301 | The AI That Builds Itself - And the Government Hand That Pulls the Plug, the largest IPO in History, and AI companies legally liable for hallucination outcomes, and more AI news for the week of June 12, 2026

Hello and welcome to a 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 before I get started with this week's episode, which is really interesting, I owe you an explanation. Why the hell did I do a regular episode on Saturday and a news episode on a Tuesday when for the last, I don't know, two-plus years, since we started adding the news episodes, it was the other way around?

the reason is obviously episode 300. So it would have been really stupid to release episode 300 in episode 299 or episode 301, and so it was important to me to release episode 300 as episode 300, which happened to be on a Saturday, so that's why we released it on Saturday. I'm doing the weekend news episode on a Tuesday, so you're gonna get it on Tuesday, despite the fact that it really includes only the news from the previous week. And we will be back to the regular schedule with a Saturday news episode this weekend, and then back to the regular schedule of the specific learning how to do things with AI for business next Tuesday.

So that's a quick explanation. This episode is brought to you by the multi-agent orchestration course. We are about to launch the July cohort, which is already sold out and a little bit over that as well, and we're almost sold out of the August cohort as well. So if you're interested in taking the course and you do not want to wait all the way through September, and trust me, you do not want to wait all the way through September, you should look for the link in the show notes and sign up as quickly as possible.

We just finished a cohort and it was extremely successful. Everybody were extremely excited, was able to build amazing things. People got new contracts that they didn't get before. One person found a new job using the things they've learned, and so on.

So people are making huge updates and upgrades to their career and to their businesses using what they've learned, real impact during and immediately after the course. So it does provide actual hands-on capabilities that you can use immediately to drive growth in your business and for your personal life and career. And as I mentioned, if you wanna learn how to use AI properly in this era of AI right now, so the first half into the second half of 2026, don't waste any second and come and join the course because just like all the other cohorts, this one will sell out as well, probably in the next week or so, and then the next time you'll be able to join us will be in September.

So don't wait. Click the button in the show notes and come and join us. And now let's talk news. There are a lot of really interesting things that happened last week, but I'm going to try to combine five of them into a single story.

That single story will tell the story of where we are process, and we're gonna talk about acceleration. We're going to talk about the current latest model. We're going to talk about how that impacts a specific industry as an example of what's gonna happen in other industries as well. And we're going to talk about how that impacts money and lots and lots of money, and we're going to talk about how that impacts the government and its decisions.

So it's going to be a long, deep dive that touches and connects multiple points in order to tell a story that will hopefully let you understand where we are and most likely where we're going in the near future. So let's get started. So the first part of the story, or if you want the engine that drives all of it, is obviously developing new AI models. And Anthropic released a paper called When AI Builds Itself, and in this paper, what they're talking about is how they see recursive self-improvement and how close they think they are.

So the idea here is to basically tell the world how close they are and warn the world, or if you want, raise the flags or sound the sirens of how close they are or what this may mean, and they're saying they don't even know what the hell it means if they reach RSI. They're also saying they're not there yet. So I'm quoting right now, "We are not there yet. The recursive self-improvement is not inevitable, but it could come institutions are prepared for."

Now, before we dive into the facts that they provided, which are very helpful to understand because it tells you how close they are and where they are, just a quick reminder, just a couple of weeks ago, we talked about the fact they hired Andrej Karpathy, which is one of the top researchers in the AI space. He was one of the founders in OpenAI. He was one of the people behind the biggest spike in Tesla AI. Then he went back to OpenAI, and then he left to do his own thing, and now he's back in Anthropic working on RSI, recursive self-improvement.

so that tells you that they're definitely aiming and focusing on that topic on board researchers in the AI space. And not only that, it tells you how advanced they are if he chose to join them versus anybody else because he can join anybody he wants. So here are a few quotes and data points from the article itself telling you how advanced AI became in creating code to create a new version of AI. So right now, more than 80% of the code Anthropic merges back into the code is written by Claude itself.

This is up from single low digits before Claude Code launched in February 2025. So in less than a year and a half ago, it was single digits. Now it's 80% of the code that writes the next version of Claude is generated by Claude itself. A typical Claude Anthropic engineer now ships eight times more code per day than in 2024.

Eight X the capability to generate new code compared to just two years ago. They also talked about how quickly AI is improving in doing longer and longer term tasks. If you remember, we talked about in this podcast several times about a company called Meter, M-E-T-R, a framework to test how well AI is working from a long horizon tasks capability. And the latest information we got from them was that every seven months, AI double- doubles the amount of time it can work on a task.

it's not the amount of time it can work on a task, it's the amount of time it would take a human to work on that task. And in the METER test, they are looking at a 50% success rate, which sounds really low, but it doesn't really matter because success rate in every test point that they're doing, and all they're testing is to see how long, how much time the AI can work to achieve that level of success. And Anthropic is saying that based on everything they're seeing, now AI's ability to do long-term horizon tasks are doubling every four months, meaning that three times a year, AI is doubling the amount of time that a human can do a task.

So as an example, if in the beginning of the year it was, let's say, an hour, then after four months it would be two hours, after another four months it would be eight hours, and after another four months it's gonna be 12 hours. That's in one year. But that pace is also accelerating because as I said, the last information, they had data points that we had was every seven months. Now it's every four months, which means it's gonna be less and less as it move forward.

Now, the specific information they provided is that in March 2024, Claude could do about four minutes of software tasks. A year later, it could do 90 minutes. A year after that, it's now 12-hour tasks that AI can do on its own, and they're talking about week-long tasks by 2027. That's next year that you'll be able to let AI loose on a task that takes it a full week.

Now we're gonna talk about what that means compared to human work, including in research, so you'll understand how impactful that is To give you an idea what that does to software development, they give another data point that says that in April 2026, Claude on its own shipped 800 plus fixes that cut a class of API errors by an order of magnitude of 1000x. So three orders of magnitude improvement in the amount of errors it had by shipping 800 plus fixes. And the work supervisor engineer who supervised the process estimated that it would take a human four years to do, and it took Claude one month to complete the task Now, Claude success rates, so now I'm talking about quality instead of quantity, hit 76% in May of 2026.

That's 50 points up in just six months in doing the same thing And now that I said, what does that mean for research itself? So on a fixed research optimization task, which is a parameter that they've been tracking, they're saying that Claude went from a 3X speed up in May of 2025 to a 52X in 2026. The model of the latest one, the 52X, was Mythos Preview, the one they did not release to the public yet, and we're gonna talk about the release and the fallback in just a minute. They're saying that a skilled human needs four to eight hours to hit 4X And now Mythos can do it in 52x.

or the way they stated it, and I'm quoting, "Claude has gone from super helpful to superhuman in under a year." So what do humans still do in the process if Claude is getting so much better? They're saying that the main thing is taste and judgment in the research, choosing which problems matter, to trust, when to quit when they think it's a dead end. So things that just require nuanced experience and a feel to where things are going from a research perspective.

But they specifically said that they do not think this is a technical or a scientific gap. It's just a matter of time until Claude can do that as well Now they've broken the potential impact to three different categories, or three, if you want, potential futures that may happen and what does that may mean to society. So the first option is that AI development stalls, which is very unlikely, but it's a possibility. But what they're saying, and I agree with them 100%, that even if AI development is frozen today, meaning we make no additional capabilities possible through AI, a 100-person company can do the work of 1,000-person company because every employee can sit on top of a pyramid of agents that will do specific work.

Again, Anthropic themselves are saying this is very unlikely that it would happen because they definitely see AI continuing to grow and improve. Option number two is compounding efficiency gains, so AI development gets mostly automated. set the direction, and then 100-person company can do the work of 10,000 to 100,000 employees. Anthropic says this is the most likely, at least where they see things right now So again, just put things in perspective.

That's 100 people, company doing the work of a really large enterprise with tens of thousands of employees. And then they're talking about full recursive self-improvement, which AI basically design its own successors completely autonomously, and then the pace just becomes faster and faster because each version gets better, designs the next version faster, and so on. And humans basically move only to oversight and verification, and even that is very questionable because the pace is gonna be so fast and the amount of new code is gonna be so vast that it's going to be very hard to do.

And what they're saying, and I'm quoting, "We do not have good intuitions for what this world would look like." So the most capable researchers in the world that are spending their time thinking about this day in, day out do not know what is going to be the outcome of the thing they are developing and that we will most likely get, if not from them, then from somebody else Now because this is what they believe, they said the following, and I'm quoting, "We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development."

This is not the first time we are hearing this. We heard Yann LeCun talk about this. We definitely heard Demis Hassabis, talk about this multiple times on how he thinks we are running way too fast, and you heard me say this multiple times. But the reality is because there is a race between the US and China for world dominance, and none of these parties will stop because they don't believe the other party will stop, and they don't wanna be the second one to get to really powerful AI capabilities So think about the nuclear arm race happening all over again, only it is more or less impossible to find new silos being dug or new tests being done on advanced engines and stuff like that.

it happens in data centers that already exist, and it's very, very hard to track what's actually happening inside them. And so it is very hard to monitor if other people are actually stopping. Hence, the chances that you will believe them their stopping is very low, and hence each side will say, "There's no way we're taking the chances, and we're gonna continue running as quickly as possible." And I think this is true for both governments for sure.

So even if the companies agree on something somehow, which again then I still believe the US government and the Chinese government will find ways to run forward to improve AI faster and faster up to the point that we will have to figure out what happens when full RSI is there and it's improving itself on a faster and faster cycle But to get an idea what does that mean, then Anthropic themselves have released Fable 5. Fable is a Mythos-5 class model. What does that mean? If you remember, we talked about in the last couple of months on a new kind of model that Anthropic has developed that they've called Mythos, that they said is too dangerous to release to the overall public because it generates too much risk across multiple aspects, mostly cybersecurity, because it can find and attack more or less any...

because it can find vulnerabilities and attack them in more or less every software in the world, including operating systems and data centers and servers, and basically anything you can imagine. And so they released it to a short list of companies to test it and help them figure out what it can do and help them solve and close the loopholes that Mythos can find. So they did not release the full scale of Mythos. They've spent a lot of time in trying to put it in a box and develop the right guardrails, and they're calling it Fable-5 They also released a second version called Mythos-5 that they released in more secure environment that has been vetted to be able to use it safely.

So we have two new models based on the same model that they've released previously. Both of them are safe to the public in different levels of security Now, on the benchmarks, this is by far the most capable model out there, and the longer the task it works on, the more superior this model becomes. So it is built specifically to do long, complex, bigger tasks all on its own, and it excels in exactly that And if you want to understand what that means in the real life, Stripe, the company, had the chance to use this new model, and they share the following, that it compressed months of engineering into days.

It took a code base-wide migration on a 50 million line Ruby code base in a single day What they're estimating is that It would have taken a team a two plus months to do by hand, And Fable did it in one day In drug design, which is one of the areas that Mythos is actually used and not Fable, it accelerated parts of the protein design process by 10X and it matched or beat skilled human operators with no human assistant Nine out of the 14 protein it targeted has yielded strong drug candidates.

So again, this is a real novel science kind of environment where it can do research, really advanced stuff, significantly faster from a speed perspective and at or above the level of the top humans in the world in that task Now the model is not cheap. It costs ten dollars of - to a million input tokens and fifty dollars per million output tokens. However, that is less than half the price of Mythos Preview when it was released just a few weeks ago. And if you look at the broader picture that we are seeing time and time again, that more and more advanced AI is becoming cheaper and cheaper to use.

Again, this is more expensive than Opus 4.8, but it is cheaper than the first version of this that came out and was available not to the public, but to specific companies. And I assume the price will keep on going down and the capabilities will keep on going up because that's what has been happening so far. That being said, if you are going to run this for the amount of time that it can run, so hours at a time, and you're paying fifty dollars per million tokens, expect to get a really big invoice at the end of the month so far we talked about the technical aspect of this and the potential implications of that and the models that it is able to develop faster and faster.

Now let's talk about the financial impact of this. In order to make this thing work, to develop these kind of next versions frontier AI models, it requires a lot of money. It also requires a huge amount of money to have the compute to allow the rest of the people in the world, us, to be able to use these new capabilities because they're consuming a crazy amount of tokens when they're doing these long-term tasks. And this is why, as we know, three out of the four main labs are going to go through their IPO.

One already did, SpaceX. We're gonna talk about this afterwards. And the other two has independently applied in a confidential way for their IPO. We're talking about Anthropic and OpenAI.

OpenAI went first. Anthropic went about a week later So let's talk numbers a little bit to understand what is that generating from a valuation perspective. OpenAI recent formal valuation was 852 billion post-money based on their latest round. Anthropic has surged over the one trillion in secondary markets when they raised and sold stock.

They just did another bridge raise to push them beyond OpenAI's valuation and as a bridge, if you want, to their IPO slot Also, as we mentioned two weeks ago, Anthropic is stating that they most likely gonna have their first profitable quarter, which is two years ahead of their plan for profitability, which is showing you how crazy fast their revenue is growing But they are still burning through crazy amount of cash, and as I mentioned, they just raised another $65 billion with 36 billion already allocated in debt to complete this round that, again, is gonna be the one that's gonna get them to the IPO So we're expecting two huge IPOs, each and every one of them north of a trillion.

There are some discussions about two trillion, but it's probably going to be closer to $1 trillion for each and every one of them. This is squeezing more or less every dollar of liquidity out of the markets right now, especially that it comes right at the back of the crazy monstrous IPO of SpaceX that we're going to talk about in just a few minutes What does that mean? It means that if you have some cash to invest, and I'm not giving any investing advice in this show and for anybody, but you need to think about where do you wanna put the money?

Do you wanna split it up between these three companies? Do you wanna stay out completely because you think the, valuation right now is completely inflated, or you think that AI is just in the beginning of its implementation phase and everybody will need significantly more AI and significantly more tokens, and hence the current valuation is actually very reasonable. and if you compare it to, let's say, Apple, Google, or Nvidia, there's still a lot of room for growth for the companies who are going to provide the intelligence for everything in the future so the race is not only for who develops the better model and for what purpose, but also for a crazy amount of money that is needed, not just as a trophy, but is needed in order to run these machines that will allow us, all of us, to use this technology effectively so what does this mean to the actual economy?

how will this trinkle into what we know today, and how will that change what we're doing today? Well, we got a very interesting example from Thrive Holdings. It is an AI-focused holding company that spun out of Joshua Kushner's, Thrive Capital, and they're planning to commit one billion dollars to buy up local brick-and-mortar accounting firms across the United States and supercharge them with AI operations So they've already developed an AI accounting platform that they're currently calling Current, pun intended It is powered by OpenAI, and it has already processed over 7,000 tax returns this last tax season and cut the tax prep time by one-third with a 98% accuracy, meaning it can do the basic accounting work very, very quickly or quicker for specific humans, which means every human can have more clients, do more work, and make more money by the use of their software.

There is zero doubt in my mind that while they're doing this, they're learning how to automate the other two-thirds of the effort in order to get it to do the rest of the work. What does that mean for the accounting field? it means that as these tools are being used more and more because people are gonna buy the businesses with their people, they're doing two things. They're not buying the people.

They're buying two things. They're buying training for their models so they can do the rest of the work, And they're buying their book of business. They're buying clients. So the next year they don't need the accountants at all, or they need them a lot less, and three years from now, 100% they won't need them because you'll be able to train on all the data what they're doing.

So this gives you an idea of what these models are capable of if there's enough capital behind them, and because there's shitloads of money to be made, and that's a professional term, there is going to be the people who's gonna raise the money in order to do the thing. Now, while accounting is a very obvious use case of just evaluating numbers and finding where they fit and so on, it is just one domino after the next. So we started with software because that's what the AI labs needed, and now it's accounting, and then every industry is gonna go after that.

So if you want an example of what this is going to do, this is a great example of what is going to happen. Lots and lots of money is going to be invested in training the models on specific fields until they get perfect in that field, and then I don't have a clue what the people in that field are going to do. But it is not going to do the same profession that they had before. Definitely not doing the same thing, definitely not at the same volume and scale.

And yes, can a company like Thrive Capital now take over a huge amount of the accounting field? 100%. What happens to these people a few years down the road? I don't know.

What happens to the larger companies they're competing with? I don't know. Nobody has the answers for that, and as I mentioned, not even Anthropic when they're developing recursive self-improvement that is going to make this thing look like a kid's toy Now, why accounting? Beyond the fact it is an obvious use case, CPA candidates has fell 30% in the last six or seven years.

That's a very big decline, and there's a very big shortage. so over there in that specific industry, it makes perfect sense. Again, it's a relatively easy use case with very clear benefits, with a shortage in supply, so it's a great first step. But as I mentioned, the next one and then the next one and then the next one, it's just a matter of when the ROI starts making sense for people with really deep pockets that can raise crazy amounts of money will go into these fields, and the playbook is very clear.

Let's get the people, let's get their business, let's use it to train the AI, and then we don't need the people anymore. So now we've covered the improvement of the technology. We covered where does it take us? What kind of models does it build?

We covered what does that do to obviously the labs themselves, but also specific industries, and again, it's gonna be every industry. It's just a matter of time. The last component that I wanna talk about is the government aspect of this. So I shared with you that the White House was working on a bill to monitor the progress of AI, and it was on the table and off the table and back on the table and back off the table one hour before it was supposed to be signed.

eventually, a version of it was signed. It's an executive order that is titled Promoting Advanced Artificial Intelligence Innovation and Security, and it was signed on June 2nd of 2026 Now, what the executive order actually does is it creates a voluntary framework for the labs to have their most powerful models designated covered frontier models. That's a quote from the actual executive order. if a model qualifies, the lab can give the government up to 30 days of early access before the public release of the model.

So the agencies, the government agencies can evaluate its cybersecurity capabilities Now, the government framed it narrowly on purpose, so it's only models that are meaningfully a step change forward in cybersecurity, not just any routine updates that they do regularly between one week to the other But the government did clearly say that again, this is voluntary. So I'm quoting, "Nothing in this section shall be constructed to authorize," and I'm continuing a few 10 sentence later, "a mandatory governmental licensing, Pre-clearance or permitting requirement So what they're saying is that if you want another group, let's say the government, to look at what you're doing, kind of like as another set of eyes to verify that what you're releasing is safe, we will gladly help you do this.

Now, another section of that talks about AI cybersecurity clearinghouse that includes Treasury, NSA, CISA to coordinate the findings and patching of vulnerabilities that are found by evaluating these models Now, I will say something about the voluntary aspect of this. When the government says something is voluntary, you can obviously not participate, but there might be and probably will be consequences, right? So it's not mandatory, but if you don't obey, then you might find yourself on the wrong side of the government, and we see what's happening right now with Anthropic as a good example of where that might lead and to make it even more interesting, at the same week that Trump signed this executive order, he also floated the idea of the government taking financial stakes at the AI labs.

Again, this is not new. We talked about this in this podcast several times before on the government will most likely push to create some kind of its own partnership with these leading labs, so it has more control and visibility into what is happening. Will that actually happen or not? Not very clear.

We both know that Trump did this with Intel, but that was more let's save a large US institution from going under than anything else, and it worked out very well for the government in that particular case. the same scenario right now with the AI labs If you wanna take it to the next level, Bernie Sanders suggested what he calls the American AI Sovereign Wealth Fund Act, which will be a one-time 50% tax paid in stock of the largest AI companies that will give access to the public for direct ownership in these companies And he said, and I'm quoting, "Will the future of humanity be determined by a handful of billionaires, or will AI be used to make life better for working families?"

That's a very Bernie Sanders kind of statement, but on a very high level, I don't necessarily disagree with him So what does that tell us? It tells us that the government is at least thinking seriously on both sides of the aisle of, let's call it a soft nationalization of the AI capabilities. And again, I'm not surprised. We talked about this many times in the podcast.

I will be really surprised if nothing like this actually happens. What exactly is gonna be the setup? I'm not completely sure, but this is going to be the next nuclear weapons. The fact that now it is controlled by specific private entities has benefits.

They can run significantly faster, they can raise money, they can organize things in ways that the government will never be able to do in the times they can do it, but it is nuclear weapons. We don't want any company in the world that just wants to develop nuclear weapons and decides whether to deploy them or not based on a voluntary evaluation by the government, and hence why I think this will be interesting However, this soft nationalization and voluntary approach was flipped over on June 12th at 5:21 PM.

So I told you that Anthropic has released their latest set of models that are based on the Mythos version five level of models. on June 12th, the government issued an export control directive from the Department of Commerce that is signed by Secretary Howard Lutnick that is ordering them to suspend all access to Fable 5 and Mythos 5 to any foreign national. Now, since it's practically impossible to know who is a foreign national and who isn't, right now Anthropic pooled access to these models to the public, period.

You don't have access to these models unless you're working for some specific agency that has access specifically through that So we went from a potential soft nationalization, if you want to voluntarily give us access 30 days before so we can review it, to three days after a model was released, the government ordered them to pull it back and not allow access to the model So what triggered the government to completely flip 180 degrees off the direction it took just a few days earlier?

the official reason is the government believes that someone had found a way to, quote-unquote, "jailbreak Fable-5." Now, Anthropic has reviewed the demonstration by the government, and they're saying that it's narrow and non-universal, essentially basically saying that it's not a big deal, it is previously known minor vulnerabilities, and that other public models, including OpenAI GPT-5.5, already has the ability to find these vulnerabilities, and it's not unique to the Mythos level of AI.

Well, first of all, I must say this is partly self-inflicted. Partly self-inflicted because Anthropic, when they gave the model to other companies under their Glasswing project, they said this model has huge cybersecurity risks, bigger than anything we had before. And so no-nobody bothered to go and check what other models had before. So now there's a bigger spotlight on their model Makes absolute sense.

There are two other stories that are not the official stories, but are definitely lurking in the background of all of this. One is China, and there are rumors that a China-linked group has gained access to Mythos and are using it to learn on Western systems, and that is a national security risk. Hence Hence the export control action or how it was titled, and the other is a personal vendetta, if you want, with the White House AI czar David Sacks claims that Dario Amodei refuses to fix a jailbreak vulnerability after being warned And he also accused Anthropic of growing fears in order to drive sales of its new model.

So there's also the personal aspect of this between David Sacks and potentially Dario Amodei Now, the power struggle between Anthropic and the government is not new. Obviously, we covered it multiple times here on the podcast When Anthropic refused to allow the government to use the model for anything they wanted on the military and security side, and then they got designated a national security risk. That being said, while they are designated that, all the leading government agencies, including the NSA, are using Anthropic models.

Specifically, the NSA admitted to using Mythos for things that they are doing. So the whole thing is very interesting. I'm not exactly sure what's going to happen. What I do know is that I had Mythos for three days, had the opportunity to use it for different things that I'm doing, and then it was pulled away until further notice But to summarize this long and very important, I think, section of today's episode, I will say the following.

One is that the cost for really high intelligence is falling. That being said, the cost of using it for real life is growing exponentially because of the long horizon and the agentic capability of these models. So if you look at the building block down to the Lego piece, then the Lego piece is significantly cheaper. If you look at the kind of Lego structures or machines that you can build with it, they're significantly bigger and more complex, and hence you're buying significantly more Legos, and hence the cost per company, the usage of it is actually growing and compounding very quickly.

The other thing that is very clear is that it is going to continue accelerating, and if we hit some version of RSI, which we're at the verge of anyways, we are going to see faster and faster, better and better, cheaper and cheaper models that can do more and more complex things that will trickle into the economy and the actual things we're doing. Some of it will be amazingly well because it will allow us to drive science faster than we can do right now and to diseases and solve global warming and Demis Hassabis dreams on and talks about all the time.

But on the other hand, it will have very significant implications because there are going to be the big players such as VC companies and private equity companies that are gonna dive all in and gonna gobble up this technology and use it to roll up entire industries into AI solutions that will then eliminate the need for the people... Or maybe not eliminate, but dramatically reduce the need for the people in those industries to do this actual work. And even the leading scientists, including Anthropic themselves, are saying they don't even want to speculate what that means because they don't have a clue So now I wanna jump into a few rapid-fire items, and I picked just a few.

The rest are gonna be in the newsletter, so if you wanna know everything that happened last week, go and sign up for the newsletter, and then you can pick whatever ones you wanna read. There's gonna be the links to everything in there, so if you wanna deep dive into anything we talked about, you can find it in there, and there's a link to that in the show notes. You can sign up for the newsletter. But the first rapid-fire item I wanted to talk about that I find very interesting is that a German court has ruled that Google bears direct legal responsibility for the AI-generated misinformation.

So we all know AI makes shit up. The professional term is hallucinate, right? So AI hallucinates, and yes, hallucinations have gotten better, and there's significantly less of them, but it still happens. It's part of the fact that an AI is a statistical model, and because it is a statistical model, it will never be correct 100% of the time because even if you make it better and better and better, you're gonna go from 90% to 91% to 92%, but there's still always the percentage it's gonna get it wrong.

Well, a German regional court has issued a dramatic precedent that basically rules that Google is directly responsible for false or misleading information generated by its AI overviews Now the court did this because what they're claiming is that different than traditional search, where traditional search just gives you the links and sends you to the source of the content, in the AI summaries, it is new content that Google is generating and hence responsible to verify the accuracy of the information.

And if they won't, then they will be legally liable for the whatever outcome happens. In this particular case, the AI generated false claims about a couple of people in their business saying that they're related to specific scams and subscription traps and dubious business practices, and that harmed their business, and Google was found liable for the damages to that business. Now, is this the final statement? Of course not.

This is just a regional court, but the precedent is very important. It will be very interesting to see how this evolves, because if it does evolve in this direction, it means the AI companies will have to develop mechanisms that will verify the outputs of every single thing they're spitting out, which on one hand will remove or reduce hallucinations almost to zero. On the other hand, will cost us a lot more money because we will have to pay for the re-verification mechanism. Can the verification mechanism run on a much cheaper model that can not add a lot of money or a lot of tokens to the process?

Yes. Am I building stuff like that for myself already? Yes. Does it make sense that this will be the future?

Probably. But from a legal perspective, that is a very interesting case, and I will keep on updating you as this moves forward Another interesting piece of news back to Anthropic is that they just released a fully available to the public version of what they call Claude Managed Agents, which allows to accelerate production deployment and cutting to first time token by 60% by using this infrastructure So Accelerated Development and Infrastructure Management, Claude Managed Agents enables teams to deploy production-grade AI agents in days rather than months by providing a performance-tuned agent harness and managed production infrastructure, eliminating the need for teams to burn development cycles on security, state management, and permissioning The idea here aligns with a lot of things that we discussed in the past few weeks, that the models today are good enough.

And again, Anthropic said it themselves in the first topic that we talked about today. Even if AI stops today, it is still extremely powerful in how you use it. What's the harness? What's the infrastructure?

How everything connects makes a very, very big difference. And I see this when I teach courses and workshops to companies, when there's people there who already understand how AI works, and they know how to build agents, and they know how to build skills, but they don't necessarily know how to put it effectively all together. Same thing in my courses. The recent course, the multi-agent orchestration course that just ended, most of the people there knew how to build agents and knew how to build skills and knew how to put them together.

And yet they've learned a lot because the way you can orchestrate it all makes a very big difference. And what Anthropic is doing here is providing this out of the box to companies in a way that will allow them to develop, test, and deploy agents safer and significantly faster, which means we're gonna have more and more agents if that wasn't obvious before. Another big piece of news from last week is Apple finally launched a worthy AI capability in WWDC Twenty-Six. So first of all, they finally introduced the redesigned Siri that is powered by Google Gemini under the hood, and now it has real true conversational capabilities.

It has visual intelligence compatibility. It has a standalone app alongside with existing integrations that Siri already has And it is a very big step forward, at least on paper, because we don't have access to it yet, compared to the existing Siri. That being said, something that I'm very curious about this whole thing, they are using a highly customized Google Gemini model. So technically, Google now owns the entire personal assistant market on phones or very close to the entire because it is going to run as Gemini itself on the Google phones, and it's going to run as Siri on iPhones, which means they're going to have access to everything that's happening.

It is obviously privacy first as far as the strategies, let's just like Apple does everything else. Or like Fred Federighi, who is Apple's senior vice president, emphasized that, and I'm quoting, "Data is only used to execute your request." In other words, like everything Apple is doing and everything they did so far, including the AI space, the data and privacy and security are at the top of their mind Search was completely rebuilt from the ground up and providing awareness and complete controls into what data Siri and other AI capabilities have access to They also added photo editing and AI dictation tools.

Again, nothing amazing. It's not different than Google already has, but they are closing the gap and adding more and more AI capabilities built into iOS 27 And the final piece of news for today is SpaceX IPO, which became the largest IPO in history, made Elon Musk the most valuable person in history by a very big spread. So the richest man on earth, with a valuation of 1.75 trillion in personal wealth So SpaceX itself raised 75 billion in an IPO and made Elon Musk a trillionaire Now the stock itself rose from $150 a share on the NASDAQ and closed by $160.

95. It since then came down a little bit But this gives SpaceX a lot of money, and SpaceX now is also xAI, which means this AI company will also have access to a lot more capital, which will be very interesting to see what they do with it. That is it for today. We'll be back, as I mentioned, on Saturday with another news episode, and then we'll be back on next Tuesday with the regular episodes showing you how to do things in your business with AI.

Final reminder to book your seat at the next cohort of the multi-agent orchestration course that starts in the beginning of August. And until the weekend, have a great rest of your week.