180 | Will AI get out of control by mid 2027? Super-Agent is showing off amazing results, Google AI is taking center stage, and many more important AI news for the week ending on April 12, 2025

0:00Hello, and welcome to a weekend news edition of the Leveraging AI podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, grow your business, and advance your career. This is Isar Maitis, your host. And every week, I'm hoping there's going to be some slowdown with a week that we won't have a lot to report, but this is not going to be this week. There is an explosion of stuff for me to share with you. Our main topics are going to be everything that Google shared in Cloud Next 2025, the big event they just held this week. We're going to talk about a doom and gloom projection to where I is going by 2027.

0:35We're going to talk about major progress with AI agents. And then we're going to dive into multiple rapid fire pieces of news, including a lot of new and interesting releases from the biggest players and small ones as well. So let's get started.

0:53Google held its Cloud Next event this past week, and as expected, it was huge and packed with AI related announcement. If you go back a year and a half ago and a year ago on this podcast, when Google's respond to the chat GPT moment was, I'll be gentle, embarrassing, I said that it's not a good idea to bet against Google in the AI space because they have everything they need in order to win this race. They have the capital, they have the data, they have the human capacity, they have and can attract the best minds in the planet. They have the compute capabilities, the distribution, the tools, everything you need in order to be successful in the space.

1:32And they're the ones that invented the transformer that actually led this entire revolution. So they have years and years of research before even the Chachipiti moment. So it was very obvious to me back then that they will take potentially the leading position. and it seems that this is the direction that it's going. If you've been watching the Chatbot Arena chart in the past few months, you see that Google is on top at number one and sometimes number one and number two most of the time. You know that they now have the best video generation tool with Veo and so on and so forth. And I think slowly they will take over more and more components and more and more capabilities in which they are going to be the leading force in the world and they will have the ability to combine them all together into unified environments.

2:16And this is more or less what we've seen at Cloud Next. It was a spectacular presentation of their ability to combine multiple capabilities together, all under the Gemini umbrella. And we're going to go very quickly on some of the things that they've shared during the conference. First of all, they introduced Ironwood, which is a seventh generation TPU, which is their AI chip that is now built specifically for inference, which will allow the test time computes, these reasoning models to run more effectively. and it is going to be available on Google Cloud later this year. Compared to prior generation, it offers five times more peak compute capacity and six times the high bandwidth memory capacity.

2:54And overall, they're saying it's producing 10x faster results of every process that it's doing, which is obviously very impressive and potentially will present an alternative to NVIDIA GPUs. But they've also introduced the whole creative side of this, which was really interesting and impressive. So they introduced Lyria, which is a new addition to Vertex AI. That's their new voice model. So now in combination with Imagine and VO2, which by the way is getting a new VO2 and the ability to generate music, now allows creators in Vertex to create anything they want. Voice, images, videos, speech, music, basically any creative thing you want to produce, you can now produce in one environment within Vertex.

3:36This is nothing like anybody else has. And that is obviously very powerful. they're claiming that some of the introduction videos that they had in the transitions to each of the speakers were done with the new version of VO2. And it's absolutely mind-blowing. It's worth for you to watch the first few seconds of every presentation that was done there, just to see what was done. I'm sure that required a lot of editing and post-production and stuff like that. But even if just components of this were produced with VO2, it's incredible. They're introducing a lot of new tools to Gemini in workspace, bringing more and more capabilities into Docs and Sheets and Meet and Chat and all the stuff that is available in Workspace.

4:11I must admit, I'm using more and more of Gemini within the different tools. I'm just praying for the day that it's all going to come together and it's not going to be Gemini for Docs and Gemini for Sheets and Gemini for Drive, but actually just Gemini that will know everything that I'm doing, including my emails and everything. I guess we'll have to wait a little longer for that. I really thought we're going to be there by now, but I guess that's more complex than I thought. They also introduced some additions and improvements to AgentSpace, which is their platform that allows companies and enterprises to develop agents.

4:39And they announced that they have a growing AI agent marketplace where you can go in and see what partners of Google have developed and is making available. And you can buy agents from different people. This is a new vector of the economy, if you want, of companies that will develop these agents. And you'll be able to go and buy them as an enterprise and use them within the Google ecosystem. They also unveiled what they call Agent Development Kit, an ADK versus an SDK from traditional software, which is an open source framework for building agents while maintaining control over their behavior.

5:11And they also introduced A2A or Agent2Agent, which is a new open source protocol that gives agents the ability to talk and collaborate with other agents regardless of the vendor. So if you think about the huge success of MCP in the past few months, and we talked about this multiple times, it became a huge success because it is open source and it allows agents to connect to multiple data sources. Well, now this new framework from Google this time, MCP is from Anthropic. So this new framework is targeting the ability to agents to talk to other agents. With the amount of hunger there is for this kind of capabilities right now, I assume this will catch in the same speed as MCP did for connecting to data and tools.

5:52They also shared a new unified security solution called Google Unified Security, which brings together all their security capabilities for the products, threat intelligence, security operations, cloud security, and secure enterprise browsing. So they're building a whole suite of security tools around their AI infrastructure, which makes perfect sense if they want to see this successfully deployed across multiple enterprises. And they're expanding the access to their WAN solution, their wide area network, from being just for internal use across Google facilities to be available to enterprises for significantly faster capabilities, services, and search across everything Google.

6:29So bottom line, Google is showcasing why they are the leading developer of AI solutions in the world, how their AI cloud solutions are going to be in the forefront of AI capabilities. And that's obviously in a fierce competition with AWS and Azure. So it'll be interesting to follow. But definitely, this was a very clear statement by Google on how they see themselves in this race. And from all of these exciting news that every time I see one of these events, I always think of Cybernet from the Terminator movies. There seems to be people who agree with me and actually know a lot more than me and have taken the time to actually put this down on how they see the trajectory of AI moving forward and what its impact might be in the world.

7:11So two AI experts, Scott Alexander, who's the author of Astral Codex 10, and Daniel Cocotaglio, who is a former AI researcher, has created a website called AI27, which shows a timeline from now till 2027 and what they think might happen during that timeframe. Now, I found it through the Dwarkesh podcast. So those of you who don't know, Dwarkesh is a brilliant interviewer and he has a podcast where he has access to basically anyone in the world. And he does a lot of in-depth technical interviews with a lot of interesting people. So he brought these two people to an interview at his podcast. I'm warning you on two things.

7:49One, it's a three hour interview. So even at one and a half speed, you need quite a long drive. but also that it's very scary and uneasy to listen to, and in some cases a little too technical. But I will summarize most of it for you. So they're trying to portray how super intelligence will develop and what might be its impact on the world. The key concept that they're talking about is that once we hit super intelligence and AI agents can build AI agents or AI systems can build new AI systems, we're getting to intelligence explosion, basically a self-sustained feedback loop that keeps generating better and better AI completely out of our control.

8:30And they're claiming we're going to get to that point around mid-2027. Now, the way they've done this is they have identified several milestones that are required in order to get to that point, and they try to estimate the time it will take to each and every one of these milestones. Hence the website, which is called AI 2027, and you can go and look at, is actually really, really cool, extremely well done, beautiful graphics, charts on the side, but it has a timeline on the left. And as you're scrolling, you can kind of see where you are in the timeline and how we might hit these milestones. In addition to the agentic world, they're also claiming that about a year after hitting superintelligence, the superintelligence itself will allow us to build production capacity to generate a million robots per month.

9:15That sounds completely sci-fi to me, but a lot of things in AI sound completely sci-fi to me, which doesn't necessarily make them wrong. A million robots per month, meaning that very, very quickly, you can take over more or less any task and any job in the world because you're generating 12 million robots every single year that can take, each and every one of these robots can do the job of at least one person. that shows you how quickly it's going to change the way we know the world. Now, they're talking about that once we hit the point that these AI tools can run 50 times faster than humans in doing cognitive tasks, what it basically means is that they can do in one week the work a human does in an entire year.

9:59That's basically what 50 times faster means. And you can spin up as many of these as you want as long as you have the compute power. So the combination of these two things shows you how quickly things can get out of hand. This basically means that even if it does not accelerate beyond that point, it just gets to the 50x speed, it means that in one year, it can generate the innovation humans would generate in 50 years. That is obviously something we cannot even grasp. And they're saying, we're going to get to that point in 2027, two years from today. Now, one of the main concerns that they're raising is what they're calling an alignment crisis.

10:39Basically, that our ability to align, or if you want, control and understand the super intelligence is not going to work. There's actually a really fascinating conversation about this by Yuval Noah Harari, who is definitely a doomer when it comes to AI. And he's saying that the relationship between us and AI right now is we're an adult and the AI is a three-year-old and we can control it and we can try to understand what it's doing and we can easily identify when it's trying to fool us. But he's saying in the very near future, it will be the other way around, that the AI will be the adult and we will be a three-year-old with our ability to think and understand.

11:12And he will be able to manipulate us in multiple ways that we can't even understand, just like an adult can easily manipulate a three-year-old. And so this is where we're going. And this is why I think that the research that we talked about several times on this podcast, Bionthropic, to try to understand the inner doing of the AI quote-unquote black box is so critical for us to be safe from AI in the future. The question is that enough, and the question is it moving ahead fast enough compared to the speed that AI is moving. Now, they're talking a lot in this timeline about the AI arms race between US and other nations, mostly China, and they're saying that the incentive to get there first is incentivizing cutting corners and generating significant safety risks, especially in alignment research.

11:58Now, Their model also talks about China stealing some of US secrets to accelerate their side of the process, what they're calling Agent 2. So there's these different level of agents that you can get to. And by stealing Agent 2, it allows them to accelerate their development. I must admit that from everything that I'm seeing from China recently, I don't necessarily think they need to steal anything from the US. They're very, very close behind, if even behind, the US development of AI capabilities. But either way, this is not painting a positive future of where the AI race is going. That being said, they themselves admit that they're not sure that this is the scenario that's going to happen.

12:31Even Scott Alexander himself is saying, and I'm quoting, I'm not completely convinced that we don't get something like alignment by default, meaning that the models won't behave just because that's the way they will behave. But the thing is, we don't know. And the truth is, nobody knows. And that's the really scary thing about this whole thing. They're also discussing the potential nationalization of AI development, but that there's a very serious problem in the decision whether it should be nationalized or run by private company. And now I'm quoting Daniel Cocotaglio. And he said, I would summarize the situation as governments lack the expertise and the companies lack the right incentives.

13:07And so it is a terrible situation. And I tend to agree. So what is the bottom line? The bottom line is this is one possible scenario, right? I don't know. And nobody knows if that's the scenario that's going to happen. We might have full control. Everything might move a little slower. There might be hurdles that will slow it down, whether from government or from people or even from specific industry. But the reality is it's a possible scenario. And that possible scenario is coming in two years, meaning they're claiming that within two years from today, mid-2027, AI will be able to spin up new AI that is so powerful that we have no ability to control it.

13:45And he will also be able to generate physical AI robots at a speed that nobody thought will be able to generate at that point in time. And my goal here is obviously not to scare you, but just to paint potential versions of this future. There's obviously the flip side of that. If you listen to Dario Amadei, he paints this future of abundance and everything is good. And we solve world hunger and no diseases and so on and so forth. Where is the future? I'm not sure. I must admit, I am more probably on the doom side with everything that I'm seeing and the speed that I'm seeing things are moving and the lack of control.

14:18Do I think it's as doom and gloom as AI27 presents? I don't think so, but I really do think that we all need to work together to educate each other and to find ways to control this before it's too late. We talk a lot in this podcast about the importance of AI training. multiple research from leading companies has shown that this is the number one factor in success of AI deployment in businesses, large and small. I'm excited to announce that we just opened the registration for our spring cohort of the AI Business Transformation course. I've been teaching this course for two years, starting in April of 2023, and hundreds or maybe thousands of business leaders has went through the course.

14:58We had people in the recent cohort that ended in February from India, the Emirates, several different countries in Europe, South Africa, many places in the US, Canada, and even Hawaii. So regardless of where you are in the world, this could be a great opportunity for you. In previous courses, we had people as far as Australia and New Zealand. So weird hours of the day, but still getting a lot of value from this course. The course is four sessions of two hours each spread over four weeks on Mondays, noon, Easter time, starting on May 12th. if you are looking for ways to accelerate your personal knowledge and career or to change the trajectory of your team or your entire business, this is the right course for you.

15:39It is really a game changer. And within four weeks and only eight hours with some homework, you will dramatically change your understanding of how to use AI in a business. We give multiple hands-on examples and use cases and teach you the tools and the processes on how to use them. And we end up with a detailed blueprint on how to actually implement AI successfully business-wide. So if this is interesting to you, go and check the link in the show notes. You can open your phone and click on it right now and go and check all the information about the course. And because you are a listener of this podcast, you can get a hundred hours off of the price of the course with promo code leveraging AI100.

16:14I would love to see you join our course in May. And now back to the episode. Now, within those lines of doing it in a safer way, DeepMind released a 140-page paper that is discussing safety development of AGI. They're predicting that AGI could arrive by 2030 and that it could support drug discovery, personalized learning, democratize innovation, enabling small organizations to tackle big challenges, address multiple big problems in the world. Those of you who have been following Demis Asabis, who is the CEO and the founder of DeepMind, you know that's his goal in life, is to solve really big problems for the human race and the world.

16:53And then putting up this paper that is called An Approach to Technical AGI Safety and Security, the goal of it is very, very clear, is telling the world this is where we're going, but this is how it needs to be done right in order to reduce the risks. And the risks that they're highlighting is misuse. So basically people using AI to do bad things in the world. Misalignment, basically what we just talked about when the AI decides to do its thing and goes rogue versus listens to the humans and what we want it to do. And then accidents, basically creating situations where AI generates negative outputs by mistake.

17:24And they're pushing for more robust training on the safety side, monitoring in a different level than we have right now. And human in the loop checks to prevent harm in every step along the way. They're also calling for strong regulation because, and I'm quoting Shane Legg, who's the chief AGI scientist, he's saying this will be a very powerful technology and it can and should be regulated. And they're pushing and urging for global cooperation on that aspect. The problem with this is that they're claiming that we have time until 2030 and you have other experts who are claiming this may be coming this year or next year, or as we heard, superintelligence by 2027.

18:00So the timeline that they're predicting might not be aggressive enough to stop whatever needs to be stopped, if it needs to be stopped, in time. To make things even more interesting or complicated or scary, depending on your point of view, a new MIT study that was published on April 9th found that AI models lack coherent value system. Now, that's good and bad because on one hand, they lack value system. That sounds really scary. But on the other hand, the research found that they lack human-like priorities like self-preservation, which is something that was actually assumed that they do have. The biggest thing that they found, and they researched multiple models from Meta, Google, Mistral, OpenAI, Anthropic, and they were testing them across multiple views to see are they fully aligned or not compared to human values and other guidelines.

18:47And what they found, and I'm quoting Stefan Kasper, who is the MIT doctorate student, who is one of the co-authors, he said, for me, my biggest takeaway is to now have an understanding of models as not really being systems that have some sort of stable, coherent set of beliefs and preferences. In other words, despite all the training and the alignment and everything that we're trying to do, they do not have the same core values that drive humans to make decisions and take action. And that is a very scary thing, especially with the level of acceleration that's happening right now. Now, let's combine this with everything that happens in the agent world.

19:24And there are a lot of agent news. And we talked about this many times that 2025 is the year of AI agents. We talked last week about Emergence AI and their new announcement that their system now allows AI agents to build AI agents on the fly. So basically, you give it a task, it understands what the task needs to be, and then it creates agents, sub-agents that will do the sub-tasks that needs to be done in order to complete the main task that you gave the agent. And their CEO, Satish Pillai, has said the following. This is a step towards AI that evolves itself. This is literally what they're saying that they're doing.

19:55They're not trying to hide it from their perspective. That is the goal. Combine that with multiple new AI agents running in browsers and controlling them, and we'll give several different examples in the show in this episode. The first one comes from Opera, the browser company. They just introduced Opera Browser Operator. It's another AI browser controller, but very different from OpenAI Operator and the Cloud solution. This is built into the actual browser, meaning OpenAI and Cloud are actually taking screenshots every second and trying to analyze a screenshot to see what it's doing in order to decide what to do next and where to move the mouse.

20:28This is built into the actual browser itself, which allows it to be significantly faster and more accurate with everything that it's. Now, there's goods and bads in the fact that it's not image-based, but it's actually looking at the browser. As I mentioned, it will allow it to run faster, significantly less lag. It also allows to reduce vulnerabilities because the browser actually understands what's actually happening. But it still opens the door for complete misuse of your browser to do, well, anything the agent decides to do. Staying on the topic of agents and their potential impact, a new interesting startup called Artisan is developing an AI sales agent.

21:02That's what they're developing. They just raised a 25 million Series A, but the big deal that made them really known is that their ads are saying literally, stop hiring humans. That's what the ad says. That's obviously very controversial, but the goal and the direction is very, very clear. They want to replace human salespeople. They're claiming that their flagship product agent, Ava, now only hallucinates in one email out of 10 ,000 emails that it answers. That's significantly less than humans, right? So we're saying, oh, it still hallucinates. It's one in every 10 ,000. Well, people make mistakes when they write emails as well.

21:37And I'm sure they make more mistakes than one in every 10 ,000. And they're now coming up with new AI agents. One's called Aaron for Inbound Messages, and the other is called ARIA for Meeting Management. and the three are obviously going to work together in tandem to do well everything salespeople need to do. Let's keep going. A new release was made by Palo Alto GenSpark, which just launched what they call SuperAgent, which is an AI system that autonomously handles tasks like anything in the browser. They are basically doing what Manus from China is doing and they were even able to score a higher score on the Gaia benchmark, which tests these autonomous agents when Manus scored 86 % and they scored 87.8%.

22:16For those of you who don't know what Manus is, we talked about this, but these are fully open, do everything agents that are really general in their purpose. It's not geared towards being a sales agent or something else. And they can do everything from browser rate to write reports to creating web pages to writing software and combining all of it in a multi-step process. I've seen two mind-blowing demos this week of things that GenSpark can do. And it literally blew my mind in one of them. it created a website that allows to enter specific information and get outputs for specific vendors of a specific company.

22:52And this tool doesn't work in seconds like it used to from GPTs. It actually takes hours to work sometimes, but it can spin up these amazing solutions. The other demo was requesting to create an ebook that can be used as a lead magnet for a specific target audience. And it went and it created this amazing ebook. It's beautiful. It's designed well. It has graphs and charts and the right color scheme. And it's very visually appealing. It looks completely professionally produced, which means it looks like the data is real. I don't think anybody will test it, which is part of the problem because these tools are going to generate these really high-end products.

23:26I really doubt anybody will go back and test whether the data behind them is actually accurate because what the tool did in 55 minutes may take you two days to test and you just won't do it. And you will just believe the information that is in there. And that is another very scary aspect of this whole thing. But this is where we are going with agents. This super agent connects to nine different large language models and 80 plus tools that it can use to do the things that it is doing. And to tell you how far behind the government is connecting the dots to what was said by the authors of AI27, in the ASU GSV summit this week, which is potentially the US biggest education summit in the world, the US Secretary of Education repeatedly called artificial intelligence A1 instead of AI, meaning she doesn't even consistently call the term correctly when she's speaking about this.

24:16Now, she might have been excited and maybe making mistakes, but the fact that she made that mistake multiple times, I've used the combination of the two letters AI probably thousands of times in the past two years, and I did not say A1 even once. So what does that tell us or hints to? It tells us that the people who are driving some of the biggest and most important decisions of our generation, including how AI will be integrated into the education system, which is a huge opportunity and a huge risk, do not understand basic concepts of what AI is and what its implications might be while industry is running at full steam ahead and keep on accelerating every single day.