Transcription
All right, let's get right to it. What were the AI stories that mattered the most this week?
And new this week, I'm going to put in the post a special prompt for you to grab the implications of these stories depending on your company, your role, your domain of interest. You're going to be able to go like a step deeper with that.
Story number one, OpenAI is going proactive. And this has a lot of pull out thread implications I want to get at briefly. The launch is called Pulse, and it offers proactive AI assistance based on what you most recently talked about. So, think of it as sort of like an Instagram stories reel, but it's tuned to your previous chats, specifically very recent chats, like the last day or two. It's available to pro users initially. I've tried it out, and what I've found is that it's a very, very seamless, almost eerily relevant experience. It changes my behavior in some interesting ways though because it's pushing me now to start to have conversations a day or two in advance of when I think I need to have them to plan for work so that I give Pulse a chance to work overnight and give me interesting insights the morning when I actually have to do the work. It's always a good sign when a product launch is so impactful it immediately changes your workflow. And that's what I found here.
But the implications go beyond sort of pro user availability. What you see with Pulse is ChatGPT investing fairly transparently in an ad surface. It was always going to be challenging for ChatGPT to keep an aura of objectivity in individual conversations that users initiate. It will be much easier to position ads inside Pulse because Pulse is already proactive. Pulse is already an experience that you didn't necessarily ask for but ChatGPT is offering to you. So if they slide in a card in that sort of Pulse format and the card is sponsored, you can just ignore it if you don't want it. Right? It's a very simple ad experience. And of course at the same time as they launch Pulse, we notice that OpenAI has opened up an ads monetization role at ChatGPT. So something is coming there on the ads front and Pulse seems to be related to it. We will have to kind of put a pin in that and stay tuned to see where they go with that.
One final implication I want to call out on the Pulse story. This is the beginning of an entire new arc where we will see AI become more proactive. This is one of the first widely consumer available proactive AI experiences. And I want you to keep that in mind because a lot of the other threads we're following over the course of the last few months point to a 2026 that is more AI proactive versus AI reactive and major model makers are all building in that direction.
Let's get to story number two. Microsoft is diversifying Copilot with Anthropic models. And this is really fascinating because Microsoft has been known for leaning in on OpenAI for a while. But as we discovered in the last couple weeks, Microsoft and OpenAI have finalized a new looser partnership format. And it's not super surprising in that context to see Microsoft go for another major model maker as essential to their suite of tools. And I think it's the right call because if you look at how Anthropic actually performs, how Claude Opus 4.1 specifically performs on work that Copilot people do, slides, sheets, looking at docs, that's work Opus 4.1 does very well. It's not just me saying that. Ironically, it's ChatGPT saying that, too. OpenAI completed a study called GDP Val this week that actually tests major AI models against economically useful work tasks.
Now, before you run away with the wrong headline here, this is not the same as saying it's testing major model makers against economically useful jobs. These are very, very limited tasks where the context is prepared by an expert in a neat little package and then an expert prepares a gold standard solution and then the major model makers run their models against the problem space and see how their result compares to that gold standard in a blind evaluation. Fine. It's not real work in the sense that it's not as messy, but it is real work in the sense that the tasks are real and designed by experts in the field. That being said, OpenAI sponsored all of this and set it up and ran it and even they admit that Opus 4.1 from Anthropic is better at doing that kind of economically useful work than any of ChatGPT's models, which is a huge endorsement for Opus 4.1 and the Anthropic team and one that I personally have found correct. Like I prefer Opus 4.1 for preparing a slide deck, for creating a a spreadsheet. It is just much more useful than the ChatGPT models at this point. And so it's not surprising to see Microsoft pulling the Anthropic models into Copilot as they evaluate that. There are heavy rumors that this is going to get even better shortly as we suspect that Anthropic is on the verge of releasing something like a 4.5 version of Opus, which would presumably be a step forward here. So stay tuned for that. That may be coming in the next couple of weeks. We will have to see.
Story number three, Meta is exploring a Gemini partnership for ad targeting. So Meta is in early discussions with Google Cloud about integrating Google's Gemini AI models into Meta's ad operations. And Meta employees have explored how Gemini's multimodal capabilities like Nano Banana could refine algorithms that match ads to users on Facebook and Instagram. So basically, can you take the sort of Facebook, Instagram, social algorithm secret sauce that Meta has and marry it to the image generation capabilities potentially or the text generation capabilities that Gemini has.
This is a huge step back for Zuckerberg and Meta. And I know you might not have expected me to say that, but hear me out. They have invested an enormous amount of money in their own AI. Zuck just got done making headlines for the largest pay packages in history for his AI team and then very publicly lost about a half a dozen researchers who stayed briefly at Meta and left for undisclosed reasons but everyone guesses the culture. In this situation, Zuck is having to admit that Llama, his own homebuilt model, is just not good enough for this task and he's having to go to Google. This underlines how the race for AI has narrowed over the last year or so. A year or two ago, Llama would have been right in the race with the top leaders. It's just not now. It's just not. And the race has narrowed really to Google and OpenAI and Anthropic. And I know Grok is really trying, but like Llama's not even in the conversation. And even Meta is admitting that. And that is the undercurrent here. And that is a really big question mark for Meta because Zuck has not publicly suggested he's walking back his big dollar investments in AI for the future. He's still planning to spend hundreds of billions over the next few years on AI. Where is that money going to go? What do they anticipate? Is this a story sort of like Apple's chipset story where they're going to spend a vast amount of money trying to catch up and build their own chipsets and eventually they do? Or is it a story where Meta is going to invest this money and then gradually kind of roll it back as they realize they can't catch up? We don't know where that's going to go yet, but for now, it looks like Meta is not in the driver's seat, even in their own business, on AI.
Story number four, OpenAI is continuing to make headlines for how much they are spending on data centers. They are aggressively building capacity in ways that sort of it boggles the mind, right? The number, the the size of the numbers involved in terms of power generation, in terms of dollars. We are now over $400 billion in investment over three years in the Stargate project, which is their sort of flagship project. And they announced with Oracle five new AI data centers as a part of the Stargate project. What this is suggesting to me is that not only is Stargate not just a headline, Stargate is an umbrella term for the three to five-year vision that OpenAI has for a massive X or so scale up in compute versus present state, maybe more than 100x scale up in compute when it's all done. And they are not having trouble attracting the investment they need to get there. In fact, some of the stories over the last few weeks around Oracle's investment in OpenAI, around Nvidia's investment in OpenAI, this week around SoftBank leaning in on financing for Stargate, it all adds up to Sam Altman being able to attract the capital he needs to get this power generation vision done. And his vision goes beyond just building stuff. He wants to get to a point where construction of additional capacity for AI is fully autonomous. And there was a blog post about that that while it's aspirational is important to consider. Sam expects that he will be able to have in three to five years entirely autonomous data center production. So robots will put the data centers together. Robots will bring the chips in. There will be robotic chip fabs and they will be able to autonomously stand up new data centers to match the capacity needs of AI. You would only do this if you were expecting not just a 100x gain in demand for AI, but a thousand or 10x gain. And so Sam's vision is predictably incredibly grand. And that blog post gives us a sense of why these other companies are willing to put so much skin in the game, so many dollars to scale up on the data center side. What this suggests is that the demand story for AI remains intact and that we are going to continue to see aggressive buildout.
Let's get to story number five. Kimi is a Chinese AI company for its Moonshot AI's model, right? So the Chinese company's named Moonshot and they produced Kimi K, which is a trillion parameter model. It's very, very good and it is now available as an agent. It's called Okay Computer and it was launched on the 24th of September. It allows Kimi to access its own virtual computer environment with a file system, a browser, and a terminal. And it can autonomously execute complicated multi-step tasks for you. And so it can transform transform chat requests into websites or data dashboards or production documents, what have you. This gives us a glimpse of do the work assistance. But critically, it also is pushing the edges of AI agent autonomy and it's doing so from a Chinese perspective. And what I mean by that is that it is reminding us that the Chinese model development evolution continues to push the edges. And if you're like deep in the weeds technically, Kimi K2 also set a new bar on efficiency for training. You can look up Mu Nuon if you want to get deeper on that and you will see a lot on sort of how the K2 model was trained and how efficient they were at training the K2 model. That's just a little sort of nerd sidebar there. But this is really the larger story that the Chinese companies innovating on AI are pushing the edges forward, especially around open source. And they are very, very intentionally pushing straight to agentic. It's sort of a leapfrog motion where they're not just satisfied with LLMs, they're going to push into advanced AI agent capabilities. And while US enterprises and European enterprises may have enough data concerns to say, you know what, we don't trust a Chinese hosted Kimi K2 to do Okay Computer agent mode for us. Individuals will not have those constraints. And if the product experience is smooth and useful, individuals are going to start using this. So I look at Okay Computer as both a push on agent mode for the industry as a whole and you'll see US companies start to follow suit. I also look at it as a directly useful tool for individuals and potentially some small businesses that aren't too worried about the data side right now.
Story number six, Adobe is buying instead of building. Adobe has taken a beating recently in the public markets because their AI products are widely perceived to be missing the mark. And so they announced comprehensive third-party model integration across Firefly, embedding what the Luma Ray 3 model for video and adding a bunch more opportunities for model support from Google, OpenAI, Ideogram, and others. In other words, Adobe is deliberately importing best-in-class capabilities rather than trying to compete anymore with proprietary Firefly models. This feels a lot like the Meta story. Adobe is basically admitting their own AI strategy is dead in the water and they need to bring in other models.
Now the implication here is that Adobe is going to eat margin to do that. Notion very publicly has said during the Notion 3.0 rollout which also happened this week that they were eating their own margin and that they were costing their gross margin about 10 percentage points to roll out AI. So they went from 90% gross margins to 80% gross margins as a business to deliver AI functionality, but they felt it was worth it for the long-term value of the business. Adobe hasn't revealed their gross margins, of course, but I suspect that is the kind of impact they're going to have when they actually bring in these third-party services and start to pay for them and start to essentially have the brains of Adobe Firefly and other tools built by other people. So this is a story where even traditional SaaS companies are starting to have to let other models in the door because it's just too expensive to compete on the model side directly. Very, very interesting implications here for other businesses that need intelligence.
Finally, story number seven. California is advancing AI safety legislation. So, California's legislature passed Senate Bill 53, which requires major AI developers to publicly disclose disclose safety and security protocols and establishes whistleblower protections and incident reporting. It also creates Cal Compute, which is a public computing cluster for AI research. The reason this matters is because the major model makers are all based in California and because California is such a large and influential state in the federal system. And so I would expect that we will see actual implemented AI safety and transparency requirements at Anthropic, at OpenAI, both of which are based in San Francisco, and at Google, also based in San Francisco. And we are going to start to see public disclosure and discussion around what that means and probably implications for other states that may be looking at AI safety. California is setting the bar for the nation as a whole here, but it is also directly impacting the daily operations of the biggest model makers on the planet. And so if we're looking ahead at national AI safety legislation, California is showing a sort of way to go there, a direction to head in. But it's also going to require companies to change their behavior immediately. Now, OpenAI has said they're neutral on this. They don't have an opinion. Anthropic has endorsed the bill publicly and Google hasn't really given a signal that I've been able to find. Regardless, once this is signed, these model makers will have to comply. They will have to start to show that they have safety and security protocols in line with Senate Bill 53. They will have to show that they have whistleblower protections. There will be new compliance loads. This has implications for companies purchasing from those model makers, too. And that's not been figured out yet. So, stay tuned for more there. But I wanted to call out that that's a landmark in the US on AI safety legislation and one to watch.
As I noted, you can get more implications on each of these stories in the prompt that I prepared for you in the post. Have a great week. Cheers.