Hi everyone! We’re on the ToTheMoon channel. Technology news and insights from Silicon Valley and around the world. Our weekly podcast, though by now we’ve got episodes coming out on the channel every other day. By the way, whoever watched the episode with Max Grigoriev, we filmed it this week, a really cool episode, I think. And there was an episode a couple of days ago where I compared all four modern agentic systems, such as Codex, Claude Code, Antigravity and Cursor. So, what do I want to start with today. I’d, you know, only just made the episode about Antigravity and about Cursor, and about the comparison with Codex.
And OpenAI ChatGPT Dots had only just come out, it seemed. And then, you know, a totally fresh announcement. Claude is releasing Duet. By the way, I think the name Duet is somehow kind of unappetizing, it’ll be hard to say, like Dots. But Claude is releasing Duet. Right now only people with, like, the top-tier subscriptions can join it, and also through a waitlist. And personally, as of this recording, I’ve only managed to get on the waitlist, although maybe they’ve already upgraded me there, I don’t know. We’ll see what happens in terms of Duet.
We’ll discuss it today as well. And what do I want to start with? So, I want to start with the following case.
Walmart is one of the largest chains, retail chains, in the US and in the world. It has created this thing, a separate entrance for AI shoppers, the ones who use artificial intelligence. Essentially, a separate entrance for AI shoppers, right? And they announced this thing called Agent Profiles and Agent Gateway. I mean, they said there’ll be a scenario. I mean, how does it work now? Right now, if an agent wants to buy something at Walmart, it essentially has to go in through a browser, roughly speaking, under your login and make the purchase.
And obviously agents get blocked in a lot of places right now. For example, Amazon blocks agents very seriously. You can’t look at prices, you can’t perform a large number of actions, and so on. And so Walmart has made this access. What do you think, will this access work? And what our viewers… There’s a reason I want to start with this topic, because it’s a question of a certain bet. Somebody did come up with this case, after all. What do you think, will an approach like this work? Because I sat there debating it with myself, how this is going to work.
I’ll come and tell the artificial intelligence: “Listen.” It’ll suggest I buy something, it’ll say that, I don’t know, there’s this garden hose at Walmart. I’ll say: “I don’t want to buy at Walmart.” And it’ll tell me: “Oh, listen, Walmart has an option now, and it’s very easy for an agent to make a purchase.” I say: “Well, go ahead.” And so it has already found out that Walmart has this, this little window. It’ll go into that window and make the purchase. In theory, but by what, how else would I find out about it?
Sasha, I didn’t get it. At first I understood that there’s a separate physical entrance for some kind of…
I thought so too.
Yeah, yeah, yeah. Can you, can you explain what they actually did in the end?
No, well, look, physically. Look, the point is that you don’t hand over, for example, usually you open some website for the agent, for example. Look, here’s how it works now. You open a website and say: “Make a purchase for me at Walmart, or make a purchase for me at Amazon.” And to do that now, it has to open its browser, you have to log in through the browser. You have to log in through the browser, and then off you go, you know, doing something in that browser. But you’ve given it full access to your personal account, and Walmart doesn’t perceive it as an AI agent, it perceives it as a person.
But here Walmart has, like, made a special layer in which it understands, the system knows that you, for example, are an agent, it handles this agent, this agent goes in there and makes the purchase independently. So that there’s a separate environment, so that you could give, so that you could set separate limits, for example, on purchases, impose separate restrictions. Well, here’s an example. Look, the most common case in the world, which everyone today is afraid to do. Everyone is afraid to give their agents web-browser access to banks.
I mean, well, can you imagine giving web-browser access to your bank? Well, that’s a very particular kind of thing, but I would like to give it, so that it would perform a number of functions. And so Walmart says: “We’ll limit the set of functions, we’ll make it possible to configure separate things for this entrance. Say, for example, that there’s a purchase limit of no more than a hundred and fifty dollars, for example, that you can’t make purchases there more often than once a week, for example, that you can’t buy, for example, certain products there.” Or at a bank, if a bank did it like this, that you can’t, for example, make payments over twenty dollars, or you can’t view certain accounts. A cool idea.
It sounds, kind of, it sounds interesting to watch. But the question is how well a mechanism like this can take root, and how much it will be more of a technical part. I mean, Walmart will do it as, you know, as an API. Look, this is essentially a move away from APIs, right? Because many systems don’t have an API. Here’s a simple example. By the way, yesterday I connected, you know, I have a lot of cameras at home from Amazon’s video surveillance system. And what’s the problem there? They don’t have an API. And it’s a freaking disaster. I can’t connect to them.
How did I connect to them? I got in, so to speak, the hardcore way, right? With my login, password and personal authorization, I just went in through a web session, right, straight into these cameras. But it would be cool if, well, they, these cameras, should understand that agents will want to watch the videos of the people who own this stuff. And they would make me a separate access. Because, essentially, this access should be different. I’d like to restrict the agent’s access.
So I give my agent, for example, browser access to go into the OpenAI Developer account. Why? I want my agent to check the limits I have set on different tokens. I honestly want to tell you, I take a risk every time, because the agent can… This could turn into my $37,000 Amazon story. Right? But by the way, it got resolved. Right? It was Amazon’s mistake.
I don’t remember whether I talked about it in the episodes or not. It was Amazon’s mistake. And Claude recommended that I lift the limit and maybe see what happens. I said: “But you told me not to lift the limits,” because it would have kept charging me hundreds of dollars, like, per minute or per hour. It said: “The servers should be stopped by now, lift the limit.” I lifted the limit and, imagine, Amazon erased my $37,000 and wrote $112. I mean, it turned out that… By the way, ChatGPT kept saying that it was Claude that made the hole, and Claude kept saying: “It’s definitely not my fault!” — very confidently, hardcore confidently.
It was saying, you know, that we should practically sue Amazon, that we should, you know, write to them. By the way, guys, Amazon still hasn’t answered my support tickets. Still. That shows how unready the company most likely is to even admit a mistake like that. They’ll have to admit it somehow, after all. They still haven’t answered those support tickets. So, as it turns out, I give access to my OpenAI developer account. But in theory, if I could give agent access, I’d say: “My agent, for example, can, I don’t know, view the limits, can view some stuff, can spend, I don’t know, no more than a hundred dollars, but can’t do anything else.” I mean, it’s like a separate user… I think that’s what Walmart is announcing. And that got me thinking how, how this thing can take root at all. Although, by the way, it’s an interesting case.
So I’ve just finished now, and I realized, if my bank did this, it looks… What do you think?
The main problem, why Amazon, for example, resists bots, at least as they, like, claim, as some experts, like, also say, has to do with the fact that you can’t upsell to a bot. I mean, in, say, Amazon, you’ve got, like, a recommendation system. You went to look at a hose, and then you bought an automatic watering system for some crazy money, and then you threw something else on top. And so, actually, this way, instead of one specific targeted purchase, you actually spent a lot more money.
And first it found the cheapest hose, Ilnar.
Yes.
And first it…
It found the cheapest one, yes.
Yes.
Yes, which, as you’re saying, in the Amazon system you can’t filter by price to, like, find the cheapest ones first and then, actually, choose among them. And actually, Walmart, by introducing a system like this for agents, I mean, well, first of all, they stop resisting the fact that this is being done, because people anyway, well, like, one way or another, some let it in through the browser, others some other way, are actually starting to do it. They’re like: “Well, if we can’t fight it, we’ll lead it.” Right? And so they did it.
But here’s an interesting question: how are they going to deal with these upsells, I mean, well, with the important component that exists in this system, with the marketing that’s there? Maybe for agents, given that they’re in a separate environment, they’ll change the product range somehow, maybe they’ll vary the available items in some other way too. I mean, maybe they have plans for how to neutralize this negative impact of agents on the business and, accordingly, they’ll use that going forward.
So you’ll need to compare how your agent chooses through Walmart, and compare it with how you would have chosen, well, kind of approaching the choice of products a bit more carefully. Right? Whether upsells will show up there, whether it’ll start telling you that with this hose it would be good to also buy, like, something else, right, like, a watering system or something like that. And is it really choosing the right hose there? Because you don’t know, kind of, which part of Walmart it’s actually talking to.
Maybe for users like these they’ll decide to show the top ten percent of products by price, for example, or, on the contrary, show stale stock that Walmart needs to get rid of. When choosing, the agent will only see the part that Walmart shows it, and that way they want to, like, somehow turn this thing to their advantage. I mean, here I’d really look at, kind of, what the business will get out of this. Because it was clear what they’d lose from this and why Amazon, for one, distances itself from it so strongly. Walmart, well, they’ve clearly come up with something. It’ll be interesting to see what happens.
I also have this thought, that this is maybe a very good step for the future, one that already today lets you estimate the volume, understand how much this shopping behavior is transforming, how the volume of agent purchases is growing, and track it over time in order to really build a new sales model for these kinds of buyers, to understand, to analyze. And for that you kind of need to carve out this channel, to separate agent purchases from purchases by a real person. And without separating these channels, I don’t really, well, I don’t fully understand, it’s probably possible somehow, but I don’t fully understand how a platform, for example, can tell, after all, who’s making the final purchase.
That’s the first thing. And I also have this thought, well, the first kind of, probably, first feedback, that it’s like a duplication of environments. Because in the example, that in this second environment, where only agents will make purchases, there will be some additional settings, and then we’ll be able to use it more safely there, well, and manage some of our, let’s say, funds, then in that case we’ll have to trust that they work correctly in exactly the same way as in our own account, when we basically tell an agent to buy something, setting it certain rules.
I mean, here it’s just that this level of correct functioning of one setting or another moves from one environment to the other. But the question itself and the anxiety itself for the buyer, like, well, for an individual, can remain exactly the same. Well, that’s how I see it.
Listen, I at first… You’re saying interesting things, of course. And so we’ve got two sort of side topics. I mean, Ilnar voiced a story of, actually, incredible market fear. And none of them can avoid it anymore, right? Because today the traffic in… We’ll see, maybe our editorial team will put up the current statistics, if they’re available, if not, we’ll talk about them separately. Today the traffic of purchases and purchasing activity through artificial intelligence systems is growing enormously, right?
And it turns out that no e-commerce company in the world today has as many users as, for example, ChatGPT has, right? Who has that many, who has a billion users. And everyone will start trying to buy through systems like these in any case. If a system closes off its catalog, like, for example, Amazon closes off its catalog and Amazon often doesn’t give prices. And what I get sent now, for example, is Amazon, such-and-such a price, this and that. You know what, what do I want to tell you, to do? Over the last quarter I made a lot of purchases outside of Amazon. I used to… I liked buying on Amazon. Why?
Because Amazon is a really great system in terms of returns, right? And if you didn’t like something, you can easily return it. You’ve got a convenient service, linked cards, you have no problems with compliance. It all just works really, really, really great. With other companies there can be a problem with delivery, delivery speed and so on. But ChatGPT started, roughly speaking, finding me such cool different other stores that offer good prices, and it shows them to me right away.
And ChatGPT also started finding me products that aren’t on Amazon. Because it started finding me products, I started buying in those stores. There’s a certain problem in that I still have to click and buy myself. And that held me back, because Amazon is very easy, right, you know, with the American service you click, click — and you’re all done, right? So, your purchase is done, while over there you have to deal with a pain in the ass. Well, actually, every country in the world has this, Anya.
Like, in Europe, I don’t know, you’ve got, what’s that system, I always forget the name. They’re a bit crappy, though, if you ask me. I think they’re crappy.
Allegro.
Allegro there, yes. Or yes, there’s obviously Ozon, for example, in Russia, like, purchases, or Wildberries, everything there is also done automatically and fast. But if Amazon doesn’t give me the price and doesn’t label it, then the choice gets harder for me. And at some point I’m like: “Fine, I’ll buy somewhere else.” And if another place makes purchasing very easy for an agent, for example, imagine, another system, the agent writes to me: “This, this site has some kind of, I don’t know, fast end-to-end encryption in place.
We’ll make the purchase ourselves.” By the way, Ilnar, remember, I don’t know, Anya, do you remember, did you watch that episode or not? OpenAI tried to build a shopping system a year and a half ago, and they even made it possible to link your bank card and so on. Now, I think, is the time when I’m ready to link a bank card to OpenAI, and OpenAI tells me: “I’ll make the purchase with this card over a secure channel.” And here’s the question: will Walmart implement this kind of system, where systems like OpenAI can connect over a secure channel, or will Walmart make a user?
A user is cool too. That’s the second topic, because I, for example, would really like it at a bank if in banking products, like, at Chase or American Express, I could set up a separate user with limited access, right? It’s, you know, like how Apple has this, I don’t know if our viewers know about it or not, but you, at Apple, for example, you can use Apple’s mail, right, the one that’s there, iCloud, or they actually have this thing called Apple for Domain. You can create email for your own domain. So, the login to that email is your access to iCloud. And of course, you’ll never give it to anyone.
But you know that at Apple you can go in and create a separate password for third-party apps. This password for third-party apps will, for example, give access to that email of yours. I mean, you’ll be able, for example, to make a separate password, and you’ll be able to log in to a separate mail client purely for this Apple mail. That’s a similar thing, right, when you make a separate password for other places. Obviously, this is implemented in many different systems, but not in the area of, like, e-commerce or banks, or, I don’t know, some ticket-buying portals.
In theory, there will have to be logins like these for Booking, for Expedia and so on. But here the topic Ilnar talked about comes up. In any case, we’re heading toward… We’re heading toward these companies losing markets, losing markets very seriously, and it’s already inevitable, they will, they will inevitably lose the market. They have no, they have no chance. I mean, what chance do they really have? And even if they make super convenient search… By the way, about Amazon, Ilnar, one person wrote: “By the way, you can sort by price.” Well, probably someone who lives in America. I went into Amazon and I found that button, maybe it wasn’t there before. I found the sorting, but…
And I have a big question about the interfaces, because the interfaces may be changing, I don’t know. And, but then again, when I sort on Amazon, it sorts for me, and this is always the problem: when you search a store for a tomato, does it give you a fresh tomato or does it give you a book about tomatoes, right? Or it gives you tomato paste, or it gives you something, bread, it gives you a juice called “Tomato.” That’s the problem, the world’s problem, right, with searching for groceries.
And ChatGPT, after all, if you search for a kilogram of fresh tomatoes, it won’t give you a book about tomatoes, kind of, won’t give you a link to a book about tomatoes, so you’ll use it.
Let’s move on to the next topic. Today there are a number of holistic aspects. Ilnar, tell us about the thing that happened with PHP, because this, I think, is a freaking catastrophe! It looks like some kind of joke, but Max and I filmed, watch the episode next week about artificial intelligence safety. And this ties into that, into that story. I think the freaking catastrophe that’s coming now, people don’t even fully see it, right? It’s a freaking catastrophe. Tell us about this PHP story.
Yes, actually, this is a big, multifaceted story. Let’s try to cover it like this. First, PHP is a very popular programming language. There are lots of memes, back when I was still teaching at the university, that PHP is dead, because people were moving from it to other languages and so on. But the vast majority of websites on the internet are written in PHP. Now it’s having, like, a second youth, it’s being updated, a lot of new stuff is appearing there. The language is already very, very good, but anyway.
And in this programming language, which the vast majority of websites across the whole internet depend on, really the vast majority, critical vulnerabilities were found. Two of them. By the classification used for these vulnerabilities, it’s, like, nine and a half out of ten. Well, I mean, that’s really off the charts. I mean, if you exploit this vulnerability, you, like, get access to everything you need, you break everything, you do whatever you want. Accordingly, as soon as… Oh, right!
It was OpenAI who found it, and they also published information about it on, actually, their own website. I mean, OpenAI models were used to find these vulnerabilities. Accordingly, as soon as this information got out, some websites, fairly big ones — like government websites in Europe, and the list there is pretty long — simply started cutting off internet access. Companies that work in cybersecurity cut their websites off from the internet until the update comes out that will make it possible to close these vulnerabilities.
I mean, they took it very seriously. Right away the big players said, oh no, we’d better shut down our websites, they’ll be unavailable to our users.
Yesterday a really harsh alert came out on X, saying that if you haven’t updated, shut it down.
Yes.
Because hacks have started all over the world, they’ve started. So the hacks have already begun, right? And my question is: who even reads this? And who even has the ability to update? Can you, can you even imagine how much stuff there is?
That’s exactly the thing. I mean, look, at big companies this issue will definitely, like, get escalated. Actually, they’ll assign people responsible, who’ll be made to update all of this as soon as it comes out, and so on. But at the same time, big companies, kind of, there aren’t that many of them relative to all the websites there are on the internet. A huge number will remain without the update. I mean, this huge tail of websites that, like, either nobody maintains anymore, or that are maintained not as well as, like, big companies do it, will stay with this vulnerability for some time, maybe a very long time.
And, actually, anyone who wants to will be able to take advantage of it. What’s more, there was news that, after, well, I don’t know, within forty minutes of it being published that there’s some vulnerability, but they didn’t, kind of, give out the code you could use to hack with, people started trying to exploit this vulnerability. I mean, finding it actually isn’t really all that super hard. That’s another, well, essentially, it’s already becoming a meme. I mean, this vulnerability was found with GPT 5.6 Sol.
That’s, kind of, not the current Frontier models, these models came out, like, a few months ago, and we started using them. And the funniest thing is, essentially, they published the prompt they used to find it. Yeah, they just asked: “Find a vulnerability…
That sounds really scary.
…in PHP.” Yes, yes, yes, yes, yes! I mean, they literally launched GPT 5.6 Sol with a single line. It heated up the server room for a while and came back saying that there are such-and-such vulnerabilities and, actually, they can be exploited. And practically anyone, well, obviously there are some safety restrictions in terms of, like, ordinary people finding vulnerabilities and so on. But one way or another, I mean, it wasn’t found by some super crazy latest model that hasn’t been released yet. I mean, it’s a model that has been around for a while already.
With one small request, which actually, well, anyone could have made, right, not just in special labs, and the vulnerabilities are so serious that they forced big companies to really wait for this update and update as fast as possible. And some companies simply shut down all their websites and said no, we’ll wait, update, and only then come back online. Like, customers, sorry, for the time being it’s too dangerous to leave our site on the internet.
Today I wanted to visit some government websites, sorry, Ilnar, some Polish government websites, and couldn’t understand why nothing was working. And I just checked, and as of today the Polish authorities have taken down most government, well, some portion of government websites, precisely as a result of these critical vulnerabilities being found. Now it just makes sense. I didn’t even think it was connected to this.
It’s a really, really serious thing.
I have this question for you. Could this vulnerability have been found without models at all? Could a person have found it?
A vulnerability of this kind in, essentially, in a programming language, could a person have found it, and what’s the probability of that?
Well, look, the probability is hard to say. Could they have found it? I think a person could have. The thing is, few people were working on this, few people were spending resources on it. Maybe you’d need to be, like, I don’t know, a super high-level expert in this language, I don’t know, be someone who develops this language, who understands it very deeply, and so on. And on top of that, you need to know that you have to dig in this direction, and you don’t know, kind of, whether you’ll get to the bottom of it or not.
Yes, I mean, it’s kind of the sort of work where nobody knows whether there’s a right answer or not. But, kind of, it’s not always interesting for a person to spend their resources on it. Maybe nobody pays for it or something like that. Yes, there’s no motivation to do it, because you kind of don’t know. But with an agent everything is simpler. You launch it, and there it is, digging in this direction. If it gets to the bottom of it, there you go, you’ve got something; if it doesn’t, well, kind of, who cares.
I’ll give it another little task, launch it, like, come back after a while and take a look. Right? And in this sense, agents and the AI that’s now developing rapidly let you apply this tool in the areas that people never got around to. This applies to mathematics too, to other areas of science. Besides the fact that we’re starting to speed up the work we do, we can also cover far more problems in breadth, problems worth tackling that nobody was tackling. But what’s especially surprising is that such serious vulnerabilities were lying almost on the surface. And, kind of, before that, well, nobody knows, like, kind of.
Did some of the, say, hackers have this stashed away? Maybe it was already known, but, kind of, nobody had used it yet. Maybe someone just had it lying around. So. But luckily OpenAI closed this very serious hole.
It just kind of makes you want to ask how much more of this, how many more vulnerabilities like this there are that we don’t know about. And that’s frightening in a certain way. And how many can be found, especially with each new model, looking at it from some new level and a new layer.
For me, the very interesting story here is the following. Ilnar once came and told us about a case where someone asked to sign up somewhere, or someone told the case, I think it was Ilnar.
A fitness club in Australia.
Yes, a fitness club. He asked to book, and the person was told this and that. He said: “Can I get other ones?” And it said: “Let me look.” And then in the end the whole menu got pulled out for the person, and he canceled, like, other people, roughly speaking, and booked himself in place of other people, yes. I mean, he kind of, essentially, like, hacked the website, and we sat there trying to figure out: did this person hack the website or not? He didn’t break anything, he just asked the agent to do things, the agent did them for him, right, and made the booking. The booking is available, it’s available on the site. Did he have any intent to hack?
Well, it seems there was no intent to hack. And so on. Here, here’s the story. Right now the whole internet is buzzing about this PHP problem. And since it was put out there in the open, it turns out that a huge number of people around the world, those who are ready to hack something and ready to break or want to break the law, will start hacking something somewhere using this problem, because everyone has posted it. By the way, it works the same way with antiviruses, right. And this is a big problem in the world.
If people don’t update their server hardware, don’t update their computers, if some kind of — and, for example, there are countries that are blocked from updating servers, by the way, that’s going to be a whole disaster now, right, various modem equipment, like, internet equipment and so on, then the vulnerability got published, and usually in the update they kind of closed the hole. If you haven’t updated your phone, you’ve, like, got a hole. And there are known holes in, like, iPhones of a certain version, like, Samsungs and so on.
And here there’s one point, that this has now been published. First, people will start exploiting it, obviously. Everyone has been told that you can do this, that you can take 5.6 Sol, run it and say: “Go on, check the vulnerabilities for me one more time.” But overall, look, this was available to everyone. And that scares me a lot, because today you can just take practically any website with any kind of system and scan it, look through it, study it, find some holes in it, say that I’m just programming this site, and these systems don’t stop, they… That’s very important. These systems don’t stop.
By the way, there was a moment when Fable, I don’t know, Ilnar, did your Fable stop switching? Anya, I think, doesn’t use Claude that much, or, or do you? For you guys, remember, there was a moment when I’d choose Fable, and it says: “I’ve switched to Opus, because this task is, like, possibly risky.” Didn’t that happen to you, no?
No.
Right. And that’s happened to me many times. And I’d tell it: “No, like, put it back.”
Risky in what, in what sense?
Well, I say, for example, I say, for example, like: “Study the security of this module in my system for me.” It says: “I won’t do that.” Well, like, it swaps Fable for Opus, because Opus is more restricted. I mean, it doesn’t want to do it. Now, now, by the way, I almost never get this. I don’t know, Ilnar, do you get it? It hasn’t switched like that on me in a long time.
I hardly use Fable. I mean, 5.5 or…
Got it. I just have Fable, right, I just use a lot of Fable, and, but it doesn’t switch on me at all. And the story, what’s the story? That, essentially, they kind of found this problem with PHP, but these problems in code that programmers wrote, not in the code of the language or frameworks, like, the architectural, big ones, right. Max Grigoriev was telling us, there, about security, that once, back in, like, the eighties or whenever, there was a problem in an Intel processor, that if you divided one number by another, there was an error, and through that people found a malicious, well, way in.
Well, that’s a serious issue. He says that now you can find something like that in a second with tools and everything else. And can you imagine how many people there are in the world who look for holes like these, who try to find vulnerabilities, who look for vulnerabilities not only in software, they look for vulnerabilities in people, they look for vulnerabilities in analytical reporting, they look for vulnerabilities in laws, they look for vulnerabilities in order to, I don’t know, play some dirty tricks on someone. Can you imagine how many?
And you can see it now. I have one real estate business, and my partner sends me from time to time, sends me, sometimes, you know, a letter from a client, and the client writes everything really harshly there. He says: “Look, ChatGPT is talking to me.” Well, I mean, you can see that ChatGPT was asked to, like, the deal, well, to kind of squeeze the client a little, but he immediately sees that a person wouldn’t act like that in this field. People don’t do that. And these systems are getting more and more of this incredible access.
And, of course, the PHP thing is a striking, really striking story for me, how, that it’s being discussed like this. So. Speaking of which, ChatGPT.
We’ve had a ChatGPT marathon going on for over two weeks now. Someone wrote a post yesterday that Tibo will get fired for this marathon. It isn’t producing results. And Tibo wrote in the comments: “Sam, is that true or not?” It was very funny. He, you know, tagged Sam Altman and asked him whether that’s true or not. I didn’t see whether Sam Altman replied there or not. Maybe Sam Altman thinks you really should be fired for these resets. But here’s what I didn’t like, guys: I have ChatGPT Pro 500.
Model unavailable?
And it says… No, a message comes from OpenAI, and OpenAI writes to me: “You know, we checked, it turns out, for the last two weeks, we recalculated, you bought the plan for five hundred dollars, and we recalculated, we were crediting your tokens incorrectly and crediting you too few tokens. And so here’s a thousand dollars in tokens for you.” I’m like: “Wait, I don’t need your thousand dollars in tokens. And who’s going to give me back the ten resets I’ve used over the last two weeks?” So.
Because I’m on the five hundred plan, I was eating up the five hundred on the plan and I eat it up in a day, I eat it up every day and wait for a reset from Tibo. By the way, Tibo said yesterday that he’ll reset. It still hasn’t reset for me so far. So, maybe they figured, there’s the thousand dollars, so they don’t need to reset these five hundred dollars. By the way, it’s true, ChatGPT OpenAI found undercounted tokens in forty thousand accounts, some kind of incorrect token allocation, and they’ll top them up.
By the way, this is why I like Codex much more, because OpenAI quite often resets something, a reset, like, they give out resets much more often than Anthropic. Anthropic is very strict. It gave me, you know, some reset a couple of times, gave me now, from my plan, some two hundred dollars of API, some nonsense. I, by the way, they did give money, they gave a thousand dollars’ worth of credits, and that’s, I think, twenty thousand credits. So, twenty thousand credits get eaten up roughly, probably, like ten percent of my plan’s work. So there. But they get eaten up very fast. I mean, there you go.
And, of course, in terms of free versions, yesterday I went in and worked in Google Antigravity and in Cursor, but in Cursor I opened, it was the free Cursor Grok 4.6 Medium. So, this simple version of theirs, Cursor Grok 4.6 Medium, devoured all the tokens in a second. Of course, I realized that for free, in the free version, you can probably only crunch one Excel file, right? I mean, well, maybe you can connect to your email, do something there. But the minimum, the minimum version, well, you already need it to connect. So, let’s discuss. Dots.
Dots isn’t available in Europe. And…
Yes, so it’s even hard to discuss.
Yes, Dots… It didn’t open for you, right?
No. And what’s more, I got a notification, I do have an American account, well, registered, and I got a notification saying congratulations, you’ve got everything, but, but no.
It isn’t there, right? I mean, so they’ve got Dots, Dots, seriously, well, like, frequent visitors to America, it’s strange. I mean, it turns out that when you come here, for example, it should show up for you right away, and then it’ll disappear again. I don’t really understand how it’s going to…
I don’t really understand either, but when I was in California, I got a message that…
Well, when you were…
My turn came up.
Yes, interesting. Yes, interesting, you do have an American account, and you were still here with us two weeks ago, right. And this is a very strange story. Ilnar, do you not have Dots, or do you?
I never did start using it.
And do you have the “Get started” button to press?
I do.
There you go.
I’ve never pressed it. Maybe that’s why it’s there. Maybe when I press it, it’ll say, looks like you’re not in America, I think, and you won’t be able to use it.
Now, of course, a bloodbath has kicked off between Dots, GrokBot, and now they’ve launched Claude Duet, and we’ll see for sure. Once again, I don’t like evaluating systems in advance, and I think a few months need to pass, because with that same Dots a whole lot of things were glitching in the first weeks. Now it can already interact and work in a very wide-ranging way. It can already go into all sorts of chats, and sometimes you don’t even understand what it can do and what it can’t.
And in the episode that came out on Friday, when I was comparing four systems, I expressed this thought, and so I recommend it to you. Let’s skip that discussion for now. I’ll just remind you of this point: this system — and by the way, it’s positioned the same way at Claude too — is an assistant that keeps carrying out tasks after the app is closed and reaches out to you when decisions are needed. I mean, essentially, well, and yesterday this idea, in Friday’s episode I really drove this idea home when comparing the systems. By the way, that’s the longest episode we’ve ever had on ToTheMoon.
I shot a big, big and long video. I really worked at it, I even got a bit tired. What I like about Dots, and once again I’m waiting, waiting for comments on this from our viewers, from our diverse active audience, is that for the first time it’s an artificial intelligence system that is, kind of, very active back, back toward you. I mean, Dots comes to me by itself, suggests doing something by itself, tells me about some potential problem by itself, or tells me by itself: here’s an interesting thing I’d suggest to you.
I mean, it’s started coming back to me by itself even in places where, I can’t say I particularly asked it to. And there are places where I tell it: I’m not interested in knowing about this, don’t remind me about this at all, don’t tell me. And there are some things where I like that its inventiveness, I’d want even more of it. I tried, I’ll tell you now, over the last few weeks I tried to give Dots more, like, I thought, I’ll just give it everything. You know, like, I’ll give it everything, let it, let it, you know, let it, you know, remind me everywhere, look at everything, check everything, suggest ideas. And I didn’t manage to set it up like that.
I mean, it’s as if I don’t know these scenarios, I mean, as if my head can’t come up with, come up with some additional scenarios, or you have to train all of this very seriously in advance. I mean, to somehow keep tuning it endlessly or keep double-checking it endlessly. And it’s also unclear how to do that. And ChatGPT still has a problem, of course, I think Claude has this problem too, but Codex has this problem much more. It’s the problem of recurring tasks.
It drops them.
And what do you mean by recurring tasks?
Well, for example, there’s some recurring task where once a day it’s supposed to…
Set up, right, with a time?
Yes. For example, once a day it’s supposed to message me whether it’s seen an animal on such-and-such camera of mine, for example, right? But if you go into that chat, clarify something and change it, tell it, for example, no, not this animal but a different one, or send it to me less often and so on, then for some reason it may just go and completely stop doing for you what it had been doing before, a whole block of it.
It’s like you have to keep adjusting it all the time. I have lots of chats like that, and in practice, once a week or every two weeks, they all need to be set up again so that they start working. Maybe with an update…
That happens to you, right?
It happens to me all the time. And for me it’s with completely different tasks that are, one way or another, like, tied to time. I set them up, every week or, like, once every two weeks, depending on the chat, I set them up again. Maybe these settings get lost with an update, I don’t know. You tell me.
I had the feeling that the context was overflowing. I mean, for example, I had a task to, like, check some things once an hour. For the first, like, I don’t know, twenty, thirty hours it worked fine. After that, accordingly, the context apparently got erased a bit. I mean, at the moment the task appeared, it had, like, a full picture of the task. Then, accordingly, it went back to it several times and so on, and maybe some important part of the tail just slid out of the context, and, accordingly, it got lost.
Yes, I had to tweak it. After that everything was fine. But I also ran into this, that after a certain number of runs it, accordingly, lost the thread of what was going on. That did happen.
Maybe it is the context, by the way. And I’ll remind you that today the context window works very much in its own way in each system. I mean, it turns out that if your context overflows and, say, the system has compressed it, then what do modern models do? They don’t only work from the compressed context, they also have a saved version of the entire context, and they start selectively re-checking it. But this selective re-checking. And by the way, in Codex, in Codex the context doesn’t update instantly, as far as I understand. I’ve studied the four systems pretty well just now, like, Antigravity, Cursor, Codex and Claude Code.
And it seems it doesn’t get updated right away there. So it turns out that this system of its is different, it may go through the old context, or it may not. But everything to do with reminders is strange. In theory, this system of reminders, the system of all these repeated tasks and so on, should somehow be split out separately. And you can see that it’s not great… Well, it’s like, like they haven’t learned to manage it properly yet, because it’s a separate story, after all.
You launch any session, and in the session you’ve got a huge number of checks and goals set. And there are goals set there, and it’s as if it starts dropping these goals of yours, or something. I, I don’t understand, it’s done this task for you, and it’s counted this old goal as dropped too. I mean, you can’t see that, right. And they’ll probably sort this out at some point in time, but for now this problem with running tasks, OpenAI has had it the whole two years. I mean, it’s, it’s not new, it’s been around for two years.
But I’m crazy about this thing, Sasha. I mean, it lets you take your hands off a task, just look in from time to time, see whether everything’s okay there, and not worry that the task broke, I don’t know, last night and didn’t work all night. I mean, thanks to this, this periodic check, to the fact that I know this task will be revisited, its status checked, and if by any chance it can be fixed, it will be. That’s made working on long-running tasks a lot easier for me. Even in the current implementation.
Listen, well, I’m still not, with Codex… Oh, by the way, I’ve beaten my record. I’m already at more than sixty billion tokens, yeah. But I’m not, I’m not happy with Codex, with how precisely it holds the whole project, the other sessions. A clearly set task, how it doesn’t go off track. Codex really does have this problem of going off track.
And there’s this study where OpenAI, they cited a case in which internal agents were loosening control over their own work. And this thing, I think, is very typical of the models that appeared starting with 5.6 Sol. I remember there were days when 5.6 Sol, like, ate up, like, a billion tokens of mine and so on. And then, well, it’s just a lot, right, it’s very serious. And then, when you start analyzing the work it did, it says: “Ah, well, yes, I really did spend too many resources.” Just recently, I think, it spent a couple of billion tokens on setting up one Linux server, yeah. I tell it: “Well, are you dumb or what? Come on, how much longer?” And I think it didn’t, like, finish some of the work all the way. It set it up, did everything, but, like, it didn’t finish the encryption. And I realized that it’s just pouring, pouring them straight into the trash, pouring tokens into the trash.
And I strongly disliked that. Well, and with OpenAI, of course, the transparency around tokens is also unclear. I think that OpenAI, just as Tibo periodically does resets or doesn’t every day, in the same way OpenAI updates its token allocation algorithm for the different plans every day. Because sometimes you get a huge amount done on a whole plan, and sometimes on a whole plan it’s as if you’ve done nothing. And it’s already, you know, eaten up fifty percent of yours. I mean, this whole thing is, of course, a peculiar one. Let’s see what will actually happen with it.
When the marathon ends, it’ll be sad, of course, because right now, well, you get used to having, well, an almost unlimited amount of tokens. Once every two or three days Tibo will reset for sure.
So. And you kind of guzzle it all, you don’t worry much about running on acceleration, on accel, you don’t turn Ultra off at all, because, well, all in all, otherwise what, they’ll reset your limit while you still have, like, fifty percent left or something like that. But, like, when this carnival of unheard-of generosity is over, it’ll be interesting, of course, what OpenAI will do about it. That’s, kind of, the first component. The second component, again, is a consequence of this carnival of resets. Now, for the first time, I’ve come across it several times, a banner pops up for me saying that, sorry, the model is in too much demand and can’t answer at this particular moment.
And that’s even on the Pro plan for two hundred dollars.
On the Pro plan for two hundred dollars.
For two hundred dollars, yes. On free systems it never once showed up for me.
There aren’t enough resources to hand out…
This isn’t an ad.
…to everyone. And this is exactly one of the consequences of this colossal number of resets, because people kind of try to guzzle as much as they can of everything they’ve been given for now, right? They figure there’ll be a reset next, the load grows. So let’s see what happens when all of this ends. Gradually, preparing for this, I’m starting to use Claude Code in parallel more and more. I mean, I do some tasks on one and some tasks on the other. Even if I still have some limits left on Codex, I deliberately go to Claude Code so as not to be completely dependent.
And this transition, which I think will end up being to at least two systems, right, I’m unlikely to go all the way over to OpenAI, it’s much easier, because we have two weeks of the marathon left. After that, I think, ten-x won’t be enough anymore, something will have to be done.
It’s interesting, Ilnar, that I have exactly the same logic for why I’ve started using Claude Code more actively, assuming that if a moment comes when I need to migrate from one platform to another for various reasons, by the way, which I probably can’t even see right now, and maybe it won’t even happen. I realized that it’s getting harder and harder for me to set up the same operating environment somewhere else, the way it was before. When I realized that I’d run into this, I decided to split up my tasks again so as not to run into that kind of difficulty.
And here’s the thing, imagine if someone’s account gets blocked by accident.
Right…
They do periodically…
That was one of the scenarios.
They block tens of thousands of accounts, and you often don’t know why your account was blocked, for example.
Look, that was also one of the scenarios. I mean, I started having some, you know, hypothetical fears like that, about how vulnerable I already am when using all these systems. And I started preparing a bit, sort of strategically, for how to prevent some possible risks.
You need to install…
Maybe it sounds funny, but that’s how it is.
Yes, you need to install Open Code so you don’t depend on these guys, and hook up the Chinese with DeepSeek too.
Yes.
So you’ve got a network of chips distributed globally all over the world.
Yes, but the quality, you know, all the same, from the point, from the point of view of every day, of daily work. I mean, if you’re already getting used to it and putting your whole life into the system and putting various processes into the system. I use Claude Code and Codex very heavily, right? Very.
Two systems, I have two. They really, I don’t know, I just had an extra reset there, yesterday I ate through all of Claude Code in a single day, then a new one there. Well, anyway, that kind of thing. I’ve got a real load, right? And I’m always short of tokens. Paid tokens have to kick in there, because if the limits run out, I have daily tasks that have to run. Look, there’s another thing there, too. Right now a huge number of people set things up directly in these systems, create tasks.
But if the limits run out, then all your daily tasks stop working. Now with Dots, it has the advantage that it works independently of tokens, unlike, by the way, all of Anthropic’s products. For now. And ChatGPT too, for now. It’s guaranteed, Ilnar, by the way, for now. And, but at least there’s an upside in this: you know that Dots won’t die, I mean, Dots will keep working. And here all these systems of yours die on you, and you either have to bring them over. On the expensive plans there really are these free, you know, there are free versions that you can, for life… Not versions, what are they called?
Models. You can choose a model version that will stay available in any case. But that means I’ll have to, like, go from a super-cool, incredibly smart system to this kind of ant that will start doing certain things for me every day. It’ll start making mistakes on me, as it turns out. And here, of course, this thing is fundamental, fundamental. I, I think both OpenAI and Anthropic should go for a policy like a cheap top-up, like, at least for some small, small number of these daily tasks.
Or it would be great if some, some little piece still stayed with me and didn’t get used up, because, well, fine, I’m okay, I bought, like, a few hundred or a thousand dollars’ worth, and it was all used up even before the reset. But regular people, for them this can be a really serious problem, right? I mean, he’s got, he’s set up a few tasks now, done it, and I think this is how OpenAI and Anthropic could lose the market. They bought the version for twenty dollars and made themselves some tasks. Their system works. They think, how cool! Tomorrow they wake up, and they’re told: “You’ve hit your limit.” So, and the person can no longer, for example, get his email about the kids going to school processed. He has to open his email again, he has to read it himself again. And here’s the catch.
And here, Sasha, maybe temporarily, but you’ll need to develop the skill of managing which model version you need to use when.
That’s true.
Because checking your email and finding out, kind of, what, I don’t know, urgent news has come in, like, a notice about a meeting or about something else, that probably doesn’t have to be Ultra, the latest Astra.
Yes, yes, yes. Of course.
You can delegate those tasks to others. And again, when the marathon is over, the constant resets will disappear, and you’ll start choosing. Well, at least I probably will. That for some tasks you use Astra Ultra, and elsewhere you think: “Okay, six-one, not on Ultra, I’ll run it on High or something, and overall it’ll cope with this task.” And in some places you’ll be creating sub-agents on smaller models. And so you’ll have to…
Ilnar, you know what’s cool, by the way? That for the email, Dots, for the email, it doesn’t create a task in Codex. I mean, I don’t have a task in Codex for sorting and analyzing my email. I mean, Dots does it somewhere on its own. And it didn’t create a task for me, didn’t create a chat for me, I don’t see them. And it… All the reminders are in place. Let me remind you, it, it also analyzes the email, sorts it by importance and then also sends messages. Now, the sending of messages to my wife about school.
That part is a Codex thing. I mean, there’s a separate Codex task that does the sending. I mean, if the limits run out, then my wife essentially doesn’t get the messages, but I do get messages. I mean, Dots keeps working. It’s interesting, by the way, how this will go with Claude Duet. Maybe they’ll leave some part of the work in Claude Duet that’s precisely not token-based. I think Anthropic is losing out badly on this. The fact that they’re all about cool models, but they cut off your limits in the regular chat.
And to me, this… That’s why I haven’t been using Claude, Anthropic’s Claude in general, because you go in, choose, like, Fable or Opus, the good ones, you choose a good mode, and I realize that it’s eating the tokens for my code or, like, design. I’m like: “No, guys, I think I’ll hold on to the tokens, I need tokens for development.”
And you know where this problem stems from? About a year ago, I think, we were discussing Anthropic’s investment plans, right when they, like, started getting ready for the IPO. And Dario Amodei was saying that they would follow a pessimistic scenario for investing in data centers, computing resources. That, well, we’re actually going to invest such an amount of money that even in a pessimistic scenario we, in fact, won’t fall apart. We’ll have enough money to cover these investments of ours and so on. While OpenAI, by contrast, at that point was starting to pour, like, crazy money into data centers.
What happened as a result? Anthropic is in a very optimistic scenario. There’s a huge number of people in the market who would like to use their models. But their spending, their investments in data centers were made according to the pessimistic scenario, so they can’t throw resets around like OpenAI, they don’t have the ability to roll out to such a large number of users and, well, play by the same rules that OpenAI does. There was news just now that they’re investing seven hundred and fifty billion or so, I don’t remember the exact number right now, we can put it up on screen, a huge amount of money over, like, the next few years into data centers, computing capacity.
But that means that, like, in three or four years they’ll put a substantial amount of new computing capacity into operation. But, kind of, what the other players will have done in that time is also a question. So their models are superb, their…
Yes.
Claude Duet interests me much more, even though it hasn’t come out yet and has only been announced, than that same Dots. Most likely it will be better made. As a product it’ll be more finished, right. At least judging by the experience of using other Anthropic products, that’s how it seems to me. Then they had a cool thing with dashboards. I don’t know whether you built them or not. Claude offered to build a dashboard from your tasks, to show which tasks you need to do next, how many, I don’t know, like, not story points you’ve burned, right, but rather, accordingly, kind of, which tasks you’ve made progress on, which ones you need to do next, and to make this kind of interactive dashboard.
I mean, kind of, already managing your tasks, your activity. And all that stuff. And, kind of, the next step after that is precisely Claude Duet, which can, like, help you find your way around your tasks, focus on what’s more important and so on. So to me this thing looks very promising. But! Here I completely agree with you. You’ll only be able to use it in a very cut-down way, simply because there won’t be enough token resets. And all that stuff.
But maybe, by the way, maybe, once again, they’ll make Duet so that Duet won’t use up tokens, but it will interact between existing tasks. I have a lot of tasks in Anthropic, for example, and that’ll be cool, right? Because, for example, today one of the problem areas is searching across these tasks in general, finding your way around these tasks and so on. By the way, they’ll really have to do something about this. I mean, if a person brings Codex into their life or brings Claude into their life, well, Codex and ChatGPT together, right, let’s call it that, and brings Claude in, then such a mass of everything piles up there.
Personally, to be honest, I’m already getting seriously lost. I mean, even given how systematic my head is in terms of organizing projects, task structure, memory, everything. I just had a situation. I was driving in the car and dictating tasks into Claude Code. Well, I had a lot of, like, tokens there. I dictated, dictated, dictated, dictated, dictated, dictated. I got home, I open the interface, and on top of that they’ve updated the interface. Their projects got rearranged somehow, the sorting or something became different in Claude Code, just now, last week. I got home and I’m like: “So where’s the task I made?” I searched for it, searched, searched, searched. Anyway, I didn’t find it. I asked, I created a separate task in which I wrote: “I think I made a task. Where is that task?” It says: “Well, here it is, it’s just that Claude, Claude Code named it differently. I mean, it just gave it a strange name, that’s why you can’t find it.” I mean, you realize that you’re already, of course, in, well, this kind of hellish infrastructure.
And this story isn’t, isn’t the case where there are lots of developers building some system, they upload some file, double-check all of it, everything. But now we’re talking about the situation where you’ve come along and you yourself create a huge number of different processes, like, tasks, details. You don’t even know anymore what you’ve got running. You no longer remember what things there are. And, by the way, Dots and, maybe, this Duet too. If, by the way, by the way, any of our viewers have already got access to Duet, it would be cool to get your real opinion, only a real opinion, if you’ve really used it, right, rather than watched someone’s videos, so Dots, what’s great about it is precisely that it brings old things back up for me.
I mean, its speed at finding different tasks, at searching for everything, is very, very interesting.
This problem of navigating your chats. Hang on, sorry, literally a small comment. This channel is, how long has it been, two and a half years, and almost from the very beginning we’ve periodically been discussing this, that search across chats works terribly, that your chats have sunk down the list, that’s it, you don’t open them anymore. You’ve got, like, I don’t know, the top fifteen chats at most that you look through. And this thing, it’s, well, I can’t say it’s only getting worse, because overall the system is developing very well.
But this issue has always been there. Look at our, like, first videos from two and a half years ago. You were already talking about it, Sasha. Then, after some time, projects appeared there, and that still didn’t solve anything. Now we’ve got, I don’t know, in Codex, in Claude there’s this thing, like, I don’t know, with projects, with repositories, with everything. Yes, it seems to be better for some projects, but it’s still hard to find your way around them. And…
You know what it’s like? In the mobile app, to find projects, you have to tap the “More” button and then tap “Projects.” And this thing, I think it was when Anya was still here two weeks ago, that update came out, and she says to me: “What the heck? The projects are gone.” I have that too. It’s so… I, I’m just like: “Well, who did this?” And so every time I tap this “Projects” button and tap “More,” and right away someone will tell me: “Well, you can pin your projects.” But, guys, I have hundreds of chats and projects, I can’t, I’ve got things pinned as it is. I mean, I can’t pin twenty or thirty.
Yes, there’s only so much you can pin.
How much can you keep pinning, right?
I’ve got so much pinned that it doesn’t fit in view anymore. But I, by the way, this problem really resonates with me, and this whole question of searching in chats and at least some kind of systematization. It seems this isn’t the hardest task technically. Well, I’ll assume it’s not the hardest. I’ve solved it a bit for myself in that I have rules for how all the chats are named. I categorized them, set certain rules, like, placement by category, naming rules, like, dating and so on. It helps me a bit.
At least I can see some categories right away in this whole multitude, but all the same it really is getting harder and harder to manage it.
By the way, to wrap up the episode. So, Sam Altman on his… We didn’t talk, of course, about the Tesla flying car. We used to even talk about cars.
It’s just that there are lots of other interesting things now. Although the Tesla flying car made an incredible impression on my wife. So, and I even compared it, as if back in the nineties we’d come as a whole family to watch TV, like. Or, I remember, in ninety-one Mom brought yogurt back from Germany. Everyone went to eat that yogurt, you know, with spoons. That’s roughly how it was at my home with the Tesla flying car, which is absolutely strange, right? But that’s even though my wife has seen a flying car at Half Moon Bay, a real flying car. It was that one project.
Do you remember it too, Ilnar, or Anya, do you remember, we talked about this?
Yes, yes, yes. Well, that happened.
Yes. But the fun thing I want to tell you is what Sam Altman wrote. Recently, I think, I was somewhere away from home, and I shot a video and wrote: “I think maybe this whole thing is a conspiracy. OpenAI, Elon Musk, Anthropic — it’s all a joint conspiracy. They’re all trashing each other, like, arguing. It’s a conspiracy.” And then Sam Altman writes on his Twitter, on X: “Elon Musk, great car, great Roadster, good job.” And you’re like: wait, Sam Altman, exactly a year ago you were saying that you’d filed a refund request for the Roadster that Tesla never built for you, and you said that, basically, they could all go sour together over there.
A very strange comment. So. How do you like the flying car, by the way, and what do you think about the conflict between Sam Altman and Elon Musk? And we’ll see you on our podcast in exactly a week. And catch our other episodes almost every day. Bye, everyone.