172 | AI agents are taking over, Google is on fire with AI releases, using AI to talk to animals, and many more important AI news for the week ending on March 14, 2025

0:00Hello 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 Maitis, your host. And we're going to start the show today by talking about new technologies that are coming up and emerging and how they're going to impact the world and specifically the workforce. And then there's a very long list of rapid fire items with a big focus on Google and a lot of stuff that they've been releasing recently and potentially our ability to talk to animals.

0:23So let's get started.

0:29Last week, we mentioned a new Chinese agentic model called Manus that is taking the world by a storm. And that storm has grown dramatically in this past week. So let's dive a little deeper into this topic. A Chinese company called the Butterfly Effect has released a new agentic environment called Manus that allows you to take action within a browser based on a agentic approach to tasks. It can perform multiple tasks at the same time. Some users report 50 plus tasks all running simultaneously. And it does things like social media analysis, stock trading, travel planning, online shopping, basically all the stuff that we can do on a browser, it does on its own.

1:09Now, in addition to the fact that it can do all these tasks, it's actually scoring very high across multiple benchmarks. So Manus claims superior accuracy over DeepSeek and even over OpenAI in a benchmark called Gaia, which is spelled G-A-I-A, which is a benchmark for testing agents, which is very different than the large language model benchmarks, where it's a non-benchmark here, it actually evaluates it on real-world tasks, and it actually scores higher than more or less every other tool out there, including the top chat GPT models. Now, while it doesn't have the human-like, more conversational voice of chat GPT, it definitely outperforms when it's getting the job done and performing tasks.

1:53It is currently only available by invitation only, and it has been rumored that invitations to get access to this model has been selling in China for numbers between$100 to$1 ,500, $1 ,800 per invitation. I am sure that the company did not expect this kind of demand, and they are struggling, and they mentioned that themselves, that they're going to try to deal with that and ramp up their demand. But while it was very obvious even in 2024 that 2025 is going to be the year of agents, We haven't seen anything in the immediate user market. So enterprise-level agents have been deployed. There are multiple big tools for large corporations to develop and deploy and manage agents, but that was not available to the wild public or to small businesses.

2:34And Manus is the first shot, if you want, fired in the open-to-the-public agentic universe. And I am 100 % certain that this will not be the last, and it will probably accelerate the process dramatically. So how does it work? Well, in the back end, it is using Cloud 3.5 Sonnet, and they're also testing Cloud 3.7 combined with Quen from Alibaba to actually do the thinking process. And they're trained their own capabilities on top of that, and the agents to actually take actions on the different browsers. I don't have access to it yet, but I've seen multiple demos of people doing some pretty remarkable stuff.

3:08It's still not fully functional. There's still some sharp edges, and it doesn't always get the job done, and it does some weird things every now and then, but the direction is very, very clear. And as I mentioned about mid-2024, the year of 2025 is going to be the year of agents. And again, now that Madness has done this, I expect the big players from the US and around the world to follow suit and provide us similar capabilities. So if you think about the fact that we got browser access from Claude mid-year last year, and we got ChatGPT operator at the end of 2024, this was just the beginning.

3:39And I think tools like Madness are going to be everywhere and available to all of us across multiple aspects. As I mentioned in previous episodes, the two biggest things that these things have to solve in order to really be used as a widespread business tool versus a cool gimmick that you can use to do research on the next sneakers you want to buy is security and control, meaning having the ability to define very clear guardrails on what these tools can and cannot do in order to protect your computer, protect your data, protect your company servers, as well as stop them from doing stuff that you didn't plan to in environments they do have access to.

4:13So let's say you allow it to shop online, you want it to shop for sneakers, and now it finds a really good Mercedes to go with your shoes, it may go and buy it because it does its own thing. What prevents it from doing this is its own logic, but I think there needs to be another layer that protects beyond just the thought process of the agent itself that will define which websites it can visit and cannot visit, what things it can and cannot do based on personal preferences or company guidelines. And once these things will be in place, I think these tools will see significantly wider spread and usage because they are incredibly powerful.

4:46Connecting this to many conversations we had in the past few months about the impact on the workforce, this can do the work of multiple employees with one agent right now for free, shortly probably for a small amount of money. and it can actually follow a process. And the next step would be how to train these models, which will probably gonna be as simple as just showing them what they need to do once. And then they can do it again and again and again at a much higher speed, at a much higher accuracy, never sleeping, never taking vacation and so on. And this is not perfect yet, but it's getting very, very close to the point we'll do that, which will have an even bigger impact on the workforce than large language models and generative AI is right now.

5:24Now, to show you how big the impact of AI is on the workforce, there's another very interesting headline from Y Combinator this past week. So Y Combinator, probably the most known accelerator in the world, has reported that 25 % of their startups in their winter 2025 cohort have their code bases generated 95 % by AI. So think about what I said, a quarter of the company in the most advanced accelerator in the world are basically creating their products, their software products, by AI almost completely. That was shared by Jared Friedman, who's a managing partner over there, so it's from reliable sources.

6:02Now, Friedman also mentioned that every one of those people is highly technical, completely capable of building their own products from scratch, and yet they are letting AI write 95 % of the code. He's saying a year ago, these founders would have coded the whole thing from scratch themselves, and now they're using AI to slash development time, letting startup sprints significantly faster than ever before. Now, I mentioned to you in the past two weeks, the concept of vibe coding that is gaining more and more attention, especially around Silicon Valley and the concept behind vibe coding is that you just use your natural language and let the machine write the code while you're monitoring what it's doing and just working along the machine, letting it actually write the code.

6:42Now, the reality, this is an amazing tool that is currently helping developers develop faster systems, but it's also giving access to people like me who cannot write code to write smaller applications to do different things that were not possible before. And this means that you can do the same thing too. If you are a coder, I assume you're already using these kind of tools, but if you're not, then you can start trying them. And I'm actually planning on recording a specific Tuesday episode on the various Vibe coding tools that allows you to develop applications and different capabilities from scratch on your own without knowing anything about writing code.

7:13So that will probably come in the next few weeks. And additional support to this current situation, Dario Amadei, the CEO of Anthropic, in an interview this week, was claiming that 90 % of code will probably be written by AI within three to six months, and nearly 100 % within 12 months from today. He shared that during a console on foreign relationship event, and we'll share a few other things that he said on that event that are really interesting. But this is a very bold and aggressive timeline, right? So we had AI coding tools for a while now, whether it's GitHub Copilot or Anthropic Claude's Code, which they just released that we shared last week, or Cursor and many others that are gaining a lot of momentum.

7:52But going to 90 % of code within the next six months and 100 % by the end of the year is very aggressive. Even if he's wrong, the direction is very, very clear. More and more code is generated by AI versus traditional coders. Now, Dario says that programmers are not going to vanish. They'll just steer the ship. They will basically be a team lead running a lot of AIs writing code for them that they'll be able to monitor and guide and reshape in the way they need in order to develop their products faster. Instagram's co-founder, Mike Krieger, who is now adanthropic, agrees, and he is predicting that engineers will soon shift from writing code to reviewing AI-generated code.

8:30And a lot of other people, such as AWS CEO Matt Garman, are stating the same thing. And VDS CEO Jason Hong also shared similar sentiments. He's claiming that coding might be already dead in the water with the rapid development of these tools. And he's basically suggesting to younger professionals or people going into universities right now to go and study biology, education, manufacturing, or farming as plausible and more secure future career options compared to writing code. Again, very extreme predictions, but coming from some of the most influential and relevant people in the tech industry.

9:05Staying on the same topic of pushing the limits of what AI can do in the organization and also talking about Anthropic. Anthropic just released an upgraded console that allows to share prompt libraries and other AI developed tools across the enterprise, which basically lets people from different departments, such as marketing, customer service, and the tech team to share resources and prompts between the organization. And I'm quoting from their spokesperson, we built our shareable prompt to help our customers and developers work together effectively on prompt development. What we learned from talking to customers was that prompt creation rarely happens in isolation.

9:42It's a team effort involving developers, subject matter experts, product managers, and QA folks all trying to get the best results, which is true, right? If you think about it, if you look at a job, it's very rarely done by one person. Different people has different aspects of the job. And if you can develop prompts or one big prompt in tandem with all these people and all their inputs in mind, you can develop better, faster, more consistent solutions that will work for everyone. And now it's going to be built into the actual systems and processes themselves. Staying on the topic of changing the world and how it works and the whole impact of agents on our future workforce and technology in general, on March 11, OpenAI launched a new API called Responses API, which is a new tool set that allows to blend the simplicity and basic concepts of chat to advanced agentic capabilities and the Assistance API.

10:31So the Responses API, which is this new tool, combines three previous capabilities. One is Web Search, which allows you to get real-time answers with citations, File Search, which allows you to dig into companies' documents faster and better, and Computer Use, which allows you to actually use the mouse and keyboard to automate tasks on the computer. Going back to the Manus moment, while Manus has developed this as a end-user tool on the OpenAI universe, this immediate release is a tool for developers because it runs on the API side, but it allows to do some remarkable things and develop capabilities that were not previously possible for most companies.

11:07Also in this new tool mentioned that the new web search capabilities is now showing 90 % accuracy versus the 63 % accuracy of the previous tool, which tells you it's still not bulletproof to hallucinations, but it's a very big step in the right direction. And I'm sure this vector will keep on moving to give us answers that are more and more reliable over time. They also released a agent SDK that upgrades the Swarm framework that they released previously, and it's letting developers orchestrate multi-agent workflows with ease. So configure the agents, smart tasks handoffs between agents, and safety guardrails, as I just mentioned previously, which are a necessary component in order to actually deploy this safely in an organization.

11:48Now, while all these things are moving very, very fast, they're not completely reliable yet. As an example, the computer using agent model, while it's really impressive, scores just 38.1 % on complex tasks on the OS World benchmark. So that means it doesn't get the job done on complex tasks two-thirds of the time. This means that if you're planning to use these agents in the immediate future on simple tasks, you're probably safe or safer. And more complex tasks, you're not there yet. And this means that you need human supervision to monitor what these agents are doing. In my software company, Databreeze, that develops AI automation solutions for companies, this is a key component, which is allowing users to verify the automations that the agents are performing, and this way to reduce the errors that the system may generate.

12:34Over time, we're seeing significant decline in these errors as the system learns different use cases, but it's something that you need to be aware of, just like you need to be aware of hallucinations with generative AI. These tools will not get it 100 % of the time correct. The flip side of that, humans don't get the job done 100 % of the time correct. So all those tools need to do is get to human level errors and then go beyond that in the near future. And there's zero doubt in my mind that that's where we're going. And now to a slightly contradicting point of view on the direction that AI is taking right now, Hugging Face co-founder and chief scientist officer Thomas Wolfe is arguing that today's AI system are morphing into, and I'm quoting, yes men on servers rather than revolutionary thinkers that are needed for breakthroughs.

13:16He is specifically challenging Dario Amadei, the CEO of Anthropic, in his vision of a compressed 21st century where AI drives decades of progress into years or months. And he's saying that it's very unlikely with the current tech. He's claiming that scaling today's AI models will not create Einsteins in data centers. And he's saying that the main mistakes people usually make is thinking that Newton or Einstein were just scaled up good students. And he's saying that that's not the case. geniuses comes from questioning and not by conforming with the known facts. So he's saying that today's systems are built to excel in answering known questions versus questioning the status quo and trying new things.

14:00And hence he's saying these systems will not be able to have like the Copernicus moment who found that the earth is orbiting the sun and not the other way around by questioning the norm and the consensus of his time. So it will be very interesting to see. There are really smart people on both sides of that argument, but this is definitely an opinion that you need to know that is out there right now by people at the forefront of the AI revolution. So a quick summary of what we talked about so far. Agents are coming. They're coming fast. They're coming to the enterprise world. They're coming to the backend API universe, and they're coming to the front end for small businesses and individuals to use.

14:34There are already examples and capabilities and tools out there to develop and deploy these agents. and it's a matter of time, months, before each and every one of us will start using agents for our personal lives as well as in our businesses, whether small or large businesses. And the implications of that are so broad that nobody is ready for that, whether the implications on the workforce, what exactly people are going to do, what kind of IT resource will we need, what are the safety and security measures that needs to be put in place, again, both on the individual level and on the company level, what does it mean for websites and the internet as a whole if there's going to be less and less people and more and more agents crawling to collect information?

15:11What does that mean to basically everything we know on how digital business is run today? And that is going to change this year or worst case scenario in the next 18 to 24 months. And we need to start thinking about this as people, as a society, and as business leaders, because it has profound implications on, as I mentioned, more or less everything we know. The last topic before we dive into the rapid fire is Anderson Horowitz released their six months, the top 100 Gen AI consumer apps. That's their fourth edition, and they released it on March 6th. OpenAI is still the king and still at the top of the list with 400 million weekly active users by mid-February of 2025, doubling from 200 million in just six months.

15:51This is an insane growth that is showing how powerful the tools and the brand are of OpenAI and ChatGPT. The surprising thing is that number two, you have DeepSeek. So DeepSeek did not exist on the previous list six months ago, and they're now ranking number two on the web usage ranking. But they're not the only newcomer to the list. 34 % of the top 50 web AI products are new on the list from six months ago. This is just showing you how dynamic the AI world is. And with the ability to take an underlying model and develop something new on top of that from a tooling and capabilities perspective is relatively easy today.

16:30And with not a lot of modes, the modes are around maybe developing the big models and the frontier models, but even that is not questionable with the ability to distill models and create things like DeepSeek. But again, 34%, one third of the top 50 AI applications were not on the list just six months ago. Two categories that showing big growth on the list. One is AI video tools that have evolved from glitchy experiments for geeks in early 2024 to tools you can actually use to create small business videos or even enterprise level videos for actual business use cases and definitely for personal usage.

17:05So AI video tools are high in the category. And the other, as we mentioned before, is vibe coding tools that allows you to build applications without knowing how to code. And as I mentioned, an episode is coming on that topic sometime in the next few weeks. On the mobile side, there are even bigger surprises. ChatGPT still rules the mobile side as well. But in number two, you have Microsoft Edge. And in number three, you have Photomath. Both are tools that I, A, don't really know, and B, definitely wouldn't count as AI tools, but they are on number two and number three on the list. I will share the link to the entire list with you.

17:37I must admit, I don't know. A lot of the tools are in the top 50 on both web and mobile, and it's interesting for every one of you to go and take a look to see what other people are using. And now let's dive into rapid fire. The first rapid fire item is Ilya Saskover with his company SSI just raised another$2 billion at a$30 billion valuation, which is six times the value from just five months ago, the company that he founded just recently. And those of you who don't know the story, Ilya Saskover was one of the leading scientists in open AI. Many people are claiming he was the most important scientist in open AI.

18:12He went silent after the clash that got Sam Altman out of the company and then back into the company. And then about six months after that event, he left the company and founded SSI, which stands for Safe Superintelligence. He surrounded himself with a bunch of really talented people. And just because he is who he is, maybe the most capable AI scientist in the world today, but if not, definitely one of the top five, he's able to raise whatever money he wants. Now he hinted, and I'm quoting, that he's climbing a different mountain from OpenAI's methods with early signs of promise. So what does a different mountain means and what does early sign of promise means?

18:47I don't think anybody knows, or maybe very few people knows. They're actually asking all their employees not to even mention that they're working at SSI. So nobody has SSI on their LinkedIn profile, as an example. So they're trying to keep everything very harsh. They have about 20 employees, which is relatively little, but because it's Ilya Saskover and because he has promising starts, he was able to just raise $2 billion and get a$30 billion valuation. To make this more extreme, he's not planning any products and any early releases until he achieves SSI. So a very interesting investment of a lot of money by some of the leading companies in the world, just because it's Ilya and his idea to do something different than everything we've seen so far.

19:26Another interesting topic has to do with AI laws and bills that are put in place. So a staggering 781 AI-related bills are pending in the US states and federal level since the beginning of 2025. So to put things in perspective, the entire year of 2024 had 743 bills. And again, now we already have 781. We're not even at the end of Q1. To compare that to 2023, 2023 had under 200 of these bills. So the urgency is very, very clear. The current administration at the federal level is taking a very clear approach to let them run as fast as they can, because we need that for national security. And so it leaves a lot of work at the state level, which is not necessarily good, but there are multiple attempts across multiple states to define what is AI, what is the risk, what are the limitations, what can and cannot be done on the government level, what can and cannot be done on industry level, and so on and so forth.

20:23This will probably create a whole crazy patchwork of laws that will make it harder and harder for companies to navigate. This may slow down innovation of companies in different states and may force companies to move from one state to the other, maybe more than once, in order to be able to keep on doing what they're doing. That is not a good thing. But again, the current political situation on the federal level will probably push this even further. If you remember, middle of last year, we talked a lot about SB 1047, which was a AI safety bill that was raised in California and actually passed the vote, but was vetoed by the governor because he thought it was slow development and was slow innovation, which is one of the most economical engines of the Bay Area and California as a whole.

21:04What will happen in other states? I don't know, but it's definitely a story worth following. Another interesting big piece of news is Larry Page, one of Google's co-founder and the world's eighth richest person just launched another company. The company is called Dynatomics and the goal of Dynatomics is to harness AI in order to highly optimize design and shorter the time from idea to creation of products in the manufacturing and factory production world. I think what we're going to see is more and more such initiatives with people with background in tech and with money that will team up with other people with niche knowledge in specific area, in this case, designed for manufacturing, and will create things that will dramatically revolutionize entire industries that have been running roughly the same for decades.

21:48Now, we've been talking a lot about Google in the past few months. I told you since the beginning that I think Google is going to end up being the front runner of this whole revolution because they have everything they need. And they've been at work releasing more and more stuff, more or less every single week for the past three to four months. Some of them were more significant. Some of them were smaller. Right now, Google actually lost the top two places that they held at the chatbot arena. So right now, number one is Grok 3, which I actually really like and I use a lot behind that GPT 4.5, but they're still holding the next two spots and they're holding four out of the top nine ranked models, including one of their latest releases, Gemma 3, which we're going to talk more about in a minute.

22:27But first, let's talk about some cool new features that are finally coming and are putting together piece by piece the big promise of AI within the ecosystem of a business. So the first one is they just added Gemini-powered side panel to Google Calendar. It was available in Google Docs and Google Sheets and Google Drive. So now there's a side panel for Gemini in Google Calendar as well. And it allows you to ask questions such as, when was my last meeting with Sean? Or how many meetings do I have on Monday the 23rd? Or show me when do I have openings on Tuesday? And stuff like that. And it will give you immediate answers without you having to actually go and browse that.

23:01So think about a scheduling personal assistant that knows everything about your calendar and can provide you immediate answers. Now, this is not available on its own. You have to sign up for it on Google Workspace Labs to test it out. And then you have to activate it per user. And it's only web-based. There's no mobile support for it yet. But I see that as a very big promise. Another really cool feature that is coming to Gmail is a button that says add to calendar. So when somebody sends you an email suggesting a meeting, so it's not a meeting invite, but he's saying, what do you think about jumping on a call on Monday at 3 p.m.?

23:32Once it sees that, it will add a add to calendar button. If you click that, it will actually open a sad panel that will fill out a meeting details with everything it knows from the context of the email. And all you have to do is click yes, and it will create the meeting for you. When I mentioned earlier that this is the promise of everything, the whole point of having a unified Gemini or co-pilot in your workspace is that it knows everything from everywhere and can start connecting the dots. I thought that we will be there by the end of 2024. We're very far from that. So I was very wrong on that particular prediction.

24:05But these are early signs of exactly what might happen, where the system will be able to connect the dots between different aspects of your entire work ecosystem, whether you're in the Google ecosystem or the Microsoft ecosystem. And we'll be able to do these really cool things for us, like managing our time and connecting from what was said in an email to what was said and summarized in a meeting to what's available in our CRM and be able to work all these things together, turn that into agents, it will be able to actually handle a lot of the work for us. As a really busy CEO that's currently running two and a half businesses, I can tell you I can't not wait for this to finally be here.