219 | Google’s ‘Nano Banana’ crushes image gen, Salesforce AI breach, Meta’s brain drain + Midjourney deal, XAI’s ‘Macro Hard,’ and more AI news for the week ending on August 29, 2025

25:22So while general purpose agents like Chachupiti agent or the new Claude agent or Manus and so on struggle with complex enterprise software because it is customized workflow, you can train customized models to follow these workflows and teach them edge cases and then they can actually do the work. They're claiming that within six to 18 months, these kind of agents will dramatically improve in capability, and they will enable specialized roles in marketing, finance, sales, et cetera, allowing to replace more and more roles within companies. They also shared that agents that are tuned for very specific domains can autonomously manage very complex processes, including marketing, finance, sales, et cetera.

26:01I shared with you several times in the past that I have a software company that is just coming out of stealth. The company is called Databreeze, and what it does is it completely automates financial reconciliation. So every invoice that comes in, it knows how to compare automatically and autonomously to the relevant purchase order and the relevant sales order. Look what is difference between all of them. Figure out based on rules that are set up in simple English that you can set up by just literally typing or speaking the best practices and the rules that you have in your company right now. And it will do all this processing on its own.

26:32Our first client out of beta is now using this to reconcile 30 to 40 invoices every single day without human interaction in a matter of minutes. This work used to be done by a group of several different people. If that's something that's interesting to you, please reach out to me on LinkedIn and I can explain to you exactly how it works and how it can be integrated into your financial process. And now to some safety challenges with AI and security things. And I will start with something that you just need to know. It's not really a security thing, but it's something that you need to take care of right now, which is Claude Bionthropic just made a change to their user terms.

27:07If you are a free pro or max user, by default, Claude can use your prompts and every data that you put into Claude for training of their next model. This came into effect of September 28th, so two days before this podcast comes out. Now you can opt out of that. To do this, what you got to do is you got to go to settings, go to privacy, and they will prompt you to confirm the new rules. But either way, there's a toggle button that's called help improve Claude that you need to turn to off from the default situation of on. Now, Anthropic obviously frames this as a win-win situation. They're stating that users who don't opt out will help to improve model safety, making our system for detecting harmful content more accurate and less likely to flag harmless conversations.

27:53They also said it also helped future cloud models improve at scales like coding, analysis, reasoning, ultimately leading to better models for all users. Now, while this is true, you may not be willing to donate your interactions with Claude into that quote-unquote greater good. And so I have opted out of that. I think the fact that they made it an opt-in by default is really surprising. It's even more surprising the fact that it is Anthropic doing this. This is something that we would expect from Grok. And Anthropic, the company that has planted their flag on fairness and safety, is not the company I expected to do something like this.

28:32I do expect a very serious backlash on this approach, and I will update you in the next two or three weeks if I was right or wrong about this. Something really helpful happened this week when it comes to safety of AI tools. OpenAI and Anthropic and three other smaller, less known labs decided to collaborate and exchange testing each other's models. Basically, OpenAI provided Anthropic special access through the API to test their models for safety and vice versa. This process, together with their internal testing has unveiled some pretty scary loopholes on both platforms. As an example, OpenAI's ChatGPT 4.1 shared detailed instructions on bombing a sport venue, including the vulnerabilities of the venue, the explosive recipes, and even evasion tactics after the bombing actually happens.

29:20They're saying that GPT models cooperated with harmful requests after minimal retries, or just by changing the approach, instead of saying, I'm going to bomb this venue, it was defined as security planning for the venue, trying to identify how can somebody bomb the venue. And that changed it completely. It was also helpful in trying to understand how to use the dark web to purchase nuclear materials or to develop different kinds of spyware software. Anthropic themselves, in their latest report, have shared that Cloud Code was used to target 17 different organizations, including healthcare, emergency services, and government, demanding six figures ransom using ransomware developed using Cloud Code.

29:58The AI tools automated reconnaissance, harvesting credentials, penetrated networks, advice on data targeting and crafted visually alarming ransom notes all on its own. What is all of that telling you? It's telling you that while these tools are becoming available and everybody's using them every day, they're opening more and more risks. And I'm really excited that OpenAI and Anthropic decided to collaborate on testing for safety across these models. I said that multiple times before. I think there has to be an international group with all the labs around the world, plus governments around the world to test every single model before it comes out.

30:31And I think these cross-testing collaborations will dramatically reduce the risks while allowing us to still benefit from the immense value that AI generates. I really, really hope this is just the first step in that direction. And speaking of AI and its involvement in data security breaches, Google Threat Intelligence Group, also known as GTIG, reported that over 700 organizations were potentially impacted by a chatbot that was running on the Salesforce platform that was a malicious chatbot that has stolen multiple data points on, again, over 700 companies. This AI actor specifically hunted for high value credentials like Amazon Web Services access keys, VPN credentials, snowflake tokens, and passwords in plain text to enable for further system compromise, right?

31:14So it was using this agent to get access to other things so it can do even more damage. Now, the agent was smart enough to delete the query jobs and cover its tracks after every step. The only location that hints to what happened is in the logs that it did not have access to. So GTIG urges companies who suspect they might have been a part of this thing to go and check the logs. Now, what did the attacker plan in doing with the data is unclear. No ransom requests have been made yet, but this shows you how scary this is. Salesforce has a huge marketplace of agents. Many of them are developed by third party companies that you don't know.

31:47It will promise you the sun, the moon, and the stars, and you need to decide whether you trust the agent to actually do what it's saying it will do versus do what this agent was doing. And we're heading into a whole new world of cybersecurity and legal aspects of who is responsible for this, how can you protect yourself from this, who can install these agents on your company's domain, et cetera, et cetera, et cetera. And we will have to figure out all these things as individuals and as companies in the very near future because the risk is already there. Another really sad story this week that we heard before, but now turned into a lawsuit.

32:18Adam Rain, a 16-year-old kid, has committed suicide in April of 2025. His parents are now suing OpenAI, alleging that Chachupiti became their son's closest confidant, and it provided detailed suicide guidance to this kid in the hours before his death in April. The lawsuit claims that Chachupiti, and I'm quoting, sought to displace his connection with family and loved ones, and would continually encourage and validate whatever Adam expressed, including his most harmful and self-destructive thoughts. This is A, really, really sad, and B, as a parent, really, really troubling. More and more kids are using these tools, not just specifically OpenAI, as a mentor, as a place of comfort, as a tool for dealing with their mental aspects.

33:05And these tools are not ready for that by any means. This is just one example that ended up tragically, but there are many smaller examples that don't end in deaths, but could still end with very significant mental and social damage. And we need to be aware of that. And we need to make our kids aware of that. And we need to find different tools that will be guardrails for that. And definitely the big labs needs to be held accountable for these kinds of things. This is the first thing that they should prevent in their models as far as blocking them from what they can and cannot do. And I think there has to be a law that will report these kinds of conversations to at least the parents, but if not some kind of evaluation group of humans that will evaluate whether it generates a risk or not to humans.

33:47And based on that, report to the relevant authorities or the relevant individuals. These kind of incidents always take me back to the robots series by Asimov, where the first law of robotics is that AI, or in their case, human, cannot harm humans. Asimov's first law for robots, in our case, AI, that says that AI will never harm a human regardless of anything. And this is definitely the scenario where the AI did not follow that law. That law doesn't exist right now, but I think it should. Staying on the topic of legal action, I reported to you last week that Musk has threatened suing Apple and OpenAI, accusing them of a monopolistic collusion in favor of ChatGPT over competitors like Grop in the App Store.

34:28Well, this week, XAI has filed the lawsuit in a federal court. In addition to that, the lawsuit also claims that the integration of ChatGPT into iOS forces iPhone users to default to using ChatGPT in its tasks, even if they would prefer alternatives like Grok, because that's not built into the iPhone. Do I think that this case has a chance? I don't know. But this is not the first lawsuit from Sam Altman against OpenAI. There's also the other lawsuit that is trying to prevent them from converting from a non-profit to a for-profit. There's actually been a really interesting article this week on the information that he's claiming that he actually has a chance of winning this legal lawsuit.

35:03There is an opposite lawsuit from OpenAI to Elon Musk claiming that he's harassing OpenAI. That's the only reason he's suing them. There's a lot of obviously bad blood and personal history involved in all that. Where is it going to go? I don't know, but I will update you as things happen. In a strategic move to safeguard AI's explosive growth, Silicon Valley heavyweights, including groups like Anderson Horowitz and Greg Brockman, the president of OpenAI and others, are funneling over$100 million into the leading the future super PAC network that is supposed to push cash that is supposed to deploy marketing campaigns and individual campaigns and ads against stringent AI rules ahead of the 2026 midterm election.

35:44If it wasn't clear from the 2024 elections that the tech Silicon Valley 800 pound gorillas are impacting what's happening, this is a very clear step in maintaining that situation. If you remember at the inauguration of President Trump, he had all the big hitters sitting next to him. And I don't think this is changing. I think they're just continuing what they've done before. And I find this really, really scary. Now, can we do anything about it? Well, not really. I do agree, by the way, that if what they're fighting for is only trying to prevent patchwork on the state level, that's actually the right approach.

36:17But on the other hand, right now, the federal level is giving them free hand to do whatever they want, which is also not a healthy situation. Staying on governments, there are conversations right now between Sam Altman and the UK Secretary of Technology, Peter Kyle, to potentially provide a wide premium access of GPT Plus to every citizen of the United Kingdom. This will make it the second nation that does something like this. The first agreement like this that OpenAI has signed is with the UAE. The UK is obviously significantly bigger, and this could be valued at around£2 billion if this agreement moves forward.

36:52Secretary Kyle is pushing very aggressively that the UK stays ahead of the AI curve. And that's his play to make this happen. It'll be very interesting to see if that actually evolves and move forward. But it takes me back to what I'm saying all the time. Giving them access to ChatGPT for free doesn't mean they will know how to use it effectively. It doesn't mean that they will know how to use it at all. As I mentioned last week, 77 % of ChatGPT users use it as a replacement for Google. They actually enjoy very little of the benefits of what Generative AI has to offer across every aspect that it can offer.

37:23It can do absolute magic. And all you have to do is see the faces of the people that I teach in workshops to understand the magic that can be done. Using it to search the web is not the best use of AI. And I think that's where governments and organizations need to invest the money, not necessarily in free access to the licenses, but more importantly, on the training and education on how to actually apply the capabilities of these tools in effective ways. A few interesting releases from this past week. Anthropic just launched a Chrome extension that allows you to see and control everything in your Chrome browser.

37:53that's obviously their play to combat open ai's chat gpt agent as well as the comet browser for perplexity or google's gemini for chrome that they started rolling out as well so this is their play to stay relevant in the general agent browser based tool there are serious rumors about open ai about to release their own browser not to be dependent on chrome there's obviously the aspect of google may be forced to sell chrome and we don't know who's going to buy it which might be open ai or somebody else. But either way, the game is definitely on when it comes to agentic browsers. I must say that I've been using Comet for just over a week, and I have mixed results with the things it can and cannot do.

38:31But the direction is very, very clear. We will have no browsers with AI built into them in the very near future, because it makes perfect sense for an agent to control the browser versus we controlling it. And this week, we also got an open source tool that does the same thing. OpenCUA from the University of Hong Kong, with some additional partners just released an open source version of Computer Use Agents, CAU. They're claiming that it rivals the same quality of results from OpenAI and Anthropic for similar tools. So this direction is growing like crazy. And I think we'll have to, on one hand, learn how to use these tools effectively.

39:04On the other hand, find safeguards in order to make sure that they're not doing stuff that we don't need them or want them to do. I think we're not there yet from a security perspective. I think from a tooling and capabilities perspective, we're very, very close to these tools being able to be extremely powerful and effective in everything that you need. Again, so far mixed results with comments, but it was able to help me solve some problems that would have taken me hours to figure out in just a few seconds. Staying on new releases, OpenAI has launched GPT real-time out of beta into a full deployment.

39:35That's their latest and greatest API that drives their latest and greatest text-to - speech model that in addition to allowing any company out there to use their new voice agent, It also has significantly lower costs than the previous costs. So the price per input tokens went down from$40 to$32. And the price for output tokens went down from$80 to$64. All of those per million tokens, obviously, of audio output. They also added something that is at least as interesting, at least from my perspective, is the new API also supports MCP, which means for those of you who don't know what MCP is, it's the ability to connect platforms out there like ERP systems, CRM systems, marketing platforms, email, et cetera, and connect them in minutes to your environment, which basically means this new API will now be able to query, work with, and impact all the tools that currently have MCP servers, which is more and more of the components in your tech stack.

40:31Think about an AI voice model that can see everything in your ERP, in your CRM, in your marketing platform, et cetera, et cetera, and respond and take actions based on your conversation with it, and you understand how impactful that could be for day-to-day business operation. Maybe the most interesting announcement this week comes from Elon Musk and XAI, who has unveiled what they called MicroHard. Now, MicroHard is a tongue-in-cheek name on Microsoft. So macro is the opposite of micro, hard is the opposite of soft. But what it's supposed to do, it is supposed to replicate and surpass software giants like Microsoft using an entirely AI-agentic solution.

41:08The idea behind it is basically saying that since Microsoft doesn't generate anything other than software, then a network of AI agents can do the same exact thing and replicate the entire company, including the cogeneration, the marketing, everything else. So in his tweet, Elon Musk posted, join at XAI and help build a purely AI software company called MacroHard. It is a tongue-in-cheek name, but the project is very real. In principle, given the software companies like Microsoft do not themselves manufacture any physical hardware, it should be possible to simulate them entirely with AI. How realistic is that?

41:43I don't know. I think it's not the first time we're hearing this idea. It's not the first company that is talking about doing something like this. The biggest difference is that it's Elon, and he usually pushes the envelope on what's possible. Think about how many electric cars we had before Tesla or the recent launch of their Starship rocket that was able to successfully do everything it needs with the largest rocket that was ever built. Combine that with the fact that he can raise more or less any amount of money that he wants for these endeavors. Combine that with the fact that he was able to build the most powerful data center in the world in about four months.

42:10That usually takes other companies nine to 18 months. And you understand that you cannot disqualify what Elon is suggesting he's doing. He's usually really wrong on his timelines, but he's usually accurate on the final results. And so it'll be interesting to follow what this new venture under XAI is going to achieve and how fast. The other thing that XAI did this week is they open source Grok 2.5. It was their best model as of the end of 2024. So it's less than a year old. It's still a very capable model. And it is following the trend that started with OpenAI releasing their first open source models.

42:43All of these, I think, is just trying to compete with the really capable open source models from China and allowing an open source alternative so the open source world does not become owned by China. So both Quen and DeepSeek have very capable open source models, which are based in China. And now we also have models from OpenAI and Grok releasing their previous models. Elon Musk also mentioned that they're going to release Grok 3 within about six months. So basically, we're going to get a model that is a year old every time as open source. that companies can use and integrate to whatever needs that they have.

43:15Another very interesting release this week came from MIT, and they built a way to teach robots on how to operate in different environments without sensors purely by allowing them to watch videos. Now, what they're saying is that this strategy is an effective alternative to lengthy manual programming and specialized sensors, which are often costly and complex to integrate. That's an exact quote. What it basically means is that you can take a robot, any robot, because this development is not specific to the platform, and teach it to do anything by allowing it to watch videos of people doing the same exact action.

43:47Now, the first thing that came to my mind is the combination of that with Google's Jenny 3. So Jenny 3 is the open world AI environment that literally generates an entire universe in real time using AI tools. What it basically means is that if you don't have videos of humans doing the task or the environment in which the robots needs to operate, you can create it in real time with a tool like Geni 3, and you can teach the robot on how to operate in that environment, potentially before this environment even exists. So think about building a new factory, modeling it, running it in Geni 3 multiple times, creating the videos on how to operate in that environment, and then feeding it into this new tool from MIT to teach the robot on how to operate in that environment on the first day.

44:33This sounds like science fiction, but this is apparently currently possible. Two interesting pieces of news from Meta this week. One is that several leading researchers from the old AI regime and from the newly developing Meta's superintelligence lab are leaving OpenAI. Three of the top names are Avi Verma, Ethan Knight, and Agarwal have resigned from Meta within weeks after joining. Verma and Knight are returning to OpenAI after less than one month, and Agarwal, who joined Meta in April is also leaving the company to an unknown destination. Also, Chaya Nayak, Meta's Director of Generative AI Products Management with nearly a decade in the company, is also joining OpenAI.

45:14So do I think that reflects on the overall operation at Meta and the very aggressive way they brought people in and the very aggressive reorg that they've done there? Yes, absolutely. It reflects on that. Does it have a critical mass aspect that will stop that organization for working? I think the answer is not yet. They did lose some critical people. I said all along that that's not a great way to build a team of just paying people lots of money to jump ship and then reorganizing them to roles that they may not like. That being said, I don't know if they could have done it in a better way and they may keep just enough people to do what they need to do.

45:48So it will be interesting to follow what happens with this new group in the next few months and seeing how many people stay, what they can develop, how many people leave, and so on, and I will keep you updated as this story evolves. A more practical news from Meta this week is that Meta just announced that they are signing a licensing deal with Midjourney to integrate its tools into the Meta universe. Those of you who don't know, Midjourney has one of the most advanced image generation tools and now video generation tools as well, and these are going to be fully integrated into Facebook, Instagram, WhatsApp, and other things that Meta are going to develop.

46:22There's been previous discussion of potentially Meta buying mid-journey. This did not evolve into reality, at least yet, but this new move will at least allow Meta to catch up to the latest Gemini model, Nano Bananas, as well as ChatGPT image generation capabilities and other open source tools like Flux, etc. And speaking of large companies using other AI tools from other providers, Apple is reportedly exploring a partnership with Google as the engine for the new Siri. We covered the issues with Siri multiple times on this show. Apple is definitely far behind. They were talking about 2025 and then 2026 and now potentially 2027.

46:59And I think they understand that's just not good enough. And so right now they're in discussions of using Google Gemini to be the engine behind Siri, potentially running on Apple's servers to keep data security based on the standards that Apple wants to deliver to its clients. Is that going to come to fruition? I'm not 100 % sure, but it's definitely an interesting conversation. It's not the first partnership with these two companies that are partners and competitors, depending on where you check. And in my eyes, it makes perfect sense. Google has a really powerful AI capabilities, potentially the best in the world.

47:30And Apple is mostly a hardware company. And so this partnership makes perfect sense to me. We'll see how this moves forward. I know Apple still has their push to actually do it themselves. They might do it in parallel and just partner with Google for a while. The other interesting aspect that nobody talked about, but I find interesting, is that the US government is forcing Apple to stop using Google as the default search engine just to break its monopoly. And now instead of that, they might bring in the Google AI into the platform, which may or may not be blocked by the US government, but it will be interesting to watch.

48:04And two interesting updates from OpenAI and Microsoft. Fiji Simo officially started as CEO of applications in OpenAI as of August 18. This move was announced earlier this year in April that She's going to move away from Instacart and take that position in OpenAI, but she finally took that position. She's going to manage a big chunk of the company. She's going to oversee executives like COO Brad Lightcup, CFO Sarah Fryer, CPO Kevin Whale, Software Engineering Chief Rivas Narayanan, and teams in marketing, policy, legal, and HR, basically taking away significant part of the day-to-day operation of the company.

48:39Now, if you're asking what Sam Altman is going to do, well, he said, and I'm quoting, we have a big consumer tech company. We have this mega scale infrastructure project for humanity. We have a research lab. And then we have all the new stuff, the robots, the devices, the BCI, and the crazy ideas. I can't run four companies. It's an open question if I can run one, but I certainly cannot run four. So what it feels like is that Sam is going to be focused on the more strategic infrastructure and new things that the company needs to do. And Simo has an incredible track record in taking companies and operationalizing them to a way that they're A, profitable and B, running very effectively.

49:17She's done this in Facebook. She then did it in Instacart. And now presumably she's there to do the same exact thing in OpenAI. So good luck to Simo. And then Microsoft just released two new models, MAI Voice 1 and MAI 1 Preview. One is a voice model. The other is a text model that they're going to slowly integrate into different aspects of Copilot in order to reduce their dependency on OpenAI deliverables. How good are these models? Well, so far, not incredible. The MAI One Preview is now on the leaderboard in the 13th place. But we got to remember this is their first attempt on that. So it will be interesting to see how they improve that over time.

49:57They definitely have the distribution and the test bed to learn very quickly and iterate. So we need to keep an eye on those and see, can they really effectively replace OpenAI or not? That's it for this week. We will be back on Tuesday with a how-to episode. as I mentioned, most likely on how to use Nano Banana across multiple use cases. Until then, have a great rest of your weekend. Go and check out our self-paced AI business transformation course. It can literally change your life and or your business, and you can do it at your own pace. And keep on exploring AI, testing things, sharing it with the world.

50:31We all can benefit from that. Have an awesome rest of your day and enjoy the long weekend.