed. And in that race right now, OpenAI opened a huge lead over Anthropic. So if you go to Artificial Analysis, which is a website and a company that does analysis of models, they have a huge range of tests they run the AIs through, and they know what is the cost of running through all these tests, and then they average that to completing a specific task.

59:22Speaker So right now, the cheapest model to complete a task is actually Muse, The latest model from Meta with six cents for task completion. A very close second is GPT Luna, GPT-6 Luna on max effort, which is now how I write all of my code when I am coding. So for seven cents to task completion, you can do this with Luna.

59:46Speaker If you look at GPT 6.1 Sol, so the latest model they just released, it will complete tasks for seventy-two cents per task. If you go to the cheapest option from task completion by Anthropic, it is currently Claude Opus and not Sonnet 5.5. So Opus 5.5 on max at almost six dollars, five dollars and ninety-eight cents.

1:00:08Speaker And if you go to Sonnet 5.5, it is actually currently the most expensive model per task, despite the fact it is significant cheaper per token compared to Opus, which tells you again, looking at the cost per token may not be the right approach, and Sonnet will cost you seven dollars and sixty-seven cents for average task completion in the tests that are being performed by Artificial Analysis.

1:00:30Speaker So seven cents for Luna and seven dollars and sixty-seven cents on Sonnet 5.5. And in between them, you have GPT 6.1 Sol with seventy-two cents per task. Now, why isn't that the only parameter? What are the other parameters? One is quality. So the fact that the task is completed doesn't mean it's completed at the same quality.

1:00:54Speaker And it is very hard to measure, but I assume that using GPT 6.1 Sol will deliver better quality than GPT Luna. Otherwise, why do I need GPT 6.1 Sol? It is still 10 times cheaper than using the Claude top models right now to complete these tasks. The third parameter, by the way, is speed. How long will I have to wait to get the job completed?

1:01:18Speaker Uh, which is actually easier to measure. But I'm sharing this because this is very clearly the focus from OpenAI, and they've been very successful in delivering models that are capable in delivering solid, consistent results at a fraction of the cost of what Anthropic can do right now while adding all the other bells and whistles that we just talked about in their announcement.

1:01:38Speaker now I told you, don't say I didn't tell you this is going to be a long episode, but we're off to the rapid fire items. And as I mentioned here, we're going to talk about a little bit of security, from a different angle and then dive into lots and lots and lots of new model releases, uh, that I want you to know about.

1:01:52Speaker So something else that happened that is somewhat alarming or very alarming, depending on what you think about it, is there was a meeting of the UN aimed to establish the first international agreement to control lethal autonomous weapon systems, and the acronym is LAWS, L-A-W-S. And these talks basically collapsed in a collaboration between US and Russia to remove the critical safeguards that were supposed to be a part of that agreement.

1:02:19Speaker So this was a closed-door session that lasted fifteen hours, and the main issues that were removed by pressure of US and Russia were provisions that initially existed in the draft and the treaty, and this was language that required AI weapon systems to operate predictably and reliably, incorporate ethical standards, and most critically, mandate human review of military strike targets before an attack is carried out.

1:02:47Speaker So these were removed from the language of the treaty, which are the critical, most important aspects of the whole thing Now, despite the actions of US and Russia, 76 countries expressed support in moving forward in negotiations on legally binding rules for autonomous weapons. So this is not the final state.

1:03:07Speaker The delegations are actually scheduled to reconvene and have another meeting in Geneva in November to continue discussing, uh, whether to elevate this to formally mandates, uh, that will be forced on countries around the world. But as of right now, as I mentioned, US and Russia has moved critical aspects of this out of the discussions, and these critical aspects are basically telling you that's what they're doing right now.

1:03:31Speaker They have or they're in advanced development of autonomous weapons that will not take these things into consideration, will act on their own, may or may not have the right ethical guardrails, et cetera, which is not a good thing because that's a whole new level of arms race that we never had before that is very, very easy to lose control because of it, because you do not know what the outcomes are going to be because you're not making the decisions, but the AI is On the same topic, if you remember, we mentioned that US and China are in dialogue and President Trump is supposed to meet with Chinese leader Xi Jinping to discuss AI collaboration and safety and other aspects, and I really hope this

1:04:09Speaker will actually move forward in the right direction versus what we just heard from Russia and the United States when it comes to autonomous weapons. and now to all the releases that happened this week outside of OpenAI, the biggest news that could have been a deep dive item on its own is Gemini 4 has been released or kind of being released.

1:04:27Speaker So Google is finally back in the game with Gemini 4 Argon, which is a frontier model, and it's a real frontier model that they released on September 30th. It is, from a benchmark perspective, aligned with the top models from OpenAI and Anthropic right now It is currently leads 13 of the 19 benchmark that they decided to publish, comparing it to GPT-6 Astra and Claude Opus 5.5.

1:04:56Speaker So if you want a few results, 77.9% on DeepSuite 1.1, which is a software engineering thing. It is number one on autonomous bench. it is scoring 51.3 on the Zapier automation bench, which puts it at number one. It scores 91.7 on LV-bench, which is long video understanding. It is 19.6 on Harvey's legal agent benchmark, significantly ahead of any rival that is in single digits behind it.

1:05:24Speaker It is number one on several other indices, so a very capable model, at least on the index level. There's been a lot of scrutiny on that and a lot of negative feedback about this because people said that previously when Google released such statement about their lead on specific benchmarks, the models turned out to be not that great and actually not scoring that high on these benchmarks.

1:05:45Speaker But a lot of other people who have access to the model said it is actually very capable. Now, like all the advanced models that we have right now, Argon can autonomously find, validate, and patch critical vulnerabilities, from a cybersecurity perspective Now, when it comes to writing code, it is an extremely powerful model.

1:06:01Speaker Google said themselves that Argon is now powering internal workflows at scale, both coding and other aspects. Autonomous memory optimization freed over three hundred, terabytes of memory across data centers just by running on its own, uh, and a lot of other things that they're doing internally, that has been performing very well.

1:06:18Speaker There have been feedback that this may not be still as good in coding as Claude or Grok or ChatGPT, but at least they're getting very, very close, and they were very far behind in the past few months. From a pricing perspective, it is two dollars per million input tokens and ten dollars per million output tokens with a deep discount that currently exists.

1:06:38Speaker Uh, how long the discount will last, nobody knows. But right now it is very competitive in its pricing Going back to the completion of tasks that I mentioned earlier, they are, are in a good place between GPT-6.1 Sol at 72 cents to the Anthropic models, and they've completed the tasks at $2 per task.

1:06:57Speaker So not incredibly well, but not bad either Google has put a lot of focus on safeguards and making this model extremely safe. They're talking about the fact that it is almost impossible to use this model for cybersecurity, uh, but it's also very good at resisting prompt injection, and it has very strong misalignment monitoring via chain of thought analysis and better sandboxes that they've tested it with and so on and so forth.

1:07:20Speaker Again, a lot of focus on As you can imagine, the network on X and other sources had a lot of chatter around this release. people are saying, "Well, Google is back at the frontier." Some people are saying, "Well, let's see, first of all, if they're really at the frontier," because so far, uh, they haven't given it to anybody, which is the biggest point.

1:07:36Speaker This model is currently not available to the public. It is only available to a short selected list of companies, to release it more securely, similar to what we've seen initially with Mythos and Astra as they were released, to trusted companies for testing and patching and so on. And then it will slowly be released to the world.

1:07:53Speaker So there is no ways for us to really know how it operates and compares to the leading models. But I've said that multiple times before. you cannot dismiss Google as a leader in this space. They have more compute, deeper pockets, more data, their own hardware, huge amount of talent across their labs and so on.

1:08:09Speaker They will stay at the frontier, even if they have speed bumps along the way, which is great for all of us because it means that we're gonna have more options, which means it's gonna drive the performance up and the cost down, which means we will have better intelligence, hopefully safer intelligence, at a cheaper cost As I mentioned, uh, this week, we also got Claude Sonnet 5.5.

1:08:29Speaker So on September 28th, Anthropic released a new model. It is 30% faster and 30% cheaper per task, in theory, based on them, compared to Opus 5.5, but it is only on paper, as I mentioned, based on actual tests. So f- on a million tokens, it is significantly cheaper than Opus 5.5, but it is actually using a lot more tokens to complete the tasks, at least based on artificial analysis, so there's no real savings.

1:08:56Speaker That being said, it is a smaller model and cheaper model that is aligned with the performance of a previous model that was released just a couple of weeks ago. By the way, similar exact thing with OpenAI. When we think about Sol 6.1, it came out two weeks after GPT Sol 6, which tells you that pacing the frontier may not be exactly what's happening behind the scenes.

1:09:18Speaker So should you use Sonnet 5.5 or not? I'm using now Sonnet 5.5 and Opus 5.5 back and forth across multiple tasks trying to evaluate when I should use which one, and I must admit I don't have a clear decision yet. Uh, both of them are very capable and highly powerful models that are working very well for me across a very wide range of tasks Now, switching gears from OpenAI and Anthropic for a second, Meta made a very big announcement this week that they're taking Muse, that is a very capable model that we're not talking enough about.

1:09:48Speaker We told you that the consumer version of this is growing faster than any other, consumer app ever when it comes to the AI world. However, they just released Meta Enterprise platform, and the idea is to take what they've developed in Muse and, uh, build it in a way that enterprises can build businesses around it with AI capabilities and data security, and so on In addition to the announcement, they have a new chief that will run this.

1:10:12Speaker It is, Christian "CJ" Desai, the former CEO and president of MongoDB. He's gonna be the chief enterprise platform officer, and he's gonna report directly to Mark Zuckerberg. So again, as you can expect from everything we're seeing from OpenAI and Anthropic and their focus on enterprise, Meta is going in the same way Now, while Anthropic has been focusing on the AI space for over a year and a half now, and OpenAI for over eight months now, Meta is new to the space, and it will be very interesting to see how much they can actually participate in the game and take a slice of the pie.

1:10:46Speaker As I mentioned, their Muse model is actually scoring very high across the board in a very efficient way, and since efficiency is a big part of the game right now, they may have a way in. That being said, Meta has never been an enterprise kind of company, so it'll be very interesting to see how they manage to pull this off In parallel to this, Meta also announced Muse for small businesses.

1:11:06Speaker So again, they're trying to attack the enterprise as well as small businesses And the way they're trying to attack it is through existing integration. So, uh, there's a free version and also a freemium version that already integrates 20 leading business applications such as Shopify, Slack, QuickBooks, Asana, Figma, and others.

1:11:22Speaker So the idea here is taking a very efficient, cost-effective model and giving it to you already connected to the tools that you are using in order to allow you to use it as an alternative to the leading labs models. Again, will they be successful? How effective it's going to be? I'll keep on reporting as I learn from other people who are implementing it.

1:11:42Speaker Another company that make big announcement, uh, this week is Microsoft. Microsoft are going to sunset and eliminate basically their personal version of Copilot, so they're not going to compete, for people like you and me just using Copilot as an alternative to ChatGPT or Claude or Grok or Muse, et cetera.

1:11:56Speaker but they're focusing on a three-pillar platform approach. Uh, they've introduced Mi- Microsoft Copilot Home, which is a unified starting point that is combining chat, co-work, and office integration. Code, which is enabling anyone to build solution with GitHub Copilot, and Autopilot, which is a persistent, proactive agent that completes tasks in the cloud, just like we see from everybody else right now.

1:12:20Speaker So three very distinct kind of solutions, uh, one in focusing on the integration what used to be historical, Microsoft Office. Uh, the other one is code creation and enabling people to vibe code anything they want. And the third one is always running agents in the cloud, which again, is going to be the way we probably interact with everything going back to our, uh, segment earlier in this episode Another thing that Microsoft released this week is MAI Transcribe-2 Streaming, which is the number one position on artificial analysis for accuracy So these are a set of voice models that allows you to transcribe and/or stream, voice through, AI functionality, and they are, like I said, currently ranked

1:13:00Speaker number one on the analysis of their ability to do this, both in voice understanding as well as in voice generation. And they're saying that it's not just designed for speech, but for expressive speech, meaning it can capture nuances of real voices versus more AI robotic voices They also released a high-speed voice variant of this called Mai Voice 2.1 Flash that is supposed to provide significantly faster response with a 150-millisecond latency, which is really amazing, and it is not distinguishable from human regular speech.

1:13:36Speaker Another company that made a big release this week is Manus. So if you remember the last time we spoke about Manus is when the Chinese government reversed the purchase of it by Meta. So they have been quietly working on the next version, and now there's Manus 2.0 that introduces a new completely in-house agent harness, and it is producing very capable results when it comes to delivering across the board agentic capabilities like we knew from the first Manus, just significantly better.

1:14:06Speaker And as everyone else, it now has its own autonomous cloud computing that can run in the cloud without you having to have your computer running while being connected to your email, Slack, calendar events, Notion, et cetera. So again, this is now merging into the same form factor of an always-running agent with its own cloud computer that can connect to your data universe if you would allow it.

1:14:29Speaker Another interesting release this week comes from a company called Tavis that unveiled Griffin, which is what they call a HIM, a human interactive model, which is designed for real-time face-to-face AI conversation. So think about a Zoom call or a Teams call with an actual person on the other side that is not an actual person.

1:14:46Speaker It is an AI-generated model that doesn't just talk or chat with you You can actually see a face and have a full communication that can participate in video calls. In their live study, they claim that 48% of participants believed they were talking to a real person after a one-minute video call. This is a huge increase from a 2% of people that have thought this about previous models.

1:15:09Speaker I must admit, I went to their website, I played with their model. It is really good. It is very obvious AI, so I definitely fall into the 52% that know that this is AI, but maybe because I knew I'm going into an AI website, maybe if this was a, a blank test and this was a participant on a broader call, maybe I would not notice, but I seriously doubt it.

1:15:28Speaker That being said, there is a huge amount of people in the world today that do not know this is even possible, and then they don't even suspect, and then they may think this is just a weird person versus an AI-generated. But that being said, this is definitely the direction that we're going to go. So if we're talking about communication with agents in the cloud, right now they're just being used through text and voice.

1:15:47Speaker There is no reason why we wouldn't use them with video as well, Just think about going back to what we discussed, uh, about spaces in ChatGPT, a collaborative environment where people get in and have a conversation. There's no reason why that entity will not have a visual representation and can participate in a very natural way in a collaborative environment that includes video conversations of all the participants in the chat.

1:16:08Speaker Another interesting release to speak, going back to the point that Anthropic is now good friends with the government again. Anthropic Claude for government, it is currently deployed at state agencies with FedRAMP high authorization, that features the same capabilities as commercial customers while meeting strict compliance and requirements.

1:16:27Speaker This is straight from their announcement So we went from Anthropic being the first company with a secure environment deliverable to the government, to being removed from the government altogether because they're a risk to national security, to now they're deploying a unique specific variant of their platform across multiple government agencies.

1:16:45Speaker all of that in just a few months. Another interesting release this week is HeyGen launched a AI video model. Now, it's not really their video model. They're actually packaging a Minimax H3 model, but they've tailored it together with a lot the loved and effective HeyGen capabilities into a very cost-effective video generation tool that allows you to go beyond just their avatar capability.

1:17:07Speaker I've been a avid HeyGen user, so I'm very excited to see how this is going to be integrated and how it's going to work, and what kind of flexibility will it provide through the HeyGen product. Now switching from models to just interesting announcement that happened this week, Google DeepMind unveils SynthID Bio.

1:17:23Speaker So we talked about SynthID many times in the past. It's a technology that Google have been using for a while that is integrated into everything their AI generates that allows to know that something is AI generated, at least Google can know, right? So it's a something that manipulates the way the output of the AI works, and then there is a decrypter of that knows how to identify the specific pattern that it hides into the words and the letters and the images and the videos that it generates without us being able to detect it on our own.

1:17:53Speaker we all know that Google has been active with DeepMind in doing different biological experiments. they won the Nobel Prize for that. And so they're now doing the same thing. SynthID Bio embeds a signature directly into the biological code that it creates, so it is possible to know that a specific DNA sequence is AI-generated versus comes out of nature.

1:18:16Speaker This is, on one hand, crazy exciting, on the other hand, crazy scary because I'm thinking, "Why the hell do we need this?" And once you start thinking about this, the Terminator movie comes to mind and a lot of other similar movies, which I'm trying not to think about. But the fact that this is possible is mind-blowing to me.

1:18:33Speaker The fact that you can make minor small changes to a DNA that will not change the way the DNA actually works but is detectable by an AI system is absolutely mind-blowing. And like I said, this is a capability that Google announced this week now a few interesting physical AI announcements. Meta just launched, and that's a couple of weeks ago, I just didn't get a chance to talk about this yet, uh, released a new set of Ray-Ban glasses, and these are the first set of glasses that actually has a projection on their screen.

1:19:02Speaker These are extremely amazing in the things that they can do and how small the form factor is. They still look a little bulky, but they provide capabilities that nobody else has right now. The price tag is gonna be $799. It comes with this little bracelet that allows you to do hand gestures that actually impact the display and control what it is going to see.

1:19:23Speaker And going back to my comment from earlier of why do I need a computer? This is exactly the direction I think everything will go. So while this is bulky and the display is relatively low resolution, so it's 600 by 600 on each of the lenses with full color, this will keep on evolving, and we're gonna get better and better glasses that are gonna be less and less bulkier.

1:19:42Speaker That will allow us to communicate with always-on agents that run on their own cloud while being able to communicate with them through voice and sound and display and make decisions, and even small hand gestures to approve things without actually having to talk in environments when we cannot do that. So kudos for this achievement for Meta.

1:19:59Speaker They were at this for a very long time. And like I said, I think that in the not too far future, we'll see more and more of these, and basically everybody is going to wear them because that's gonna be the way we're gonna get work done. Another announcement that happened is Apple's HomePod which is a smart home hub That is supposed to be released in the next couple of weeks, so October 13, and it is supposed to have a six-inch square display and a distinctive design that is going back to a, the 2002 iMac G4, and it is supposed to integrate the new AI capabilities that are now built into the entire Apple ecosystem.

1:20:33Speaker So those of you who are in the Apple world and you want a new toy for your house, this is coming in the next couple of weeks and another very interesting announcement this week is that Kodiak AI targets the first driverless long-haul operation on Texas Interstate with IKEA partnership. So Kodiak plans to deploy fully autonomous trucks with no human in the cab on a 219-mile I-45 corridor between Houston and Dallas by the end of the year.

1:21:00Speaker Now, in their testing over the past four years, Kodiak has delivered more than 1,300 loads and logged more than 750,000 miles carrying IKEA goods autonomously with a safety truck doesn't do anything stupid and kills other people. and now they're going to do this without the safety driver, and they're going to launch it before the end of this year.

1:21:22Speaker This is obviously a very scary, and yet, again, if they've done this successfully for four years with so many miles without any major issues, I guess they're ready to do that. This shows you how far we're moving ahead with autonomous capabilities involved with having real humans around it operating in a normal human environment, and this is just a step that showing us how AI is going to be in physical AI embedded into everything we're going to do.

1:21:50Speaker I think the next few years we'll see more and more robots, more and more autonomous cars, and so on in more and more aspects of our lives. There's a lot of other news to report this week, but I think we went long enough today, so I'm gonna end it here. there's more stuff for you to read in our newsletter.

1:22:05Speaker So if you're looking for additional news, you can find them over there. At least skim through them and see what you find interesting, uh, and what you wanna read more about. We will be back on Tuesday with another deep dive into how to perform a specific task with AI. Don't forget to share this podcast with other people who can benefit from it if you find it valuable, and don't forget to check out the next cohort of the multi-agent orchestration course.

1:22:28Speaker It starts in less than a month, and it is probably going to be the last one this year. So if you want to know how to build powerful autonomous solutions for your business with AI before 2027, and you should, come and sign up for the cohort. The links for that are in the show notes. That's it for today.

1:22:43Speaker Have an amazing rest of your weekend.