334 | AI just got 50x cheaper 🤯, Buy any AI model from anyone (Coming soon), “The world should accept some bad things happening” (Sam Altman), and more important AI news for the week ending October 9, 2026

0:00 Hello and welcome to a weekend news episode of the Leveraging AI podcast, a podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Metis, your host, and this was another week in which a lot of small things happened, and it's another week that shows several very clear trends, and we're gonna focus on two of them, or two and a half of them.

0:23 in this episode, we are going to talk about how companies, leading companies are now letting you run other companies' models through their platform, and how critical an aspect that I've been predicting for a very long time now, is actually coming to fruition, where you can run multiple models through a single subscription or even on an agent platform and use whatever tools you want in the back end, or sometimes maybe not the tools you want, but other tools and not just the owner of the platform.

0:54 we're going to talk about how America is currently competing on the cheap end of the model scale and not just on the frontier, and that has been very, very clear this week that this is the direction. And again, it's not a new direction. It's just the first time that the focus of the US has been on that.

1:13 We are going to talk about AI safety again and what's actually happened in OpenAI, with a few people departing and a few people getting fired, and how does that impact, safety and what people are saying about it. We're going to talk about the top 100 GenAI consumer apps. It is the seventh edition of this report, and we're gonna dive into who's in the lead, what changes, what happens, and what does that actually mean.

1:36 And then we have a lot of new releases of models and of new features within models, And we're going to also talk about how AI currently makes novel discoveries in science in two different aspects of science and how that may impact our future. So lots to cover. Let's get started The first topic I want to talk about, that again is a trend and not really a topic, is the fact that it is becoming very clear to everyone that the competition is not something you can ignore, and the fact that you're OpenAI or Anthropic or SpaceX or Gemini does not mean that the people who are willing to pay you will pay just you, and they will go and look at other options as well, and the fact that you need to give them

2:19 options. When we started seeing lots of new routers this year and how successful it's becoming, I said multiple times that the large labs will have to follow suit. They will have to follow suit because they don't have an option. And this is now happening, and it is happening from multiple directions.

2:36 So this week, Elon Musk announced that his AI agent tool, GrokBot, which I really, really like and I'm really enjoying, using, will also run third-party models to give you the best bang for your buck The exact quote from his post on X said, "Important note regarding Grok@bot. Going forward, SpaceX will use the best back-end model for any given task, including Claude Opus 5.5, Midjourney, Suno, and other leading APIs, whatever is most likely to give you the best outcome."

3:08 So this is a post from Elon on October 7th, which is basically saying two things. A, he's saying that they will try to give you the best results and not necessarily to force their own model, and B, that they're integrating stuff beyond just the large language model. So the fact that he chose to mention Midjourney and Suno by name, which are more creative kind of apps, tells you the direction that they're going.

3:31 They want their model to be good at everything. They understand their model will never be good at everything better than all models for each task, and they're allowing other models in the back end to provide the solutions for that. Now, he also said that they're going to focus simple tasks on answers that are gonna come from the small model, or for the next variation of Grok.

3:53 But that does not change the fact that they're going to also route some of the needs to other models, either cheap or expensive or better or specific to specific things like Midjourney and Suno. Now, the other interesting in this statement is that the first model he chose to mention is Claude Opus 5.5.

4:13 If you remember, not too long ago, Elon said that Anthropic are evil, nothing short of that. But it's not gonna be the first time that Elon makes very aggressive statements about somebody and then changes their mind and decides to work with them. Now, the other interesting thing from Elon that relates to this is if you follow the money and how SpaceX is making most of its money right now.

4:33 So SpaceX AI unit, so just the AI part of it, did $2.56 billion in Q2. That's up 247% year over year. Most of it, 2.19, so almost 2.2 billion, came from selling AI infrastructure, basically selling compute. Only a small fraction of the revenue came from actually selling the usage of Grok. Musk himself said that only 10% of SpaceX compute may go to train Grok next variations.

5:04 So it tells you where SpaceX sees its future from a revenue perspective. It is not necessarily developing Grok as a leading model. It is developing their AI infrastructure so they can sell that. We spoke multiple times about, their wish to do that in space and how Elon thinks that is the future and how he's, by the way, currently trying to raise $40 billion to put more compute in space.

5:27 Now, there were a lot of comments about this because the way he's trying to raise the money is a term sheet of two pages, which is a lot less than people would expect when you're asking them to write a $40 billion check. But I guess the way this works is if you are willing to give Elon your money, you are willing to give Elon your money, and it doesn't matter what's gonna be written in those multiple pages or not.

5:47 And that's has always been the case. If you think about when he, bought, X, back then Twitter, how did he do it? And when he then, decided to launch, the xAI stuff and how he raised $7 billion as his seed round very easily just because he's Elon. So I assume he will also raise these $40 billion in order to put, compute in space, but time will tell how that plays out.

6:09 but going back to our main topic of this discussion is that Grokbot will allow you to use other models inside of Grokbot. Again, you won't be able to choose the models, but these models will be served as part of Grokbot in order to give you the best experience. Now, going not too far back on September 29th, OpenAI launched OpenAI Marketplace, and what that means is that eligible enterprise customers can apply part of their existing OpenAI commitment to other platforms with currently 32 approved partners.

6:44 These include Adobe, Figma, Salesforce, HubSpot, ServiceNow, Harvey, CrowdStrike, Sierra, and Decagon.

6:51Speaker This episode is brought to you by the Multi-Agent Orchestration Course. It's a course that I have been teaching personally since April of this year, and every single one of the cohorts that we have opened have sold out. And the people who have been through these courses made incredible transformations for themselves, switching jobs, starting their own companies and so on, and definitely for their businesses by being able to build really complex processes by following our methodology and our infrastructure settings that we teach during the course.

7:19Speaker The course is four sessions of two to two and a half hours each over four consecutive weeks, and through it, you can go from knowing a little bit about how to use AI and using it in ways that most people are using it today to learning how to build really complex, sophisticated systems that can automate more or less everything in your business.

7:36Speaker The current cohort that is open starts on November 2nd, and it's gonna go on November 2nd, November 8th, November 16th, and November 23rd, and we are almost sold out for that cohort as well. It is very likely that we will not open another cohort in December, meaning the next one after that is gonna be in January.

7:52Speaker So if you still wanna learn how to build really incredible automations and learn everything that I've learned by doing this every single day for the last few years, don't miss this opportunity to come and join us on November 2nd. There's a link in the show notes, And if you're going to use the promo code LEVERAGINGAI100, you're going to get $100 off of the price that is on the website.

8:13Speaker Like I said, there's a few seats open, but they are selling out fast. So if you're interested in doing this still this year, go and click on the link right now on your phones and complete the registration and the payment so you can guarantee your seat. I'm looking forward to seeing you there. And now back to the episode

8:29 Now, the other part of this is that OpenAI Enterprise customers can run open models like GLM 5.3 Flash and Kimi 3 and other Chinese models inside of Codex and charge their OpenAI commitment to that. So again, OpenAI understands that they cannot keep you hostage, and that you will go and use cheaper other models if you need to do so, and by allowing you to do it through their ecosystem, they're, A, making it easier, B, making them more competitive comparing to, let's say, Anthropic, if they decide not to do it, which I assume they will.

9:05 And it probably allows them to make some margin in between. So if the cost of GLM 5.3 is whatever, 15 cents per a million tokens, they're gonna charge you 17 cents and gonna make 2 cents on every one of these transactions without actually having any cost. So it makes perfect sense for OpenAI and again, all the other leading labs to go down that path.

9:25 Now we have GroqBot that's gonna choose the models for you. We already had ChatGPT that allows you to use your subscription dollars to use other models. Again, if you're an enterprise, that's not for everybody for now. and then we had another example of this, a more extreme example of this, very clear this week.

9:43 Hermes, or actually the company behind Hermes, News Research, raised $90 million this month, actually on October 7th. Now, I know $90 million sounds like a cheap change in the current, times where people are raising hundreds of millions and billions, but this is going to value them at $1.5 billion. Now, those of you who don't know Hermes, Hermes is an awesome, awesome agent tool that kinda does anything.

10:08 So the same concept as GroqBot and Instinct and OpenClaw, et cetera. They were the biggest craze after OpenClaw, and they have a huge following around the world. Now, in addition to the fact they just raised a significant amount of money, they're a full open source model. So they are an MIT license, which allows you to basically have access to anything that you want and change more or less anything that you want in the model Their new direction, and why do they need the $90 million, is to focus on Hermes for business.

10:38 So basically allowing companies to implement and integrate this into their businesses. So all of that is nice, but how does that relate to the current trend that we're seeing? Well, the post from their founder on X said the following shortly after they raised the money: "You should be able to use smarter models in your agent.

10:55 You should be able to use cheaper models in your agent. You should be able to use local models in your agent. You should be able to use a different model for every job. You should be able to switch models in the middle of a conversation. You should be able to choose what data your agent has." And then he goes on and on and on.

11:11 Basically, the bottom line of all of this, your agents should be yours. That's how they sum it up. But the first five or six lines of the things you should be able to do with your agent have to do with the ability to choose. Choose different models, for different tasks, choose different models based on price, choose different models for whatever you wish.

11:28 And they're making it very clear that this is what they are selling. They're giving you the option to choose. And again, I think people, companies, organization will want that flexibility, and I think everybody is going to go in that direction. And I've been saying this for a while, and now it's just proving itself out Now, If you want other examples of this that existed for a while, Microsoft Copilot allows you to choose between Claude and GPT models.

11:54 It's been that way for a while now. GitHub Copilot allows you to choose different models for a while now. Amazon on its AWS Bedrock allows you to choose multiple different models and apply them however you wish. the only leading companies who haven't done this yet are Anthropic and Google. but I assume that this will probably be coming, again, because I don't think they will have a choice.

12:17 I think Google, for now, are riding the incredible distribution that they already have, and Anthropic are riding the hype and the advanced models and that so many people want to use them. But I think over time they will not have a choice. They will have to offer these kind of things. Otherwise, people are just gonna switch away from getting their licenses to just getting some kind of a router and using more other models, and they have no control and no other way to make money in the middle.

12:41 So again, it will make perfect sense for them to do the same thing. But what does this mean for your business? What does it mean for you when you decide which models to use? The first one is don't get married to any of them. So stop choosing a specific AI vendor and committing to them completely, and find platforms that gives you the flexibility to choose other solutions.

13:03 If you don't do that, at least give yourself a few options, so you can still get your ChatGPT or Claude annual subscription for your enterprise or company, but then get a few other options that will allow you to choose. So some kind of a router that will allow you to choose models for other tasks. Thing route by type of the things that I'm doing.

13:27 When I'm doing planning, when I'm doing strategic work, I'm choosing the more expensive models. When I'm doing the execution after I have a very detailed plan, I will do it with small, cheap models, not necessarily from the same vendors. And in all cases, I build everything around the environment in which I can switch in real time or work collaboratively.

13:48 in my case, mostly ChatGPT and Claude on the same environment, and I recorded several episodes on exactly how I do that. So you can go back in the last few weeks and find, the explanation and listen to those episodes separately. Now, if you are in the ChatGPT universe, you need to check your OpenAI commitment and ask OpenAI what part of your tokens or your commitment for the year can go to other models, what other models are available, and then decide how and if you want to use them for specific tasks in specific scenarios in order to give you whatever benefit you wish to get, most likely savings Now, still on the enterprise, side, your harder problem is not gonna be switching models, but more

14:31 managing governance and data security. And so this is a hard task regardless of which models you use, so there is a benefit of working with one provider from that perspective, because once you commit to a specific enterprise plan, you know what governs that plan, and going through a general router may not provide you that.

14:51 And this is another potential benefit, assuming, and again, I don't know that for a fact, But assuming you get the same data security structure if you're using LLM or whatever other model through ChatGPT, which I doubt, but maybe that's the case, then there's a benefit in staying with one platform. Now, the bottom line relates to what I said about the way I work before.

15:10 Your biggest value and your biggest benefit and your biggest differentiator is not the model you're using. It's your data, your infrastructure, your workflows, the skills that you've built, the agents that you've built, the way the agents collaborate, the agent setup and how it's connected to different tools, and so on.

15:27 This is where you are going to drive the most amount of value for your business, and being able to switch models is just gonna make that environment more powerful because we will have cheaper, better models coming out all the time. More about that in just a couple of minutes. And being able to switch as needed while keeping your data and infrastructure and so on is the key to Having the ability to quickly evolve and make changes as the needs arise for whatever reasons, either security reasons or cost reasons or speed reasons or whatever the cause may be.

15:58 Now, speaking of cost and saving money, it has been very clear since about April this year that the competition is not gonna be just who has the frontier most advanced intelligent models, but also who can provide a lot of intelligent at a very cheap price. And so far, the winners of that until very, very recently were very clearly Chinese models.

16:19 That was their way to compete. So while they were slightly behind the frontier on the leadership side, they were definitely the leaders when it comes to providing solid intelligence at a very cheap price. And this have changed in just the past few weeks and changed even further this week So what changed?

16:38 Well, two things changed this week, and one changed just a couple of weeks ago, but again, it's all going in the same direction. The two things that changed this week, one is we got Haiku 5.5 from Anthropic. We're gonna talk more about this in a minute. The other one is that we had the release of Reflections AI's Beam, which is the first open source American model that actually competes with the Chinese models or almost competes with the Chinese models on both quality and price.

17:06 And the one that happened just a couple of weeks ago is that OpenAI launched, ChatGPT-6 Luna, which is an incredible and a highly competitive model. So let's dive a little deeper and starting with Haiku. So Claude Haiku 5.5 launched on Wednesday, October7th. It is now available on the API on AWS, on Google Cloud, on Azure, and obviously in the chat itself, and it's also added to GitHub Copilot if you are wanting to code with it in that environment.

17:35 The price is 10 cents per million input tokens and 50 cents for output tokens. That's not close to the scale of any other Anthropic models in the past. Now, this is true for prompts that up to 100,000 tokens. For prompts that are bigger, it is going to be a little higher. It is going to be 50 cents for million input tokens and two and a half dollars for a million output tokens if you're on the larger prompts.

18:03 But what Anthropic is saying is that currently 90% of all prompts that come to Anthropic fall within the shorter thing. So it is a very cheap model, even comparing it to the Chinese. We're gonna do a quick comparison in a minute., Now, to tell you how cheap that is, even compared to the previous Haiku model, so Haiku 4.5, which was the latest model that we had until a few days ago, was $1 for one million input tokens and $5 for the output tokens.

18:30 So that's a 90% discount while getting significantly more intelligence. The other huge jump is that it has 1 million token context window versus 200,000 in Haiku 4.5. Now, talking about the intelligence, and this is the big deal. The intelligence score of Haiku 4.5 from Artificial Analysis Index, which is the place that I go to see how stuff ranks, is 43 compared to Haiku 4.5.

19:01 That was 26 points lower. They more than doubled the intelligence of the model, so from 17 to 43 from one Haiku model to the next On some specific aspects such as computer use, Haiku 5.5 scored 72.4%, where 4.5, the previous model, was at 15.7. From 15.7 to 72.4. Now, the really amazing thing is not the fact that comparing it to the previous Haiku, but let's just make a broader comparison and see how good this model really is.

19:40 And so what I did is I went to past models. I said, I wanted to see for Haiku 4.5, how does it rank compared to other models? And the first thing I compared it to is the top models today and the top models of yesterday. So if you're thinking about the top models of yesterday, basically the frontier models of OpenAI and Anthropic, here are the numbers.

20:01 Claude Opus 5, which was the best model in the world through this summer before the launch of 5.5 and 5.6, so it was a model that was released on July 24th and was ahead of everybody else by a big spread and was a big part of the huge spike in the usage of Anthropic, have a score of 45. Again, the new model, Haiku 5.5, has a score of 43, almost an identical score.

20:27 The cost of Claude Opus 5 when it was launched was $5 for a million input tokens and $25 for the output tokens. That's 50x more expensive than the current Haiku for a two-point difference on the intelligence score. Claude Haiku is very verbose, meaning it's not going to actually be one-to-one, so you're not gonna save 50x to get roughly the same level of intelligence.

20:53 Let's say you're going to save 25x. This is still insane when we're thinking about the fact that the flagship model was released on July 24th, and we're just in the beginning of October, and we are, quote-unquote, "pacing the frontier." Now, if you want an even greater comparison, the most capable model from OpenAI before the recent version was GPT 5.4 on extra high effort.

21:18 This model was released on March 5th, so just over six months ago, its score was 39, again, compared to, Claude Haiku of 43. So the current small, cheap, inexpensive model from Anthropic scores 4% higher than OpenAI's flagship model from just over six months ago Another example, Claude Sonnet 5 on max performance, scored 38, and GPT 5.6 Luna, so the latest small efficient model from OpenAI scores 37.

21:52 So Haiku scores six points more, and it is competing across the board very nicely with the latest and greatest models. and it competes very nicely with the previous version op models, not the small effective, not the Chinese ones, and so on. This is absolutely mind-blowing because I hear people, when I tell people I'm using, ChatGPT Luna and now I'm gonna start using Haiku more and more, and they're like, "Oh, are you crazy?

22:18 Why are you doing this? They're so small models, and they're not effective." I'm like, "They are at par or better than the best models we had in the first half of this year." Now again, a few more words about the competition with specifically Luna. So ChatGPT Luna launched on September 22nd, so not too long ago, and it is the same price as Haiku.

22:38 It's 10 cents for input tokens and 50 cents for a million output tokens. it is less verbose, so you're actually gonna get more off of it, per task. It is a very capable model. Most of the code that I'm writing right now, and it is a lot every single day, I'm writing with ChatGPT-6 Luna on high or max, and this is done because I have extremely detailed requirements written down, and so it is very good at executing these requirements, so you don't need a very smart model just to execute if you have very clear requirements defined, and the requirements are defined by more advanced models But the real story that I'm trying to tell, and that's how we started this segment, is how does this compete with

23:18 China? And if you remember, until not very recently, we talked about everybody will switch to Chinese models because they're providing you sufficient, so not the best intelligence, but sufficient intelligence at very competitive pricing. Well, let's look at how that looks like, and we'll add another parameter that I told you multiple times about, which is the cost per task.

23:37 And this is a parameter that we're getting again from Artificial Analysis, and they're calculating across the entire set of tests that they're running, how much it is going to cost to run it, and then they're dividing it by the number of tasks. So it doesn't necessarily mean the tasks you are running, but it gives you an idea of how much it's gonna cost to complete tasks when you're comparing the models between one another.

23:57 So Claude Haiku 5.5 is now at twenty one cents. JPT 6 Luna is at seven cents, so a significant discount. But again, comparing it to models from China, if you're comparing it to a Flash model, so let's say GLM 5.3 Flash, it is at twenty five cents to complete a task with a intelligence score that is basically identical to Claude Haiku.

24:22 If you're looking at DeepSeek 4.1, it has a score of thirty-nine, again, compared to, Claude Haiku forty-three, and it's gonna cost you twenty seven cents to complete each task. If you look at the Xiaomi, Mimo v.26 Flash, again, their Flash model, they're at six cents, but with slightly lower intelligence.

24:40 Now, if you go to the top-of-the-line Chinese models, so not their Flash models, but their top-of-the-line models. So Mimo version 2.6 Pro is at thirteen cents with a score of forty-six. So again, I remind you, this is compared to Claude Haiku of a score of forty-three, very close together, with twenty one cents.

24:59 So you're gonna pay a little more, but you're using an American model from a company that you know where your data is going, that has full protection, of your data, and so on. not at a 50x price, which what was the case until, well, just a few days ago, if you wanted to use an anthropic model. the same thing with GLM.

25:14 So GLM 5.3 on Max has a score of forty-five, and it's gonna cost you two dollars to complete a task. Qwen 3.8 Max is gonna cost you five dollars and forty one cents, and Kimi K3 is gonna cost you two dollars to complete these tasks. Again, compared to twenty one cents on Claude Haiku 5.5. So we are now at a state where the leading US labs now have models that are at the same level roughly as the leading Chinese models and at the same ballpark of price range, and in many cases, a lot cheaper.

25:47 So again, if you're comparing the leading Chinese models, they're gonna be more expensive by a big spread per task compared to the small models from the US right now. Now, the newer addition to the family or to this competition is Reflection. They just added and released a new model called Beam, and it is the first open source American model that actually competes well.

26:11 So we had other models before that were open source American models such as NVIDIA Nemotron, and Google and Meta have their own families on open source models, and, OpenAI released a few open source models, but all of these were not really competitive with the leading AI models. And now Beam actually puts itself, well, not at the frontier of the Chinese models, but very close to that.

26:33 So this new model was released on October 5th. It's a 501 billion parameters, open source model, and what they're claiming is it is going to be one of the most token efficient models in the world. They're saying that it matches GLM 5.2, and that it is matching the previous generation of Qwen as well.

26:54 So they're not trying to compete with the latest Qwen model or with GLM 5.3. They're one step below that, but this is just the first model they are releasing, and from a price perspective, they are in a good spot as well. Now, this model also does very well on the benchmarks. Again, they haven't released it to third-party analysis yet.

27:15 We'll probably have that next week. But based on their own assessments, Beam, the new model from Reflection, scores 80.1 on Terminal Bench 2.1, which is the same score as GLM 5.2. It is seven points behind GLM 5.3. It is seven point behind, Kimi K3. five points behind, Qwen 3.8. So again, it is not aligned with the most frontier models from China, but is not very far behind, and it's an open source cheap model that is focusing on being efficient.

27:42 It is very similar in the gap when you look at other benchmarks such as, DeepSWE 1.1 and SWE Bench Pro V2 and Humanity's Last Exam and so on. It is not comparing itself to the latest and greatest Chinese models, but it is aligned with the previous version of those Chinese models, and it's going to be very competitive on price.

28:04 So what does this mean for your business? first of all, it means that you do not have to take the risk with Chinese models if you want to run on a cheap model to do a lot of the tasks. Now, to be fair, and I said that multiple times on this show, I think most business tasks today can effectively and efficiently run on Haiku and on ChatGPT Luna, 5.6.

28:29 Why am I saying that? Because I'm doing it, and it's actually working really well, and I'm doing this across the board for myself, for several clients, and it is working for multiple tasks without any issues. What does that mean for you? Go and test it. If you have specific workflows that you're running regularly, whether you built skills for them or you're running it through whatever, automation that you're running, or even if you have prompts that you saved in a prompt library and you still haven't built skills for them, which you should, you can go and test it on the new model and just compare the outputs.

28:57 Run it twice, run it with whatever model you're running right now, and then run it with a cheaper model and see what the results are. If consistently you're getting better results after you run it six or seven times, then just switch. Never look back and forget about the frontier models because they're not relevant to the task that you were testing it for.

29:14 Now, a few things to still pay attention to. One, as I mentioned, Haiku's price jumps 5X if you are sending really large prompts. So if you're sending over 100,000 tokens, you're gonna pay more money for those. So for those, you may wanna choose a different model, going back to our previous segment of routers and so on.

29:32 you can use ChatGPT Luna for that if you have access to that model one way or another Now, the last thing that I will say that if you need the data to never leave your servers, so if you need an open source model, you now also have choices from American vendors, not just Beam from Reflection, but other models as well.

29:51 As I mentioned before, Google has released a few very capable models on their Gemma family, and Nemotron is releasing models all the time, and Microsoft have released a few, interesting models as well. So you still have American options that you can run on a server that you own, either on Azure or AWS and so on, and have full trust on the model and the where it is stored on US servers without taking any risks.

30:16 and I guess during the current times, we cannot complete a deep dive section without talking about AI security. just over a week ago, on October 1st, OpenAI fired three of its researchers, Mikita Belizny, Tomek Korback, and Jasmine Wang were terminated at the same time. OpenAI stated the reason that they were potentially sharing sensitive information with an outside AI safety organization without being authorized to do so this week we got the feedback, or if you want, the argument from the three people who got fired, and they're saying that they got fired for, and I'm quoting, "Prioritizing safety over near-term interests of OpenAI as a corporation."

30:59 Now you may say, "Okay, they're holding grudge, because they got fired, and this is why they're saying what they're saying." first of all, it's not the first time that this is happening, right? we had similar incidents going back as far as 2024 When Leopold Aschenbrenner was let go, and Pavel Izmailov, both were let go from OpenAI for similar situations.

31:18 They were both in the safety environment at OpenAI. They both raised serious flags on safety after departures. we had a long, cover of the Leopold specific incident in the article he wrote, at that time. But another big thing happened this week, which is David Robinson, was the safety transparency lead at OpenAI, he resigned, so he wasn't let go, after serving in safety roles at OpenAI for three years.

31:46 He was one of the leaders and was part of writing the safety reports for every major product launch as OpenAI, making his departure very clear that he has knowledge on what's actually happening at OpenAI. And he wrote an article to The Atlantic, and he said, and I'm quoting, "The future depends on wisdom that Silicon Valley lacks.

32:07 Wisdom about how to handle dangerous technology and, more fundamentally, wisdom about what it means to care for people." And he said more than that. he was saying that the culture at OpenAI is of releasing things, seeing what went wrong, and fixing it. He's saying that's not just OpenAI, that's Silicon Valley, and that is true.

32:26 The mindset behind a startup, any startup in the world, is run fast and break things. That's how startups become successful. The problem becomes is when the technology that these startups develop can make dramatic changes to the future of humanity, like OpenAI technology does, and then this concept stops being logical.

32:46 If the implications of something breaking might be, if we're going back to our conversation from last week and Bill Gates' thoughts, hundreds of millions could die or suffer, then moving fast and breaking things stop making sense and something else has to take the place of that in the mindset of these companies, and this is the reason that David Robinson resigned.

33:08 He's saying that this is the culture at OpenAI right now, and he thinks it is too dangerous to continue that way. Now, he's not the first person to do this. Just a few weeks ago, we talked about Jacob Coxon, who resigned from Anthropic for the same allegation, saying that we're moving too fast and that not enough effort and resources are invested in safety infrastructure, and it's not nearly close to the amount of resources that are invested in developing the next model.

33:35 Now, if you think that what they're saying is wrong in any way, all you have to do is go back to the website we talked about before, Artificial Analysis, and they added a new graph that talks about, and it's called Frontier Language Model Intelligence Over Time. It is basically showing the releases over a step time over the timeline, and what you can see is that the releases is going higher and higher on the intelligence index and closer together as we're moving closer to today.

34:03 So what that tells us is the acceleration is keep on happening. It's not just that we're getting better technology, we're getting better technology at a much higher pace than we got before, which means that a lot of money and a lot of efforts are invested in that. And what they're all claiming is that nothing close to that investment goes into safety and security.

34:24 Going back to our conversation from last week with Jensen Huang, who said it's all an engineering problem. if the engineering problem is not getting enough resources, it has no chance of being successful. So if most of your efforts is going right now towards development and a small percentage goes to safety, the chances of safety being safe is relatively low.

34:43 What Jensen said is that they're investing 80% of their resources towards safety, and I'm willing to bet that's not the case at OpenAI and Anthropic Now, if you think this is just a business decision or it's just a something that happens because of nobody's paying the right attention, it is very clearly not the case.

35:03 If you think about OpenAI's mission statement, when OpenAI converted to a for-profit structure, they removed the word safety from the organizational mission statement compared to what it was before. The change is not just rhetorical. Under the new structure, a safety and security committee, on the nonprofit side of the board, is the nominal authority to require mitigation measures, including halting releases.

35:30 But the OpenAI group PBC board, so the board on the for-profit, public benefit corporation, is now the operational center of gravity, and it does not have such a committee as part of that board. So from a decision-making process, it is very clear where the focus is Now, the other thing that happened this week that relates to safety in OpenAI was something that was said by Sam Altman himself on an interview on the Decoded podcast by Politico.

35:57 And he said, and I'm quoting, "We believe that the world should accept some bad things happening for the benefit of this technology and the people having the agency What does that mean? It means that OpenAI, and I'm sure all the other labs as well, understand the risks. They understand the potential negative implications of the technology, and they believe that this is acceptable.

36:20 They believe that the risk they're deciding to take for us across the board, and again, these could be very severe if you are taking the thought process behind Bill Gates, as we discussed last week, they believe it is a reasonable risk to take and that the downside is gonna be minimal compared to the upside that can happen there.

36:39 I believe that it shouldn't be their decision to make. Going back to my thought process on this, I think there has to be some governance on what decisions they're allowed to make for humanity and what decision they're not allowed to make for humanity. And being technology companies, there's a lot they shouldn't be allowed to make.

36:58 But if we're looking at the current administration and we're looking at the accord or the ridiculous accord that was signed last week, that is basically saying that they will govern themselves, is not seem to be moving in that direction. Now, to be fair with OpenAI, they did scrape the release of ChatGPT 6.1 Astra that was planned for DevDay, because they found it to be not safe enough And they said, and I'm quoting, "We will resume only when we're confident we have additional safeguards and alignment in place.

37:29 We are working on these now. This is not the first time we've paused to take such measures, nor do we expect it to be the last as AI capabilities continue to advance." So they are monitoring what's happening with their AI tools, and they are taking into consideration safety risks, and they are stopping releases of models when they see them as critical, and that's the level that they're defining themselves, a critical risk rating for a model.

37:55 So it's not that safety's just being ignored. I don't think that's the situation. I just think it's not getting enough resources and enough focus, because of multiple different kinds of pressures that is both internal and external on these companies However, that being said, if you want proof to the fact that they are very good at explaining how their models are safe, but in reality they're not, another piece of news that came out this week came from a company called Common Sense Media Youth AI Safety Institute, and they released a report on October 7, rating different models and how they impact kids, and their rating for ChatGPT for Teens is unacceptable risk.

38:33 That's the actual rating. The way they did this is they ran 4,000 prompts across multiple accounts for registered users between 13 to 17 years old. Again, that is supposed to have protection for kids and teenagers. they're reviewed by child psychiatrists and pediatricians to see how ChatGPT scored. ChatGPT for Teens, or that was launched just earlier this year on August 18, its core idea was to block younger generation from seeing sensitive content and hundreds of prompts related to suicide, self-harm, disordered eating tested on newly created accounts that were all supposed to generate parental control guidance, from the teens version of ChatGPT generated zero parental control notifications Now

39:21 again, to be fair, this was not completely aligned with the, quote-unquote, "rules of OpenAI." On OpenAI's defense, accounts need to be linked for approximately three hours before notifications get activated, and a lot of these accounts was fresh accounts, They might have gotten parental control notifications if somebody would have waited.

39:39 But who's gonna stop a kid from just starting an account and starting to ask questions? He's not going to wait for three hours ChatGPT for Teens is supposed to send referrals, for specific safety situations, and it failed to do so in one in four attempts. So again, there's the promise on one side telling you this is going to be safe, it's going to monitor, it's gonna have parental control, it is going to send notices to the right scenarios and raising flags as needed, and it's doing this some of the time, which is just not enough, especially when you think that these guardrails are in place Now, OpenAI's response was that they were not running the tests aligned with OpenAI's guidelines.

40:16 but Robbie Turney, the head of AI and digital assessment at Youth AI Safety Institute, who ran said, "Keeping kids safe requires rigorous independent evaluations like ours, and we hope OpenAI will engage on a substance of what we found rather than trying to sow doubt on our methodology." So what does this mean?

40:36 Well, it means first of all, that it is clear from multiple angles that safety is not the focus. I'm not saying they're not caring about safety. I'm saying it's not the focus of any of these labs, and that is very, very scary with the potential outcomes and the potential implications on all of us in multiple different ways, whether you go with the path that Bill Gates talked about of chemical and biological weapons, or really severe cybersecurity attacks, or if you're just thinking about your kids and how they're using and what they're asking ChatGPT, or if you're thinking on potential job losses and so on.

41:13 None of these things are being addressed in a serious enough way, and the current accord from the government is not doing anything in order to stop this or make this any safer. It is instead focusing in changing the name of artificial intelligence to super intelligence, versus the terminology that everybody's been using for more than a few decades And all I can say is that where there is smoke, there is fire.

41:40 Like the amount of alarms that's been going on in the past few weeks, and definitely in the past few months if you're starting from the Hugging Face incident, is growing exponentially. It is growing from the actual events that are happening. It is growing from the, quote-unquote, "feedback" that we're getting from people who are departing from these organizations.

41:58 It is growing by the words of the leaders themselves. If you think about what Sam Altman just said this week, that we need to be willing to accept some pain. I think we're at a very important point on the timeline of AI development, and I really, really hope that we're gonna start seeing some real actions being taken in order to take safety under a much more serious consideration Now to our rapid-fire items.

42:23 the first one, we're going to talk about the top 100 GenAI consumer apps. It's the seventh edition of this report Coming from Andreessen Horowitz, and they've been doing this, very consistently. And here are a few findings, and I'll go through this quickly. Again, this is a rapid-fire item, but there's a lot to cover, so I'll try to go through the main findings quickly.

42:40 First of all, ChatGPT still remains the market leader when it comes to consumer apps, with a very big spread, both on web visits and mobile active users, holding about a 2X lead over Gemini and a 6X lead over Claude in August of this year, which is when this is the last data in this report 12 monthly active and 14x lead on mobile active, users on Claude.

43:08 So again, ChatGPT is still ahead, Gemini is still number two, Claude is still number three, Now, in addition, ChatGPT also has three X more paid subscribers than Claude or Gemini, and Claude has emerged as a very clear number three, skyrocketing in the past 12 months, surpassing competitors like DeepSeek, Perplexity, and other Chinese models on the consumer subscription of things.

43:34 So Claude climbed from being absent on the first report of this, which was September of 2023, to its current number three position. They have a very successful monetization base, so seven point three percent of Claude's consumer base subscribers, to the Max plan, which is $100 a month, compared to one point three for Google and one point one for ChatGPT.

43:57 So significantly more Claude users percentage-wise for Claude versus other companies are paying-- are getting paid members. ChatGPT global web traffic share fell from seventy-nine percent in May of 2025 to fifty-three point nine percent in May of 2026. So again, they're losing a lot of market share, with Gemini surging approximately four hundred and fifty percent and Claude approximately eight hundred and fifty-five percent.

44:24 So a eight and a half X growth for Claude year over year It was also the first year that ChatGPT's AI assistant app market share fell below 50%. that happened on March of this year Now then another thing that is very clear from this report is that personal always-on assistants and agents such as Instinct and Tomo and Poke and Lindy and Town and the big players, so Meta Muse and OpenAI Dots and xAI's Grok bot are now a big deal and they are growing dramatically.

44:59 Muse, which is the most successful one of all of them because Meta is pushing it across their incredible distribution, have reached five million downloads in just 22 days and 3 million weekly, active users in the first three weeks, outpacing how fast ChatGPT, Grok and Claude, were accumulating users when they got launched The other thing that is very clear is that the big players are taking most of the market share, and in order to play as a smaller players, you need a clear niche.

45:25 So that niche could be, being a company like Figma, who is number 15 on the list because they have a very clear direction of what they're doing, or Suno or ElevenLabs that are pushing in a specific type of creativity or very well-tailored solutions such as Cursor or OpenRouter that gives you very specific kind of things.

45:44 So this is the way for smaller companies to compete with the big players right now But the biggest thing that we can learn from this new report is that the vast majority of the US population are not paying for AI right now. Almost all revenue of these companies currently come from the direct subscriptions.

46:07 So 84% of total AI native, products come from subscription. Again, we're talking about consumer models and putting aside everything that 14% of overall revenue from AI, and most of that is coming from ChatGPT's. they launched their advertising, and they just reached one billion analyzed revenue, run rate in August of this year using their 1.2 billion weekly active users.

46:31 So the main engine of the internet previously was ads, and it is definitely not the case. It is subscriptions. And as I said, the subscription, the paid subscription base right now is very small compared with the overall distribution of this channel, which means these companies will have to figure out a way on how to monetize their success, how to monetize the access that they're giving, because right now, the vast majority of people are not willing to pay, for their AI usage.

47:01 I am shocked and baffled but not really surprised when I think about it. as somebody who's paying for the max plans on multiple of these platforms and would have paid a lot more if I had to for what I'm getting out of it, it is very surprising to me that 84% of people are still not paying And I think the reason is that people are just not aware of what they're leaving on the table.

47:26 People do not know how much value they can gain for their personal lives or for their businesses by paying for the better models and for access, for more tokens. And I think that will change over time, but it might take a while Now to real rapid fire items. We're gonna go through a lot of stuff quickly, we're gonna start with recent launches and announcements of either features or new things from the leading companies.

47:49 So OpenAI just launched GPT-6 for everyone around the world. So the global rollout has started. It was mainly focused on the US and a few other countries, and now it's going to be rolled out to everybody, starting with paid tiers and then the free people as well. as I mentioned, Claude released Haiku 5.5.

48:06 We had a big deep dive into this, so I'm not gonna mention a lot about that Mistral, which are a French company that we talked about several times but not for a while, have launched Mistral Large 4, which is their latest model It is a 1.05 trillion parameter multimodal model that is designed to be close to the frontier with an open source model It is built very much with cybersecurity in mind, so it ranks among the top five models globally on the Artificial Intelligence Cyber Index, and it leads the open weight models, that is developed outside of China on that topic as well.

48:43 Obviously, it's built around agentic and enterprise workflows, so a very capable new open source models from France, from the leading European company, if you want, in the AI space Now, Anthropic shared two interesting things as far as helping people learn how to use AI. One of it is that they are going to invest $100 million in an academy to train 10,000 enterprise AI engineers by the end of 2027.

49:07 They're calling it the Claude Frontier Academy, which is a first of its kind program, backed by, again, a lot of money to train, again, 10,000 people on how to implement AI The idea is to teach people how to use AI to move from concepts to production at big companies The first cohorts of these courses are gonna be engineers from Accenture, Bain Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and a lot of companies that you already heard of.

49:33 And then they're gonna go down the totem pole until they will eventually reach companies that you and I work for and have a chance to actually take these courses Now, to be more specific, this course targets hands-on software engineers with strong software fundamentals. So that's again the current focus right now.

49:49 And again, it's just 10,000 people all around the world that are going to go through this course right now. But the other thing they announced is a program for startups that is gonna give startups access to not free, but a lot of tokens at a discount or free up to a certain amount to grow the startup on the Claude platform.

50:07 So if you are in the startup world in early stages and you want to get access to more Claude for lower expenses, that is a great plan to try to sign up for OpenAI had a big announcement this week that could have easily been another deep dive section, and they have, released a vast collection of new mathematical results that is comprising of 722 manuscripts organized into 372 result families produced by an unreleased internal model that is finding solutions to more and more mathematical problems that did not have solutions, many of them for decades or hundreds of years Now, very different from the previous Navier-Stokes proof, which OpenAI reportedly required 10,000 agents and millions of dollars

50:54 in compute, nearly all of these new 722 results were achieved with a single prompt each to a single AI agent, each averaging about three hours of ChatGPT Pro compute. Again, this is a model that we don't have access to that is probably specifically trained, to solve problems in math. Now, to be fair, within 24 hours after the release, OpenAI withdrew, a few of the mathematical manuscripts, due to error issues.

51:20 but still, a lot of them were accurate Now the world of mathematicians obviously were not in favor of the current situation. The statement said, "The future of mathematical research cannot consist only of understanding results produced by AI labs. Mathematicians must be able to formulate their own questions, develop their own approaches, explore directions that have not been selected as examples of an AI system capabilities."

51:47 The Association for Human Mathematics condemned the release What they're saying is that OpenAI exhibited, and I'm quoting, "Total disregard for the norms of scientific research." that OpenAI ignored the initial recommendation of the body that they set up in the research that advised them against testing advanced mathematical problems on internal models They argue that mathematicians did not request this work and categorizes the simultaneous publication of over 700 files as a demonstration of power rather than a genuine scholarship Now, they also called for a boycott of mathematicians to discontinue their work with OpenAI and return to how science is supposed to be around human-centered understanding.

52:30 So where does that leave us? first of all, OpenAI and all the other labs are gonna continue to try to solve whatever science problems they can. A, because it allows their models to progress. B, because there are a lot of benefits in that, right? While they're doing this in this particular case in math, I have zero doubt we'll start seeing more and more of this in other aspects of science, and AI will be a complete game changer in finding novel new ideas and concepts and solutions to many big problems we have in the world right now, and saying no to that just because it's doing it faster and without humans is...

53:06 I understand how frustrating that is for scientists, but I think the same issue is gonna happen to all of us. so this contradicting feeling that I have every single day that on one hand is really exciting and saying, "Oh my God, I can do 100 times more with my time and explore new things and be a lot more creative because AI allows me to do it," the same thing will happen in science.

53:27 But at the same time, looking at the longer horizon, it is very clear that I will not be needed to do a lot of the things that I'm doing right now to make AI do what it's doing for me right now. It will just do it on its own, and then the question of, what will I do is becoming more and more complex.

53:41 So the same issue that scientists are feeling right now, everybody else will feel at a certain point. I don't know if everybody. Most everybody is going to feel sometime in the next few years Now staying on OpenAI and new releases, they just released a new meeting plugin that automatically records, the meeting, takes notes, generates personalized summaries, allows you to engage with it in real time while the meeting happens.

54:05 it identifies next steps during calls, and And the feature is currently available as a beta for Pro and business users on macOS desktop and on the Enterprise platform. It is going to be available very soon Now different than just regular notetakers, which I've been using for a few years now, the plugin leverages the knowledge of ChatGPT of you as an individual, the work history you have, and so it can customize and be more specific and tailored to your actual business needs.

54:32 If you think about slightly into the future and you combine that with dots and spaces and everything that they're building, it is very obvious that the plan is to have an agent that is always integrated into all your work, understands everything that you're doing, is always on, and also participates in meetings and can be significantly more proactive and helpful in every scenario because of all of that.

54:52 And again, there is zero doubt in my mind that's how work is going to happen for some of us in the very near future, and for the rest of us will slowly be integrated into how we work in the next few years Anthropic has launched two really interesting features. One is called Claude Dashboard, the other is Claude Motion.

55:09 They're a new suite of AI tools that are designed to streamline data analysis and presentation for business professionals. So Claude Dashboard can connect to enterprise data platforms like BigQuery, Snowflake, Databricks, Amazon Redshift, ClickHouse, Salesforce CRM, and so on, and construct real-time dashboards using just natural language.

55:28 So if you had stock in companies that build these kind of solutions, sell the stock Because you can now do this with AI significantly faster than learning the scripting language of whatever tool you were using so far. Now, Claude Motion, the other tool, allows to create motion animations with text and charts and shapes and images, from reports and other data sources that you have to explain and convey whatever message that you want And everything it generates, every word, every number, every timing of that mini video is fully editable inside of Claude, and the resulting animation in the end is an explainer that can be exported as an MP4 file that then doesn't require any kind of AI tool in order

56:12 to play it. Those of you who knows what Gemini Notebook can do when I mentioned that last week when we talked about Spaces by OpenAI. I think the future of collaborative work and sharing information is a lot more exciting than everything we're used to right now. I think we're gonna see the boundaries that we are used to, such as you either give me a document or a PowerPoint or an Excel or a PDF, all of that is going to go away, and we'll be able to generate near real-time whatever results we need to deliver whatever message we need.

56:44 I've said that multiple times that the age of AI requires us to separate the data from the delivery platform. Right now they are one. When I'm delivering an Excel file to you, the delivery platform is Excel, but the data is Excel as well. Same thing with a document, same thing with an email, and all of that should go away.

57:03 The data will be the data, and the delivery method can be decided in near real-time because the AI can generate it on the fly and give us much more relevant results as we need them for what we need them at the time we need them, if the data is accessible On the data security side, Anthropic made several announcement this week.

57:22 One of them, and a big one, is a new flagship program that is called Critical Infrastructure Defense Program, CIDP. The CIDP brings Anthropic frontier models, to on-site engineers and threat research to defend operational technology across power grids, water systems, factories, transportation networks, and so on.

57:40 So going back to the cyber risk that exists right now, the goal of this plan is to give the people who are trying to protect the infrastructure of our world the edge, or at least some time ahead before the cybersecurity attacks come, and I find this to be something very smart. The initial hope that this is a way for us to protect major infrastructure around the world from the cyber risks that are obviously coming.

58:06 Another big announcement from Claude is that they're rolling out their Claude for Google Workspace. It is now available in public beta for all paid Claude plans, and it's a seamless in-app editing. With Claude, you can run Claude inside a sidebar next to open anything on Google, so you can work on it on documents and spreadsheets and so on And there are two ways to do that.

58:32 Again, you can either run it in the sidebar inside the Google environment, or you can run the new beta connectors inside of Claude and connect them to documents and edit them, documents, spreadsheets, presentations, whatever you want It also has more controls from a work control, so you can ask it to be in ask before you make any edits, and then it will just suggest what it's going to do.

58:53 You can accept edits and changes or make comments to them before you actually make changes to the documents And they're adding a lot of enterprise-grade compatibility tools on top of that to make sure that if you are an enterprise using the Google environment, you can still apply all the safety and all the things you need in order to be able to use it in a safe way Another interesting announcement this week comes from Google.

59:16 Google just introduced Playground, which is their experimental gaming platform that basically allows anybody to create a game and share it with the world So the idea is you go in, there's a blank canvas, or you can remix from other existing games that other people created, start from that and add and tweak different things.

59:36 It has a physics and rules engine, and you can tweak them in real time. You can customize the characters, you can customize the environments, and you have full creative control throughout the development of every aspect of the game. And the games are not just flat 2D games. They could be really advanced multiplayer, with leaderboards and community and so on, and all of that is built into a shared community and environment where everybody can go in and build and remix and play these games.

1:00:06 This is a very interesting, approach from Google to an environment. I shared multiple times that I think that the concept of app stores will cease to exist because people will be able to create things in real time as they wish. This is a first step of a gaming version of that Another big and important announcement from Google this week is that DeepMind has announced a global extension of its SynthID detector.

1:00:32 So we talked about SynthID many times on this podcast. It is a back-end unseen traceable capability to identify any AI-generated content by Google. So anything you create with Gemini, whether you're creating images or videos or text or whatever the case may be, has a SynthID identifier built into the actual content itself, so there's no way to remove it.

1:00:58 And so far, it was not available as a detection tool, and now it is. So far, Google and a few selected organizations has access to it, but now they're rolling out global accessibility in a partner ecosystem to the SynthID detector, which means anybody can detect any AI-generated content, again, specifically from Google.

1:01:18 I really wish there will be some global agreement on a standardized way to do this. right now there isn't, but I really hope that's gonna be mandatory for open source tools as well, even though in open source tools, anybody can take the code, change it, and not put the ID into it. But if there's a law that forces you to do it, then you could do it from a technical perspective, but then you'd be breaking the law.

1:01:40 Like a lot of other things, we're not there yet, but this is a step in the right direction That's it for this week. I hope you found this episode, very educational. I hope you learned a lot. I hope that you are not totally terrified from the safety things and that you're managing to find the benefit and the value in the other things that we're talking about.

1:02:02 there is definitely an acceleration in the technology of AI. There's definitely an acceleration in intelligence, and there's definitely more and more adoption across multiple fields, including science. There are more news that we did not share on this episode that are all available if you want to hear about it in the newsletter that you can sign up for.

1:02:23 If you want more of this, you can join our Friday AI Hangouts. It's an amazing group of individuals that meets every single Friday at 1:00 p.m. Eastern on Zoom. It is completely free to join, and we have deep conversations about everything in AI from strategic things to tactical things. People are sharing what they built and demoing things that they've done and asking questions and getting answers.

1:02:44 so feel free to join us, for that as well. And we'll be back on Tuesday with another how-to episode where we're gonna walk you through how to implement AI in your business in an effective way. And don't forget, there are still a few weeks in which you can join our multi-agent orchestration course. It is going to be the last cohort of this year.

1:03:05 They're all selling out, every single one we run since April. So if you want to learn how to implement AI and build really incredible advanced systems in a safe way, sign up in the next few days because otherwise you will miss this cohort and the next one will probably be in 2027. That's it for today.

1:03:22 Have an awesome rest of your weekend