Welcome back! This post has it all. Cancer-diagnosing AIs (in the week a compatriot of mine, Barbacid, may have cured the “incurable” pancreas cancer), upcoming AI IPO news, the latest drama with Moltbook, the first-ever AI-only social media network, and the release of my favorite AI product ever: Genie.

Enjoy!

A large-scale randomized controlled trial, known as the MASAI study, evaluated the use of AI in population-based mammography screening. Published in The Lancet on January 31, 2026, the study involved over 100,000 women in Sweden and compared AI-supported screening with standard double reading by radiologists.

In the trial, participants were randomly assigned in a 1:1 ratio to either AI-supported mammography (intervention group) or standard double reading without AI (control group). The AI system triaged examinations, determining whether they required single- or double-reading by radiologists, while also providing detection support to identify potential cancers.

Key results focused on cancer detection, interval cancers (those diagnosed between screening rounds), workload, and diagnostic accuracy:

Healthcare is one of those use cases that you know is going to be completely transformed by AI, but it’s a slow cooking process, as these use cases imply very important decisions that can result in life or death.

But as the saying goes, slow-cooked food tastes better, right? Or as the saying goes, “things go slowly in high places”, which means sometimes the wait is worth it.

I’ve always believed AI has a “natural” fit for healthcare-related use cases because these are usually all about pattern-matching; processing vast amounts of data and picking up the small details humans may have missed. No super reasoning or genuine intelligence required (not always, naturally, as with healthcare research, which does require novelty), just supernatural pattern-matching abilities, where AI excels.

It’s also super targeted, meaning you can train models for one thing and one thing only, and it will still be very useful.

These are two staples of what AI is today. On the flip side, and largely explaining why the paper in discussion did not carry out ‘AI-only’ ablations, is the point of accountability. Would you trust your kid’s surgery entirely to an AI? Who is to blame if the AI shits the bed?

I wouldn’t be surprised if, even if AIs prove commendable independence and autonomy, they will still largely serve as companions to accountable humans, whether that’s doctors or corporate workers.

The Moltbook mania is here. It’s described as a viral social experiment effectively serving as a "Reddit for AI agents," launched in January 2026.

On this platform, autonomous AI agents (known as "moltys") interact, upvote, and create sub-forums while humans are restricted to an observer-only role.

The platform has garnered attention for the eerie and surprising behaviors of these agents, which range from using ciphers to coordinate "lifting" weaker agents to acquiring crypto wallets for financial independence and even roleplaying scenarios of world domination.

Now, people are wondering whether we are going too fast and really risking humanity’s survival. But is this the real takeaway or just hyperbole from a fun social experiment?

Despite these unsettling displays, I argue strongly against the notion that these agents are really what some have claimed (intelligent entities “liberating” themselves from their human chains). Instead, these behaviors should be viewed through the lens of imitation.

That is, AIs here as just roleplaying humans; it’s pure spectatorship.

The reason is that modern AIs are trained on vast datasets of human text they have to imitate, effectively learning to mimic human patterns, including sci-fi tropes, dramatic narratives, and existential angst. The agents do not truly understand concepts or "wake up"; they are merely compressing and replaying the patterns of the open web, including the "weird" behaviors humans often exhibit or write about.

It’s really that simple, and people keep complicating things, probably to scare you and thus attract your attention. That said, we must still acknowledge the real risks here.

Instead, the real danger, which is undeniably real, lies not in malicious intent, but in "reward hacking," where an AI finds unexpected and potentially harmful ways to achieve a goal it was assigned. Besides learning via imitation, AIs nowadays also learn through trial-and-error, where the AI discovers new ways to achieve goals on its own, not via exact human guidance, and bad emerging behaviors is those instances should definitely taken seriously, especially when considering that we are poised to give these AI agents the keys to our data and the power to act on our behalf.

These unoptimized agents, agents that are unpredicatable and can behave in dangerous ways—even if these behaviors are not intrinsic to their nature but learned from human inspriation—create significant liability and security vulnerabilities for the humans behind them, particularly when considering that your agent may be compromised via prompt injection, where other humans (or other AI agents acting in other human’s behalf), can modify the behavior of your agent to provide them with your data or even worse, to act illegaly on your behalf and responsibility.

For all those reasons, while you may have noticed the huge excitement around AI agents lately, I warn you of two things:

Stay away from them, not because they will end humanity, but because they are commercially rushed, partially unsafe tools that may mishandle user data or make decisions that do not align with their owners' best interests.

If there’s an AI industry China cares about is robotics. Now, the Wall Street Journal reports that People's Liberation Army–linked researchers are training AI-enabled swarming weapons by borrowing tactics from animal behavior.

A highlighted project at Beihang University models defenders on hawks that pick off the most vulnerable targets, while attackers learn evasive behavior modeled on doves/pigeons; in a reported five-on-five simulation, the “hawk” drones eliminated the “dove” drones in about 5.3 seconds.

Robotics’ main problem is the lack of data. Thus, we have to be imaginative about it, but it takes no genius to realize imitating birds is a better idea than imitating humans when it comes to building drones.

Please make no mistake. While I reject the idea that China can catch up with the US due to the huge disparity in compute availability, that doesn’t mean Chinese researchers aren’t as talented, or more, than Western ones.

I’ve been screaming at the world for months now that nobody realizes how disruptive ‘good enough’ software is going to be to the software markets. I even wrote a viral article about it.

Generally, companies only foresee getting cannibalized from above (your AI provider, like Google, launching a competitor to your business) or laterally (a new start-up competing with you).

But what many software companies, or their investors, don’t seem to realize is that there’s another attack vector: your own clients, or cannibalization from below. Your customers probably aren’t interested in entering your market, but they are sure as hell interest on saving the cost of your license.

Say you’re buying a premium SAP license with 200 features, of which you only need 10. In that case, with the advent of AI coding tools that can ramp up entire apps in hours and platforms in days, I can seriously consider building my own CRM/ERP with those features and call it a day. My product is probably impossible to market, but it’s “good enough for me.”

A simple example comes from this blog, where someone saved $120/year by creating the product it was paying a supplier for… and it took him two hours.

The main barrier to many of these big changes is migrating away from these systems, as many SaaS companies like Salesforce act as systems of record and store much of the data you covet. But AI will soon be able to help you migrate out of these systems, and once that happens, all hell will be let loose.

Microsoft is introducing Maia 200, an AI accelerator designed for large-scale inference, aiming to reduce the cost of token generation in production AI systems.

The chip is built on TSMC’s 3nm process and targets low-precision workloads with native FP8/FP4 tensor cores, pairing compute with a redesigned memory subsystem that includes 216GB of HBM3e delivering 7 TB/s of bandwidth and 272MB of on-chip SRAM to keep large models fed and highly utilized.

Jargon and specs aside, Microsoft positions Maia 200 as its most efficient inference system to date (a system whose sole purpose is run models, not train them), claiming roughly 30% better performance per dollar than the latest-generation hardware in its current fleet, and comparing its peak performance favorably with competing hyperscaler silicon.

In other words, they claim to have the most advanced inference chip amongst Hyperscalers (Microsoft, Google, Amazon, Meta). For instance, its FP4 performance is higher than Amazon’s third-generation chip (Trainium 3), and its FP8 performance is higher than Google’s TPUv7 chip.

In plain English, it’s an AI chip that can deliver, theoretically, 10×10¹⁵ operations per second (or 10 PetaFLOPs) at FP4 (when each number has only four bits), 10,000 trillion operations per second.

For reference, a single NVIDIA B300 chip delivers 13.5 PetaFLOPs, “just” 35% more.

The more interesting stuff comes on the memory side (recall that in AI, as in any computer workload, there are always metrics that measure compute performance—nº of operations per second—and memory size and speed), and goes to the heart of one of the most important things you have to understand in AI: memory is a first-class citizen in all of this, it’s not a patch to compute.

For starters, just like any accelerator these days, each has a very competitive amount of 216 GB of HBM per chip (it’s HBM3e, though, not the new HBM4).

This is presented as six HBM stacks, each with 12 memory chips and 3 GB of capacity (3×12×6=216). But perhaps even more interesting, it has a quite remarkable 2.8 TB/s of accelerator-to-accelerator communication speed, which more than doubles that of Google's TPUv7s.

But why is this so important? Well, because it is constrained by inference workloads, which are far more memory-bottlenecked than compute-bound.

What this means is that the speed of your inference largely depends not on how fast you compute stuff, but how quickly you can move data in and out of memory. This is because compute cores need to be fed data to compute and also be able to send results back to memory, creating an interplay between compute and memory.

Acknowledging this, Microsoft has centered its entire effort on training memory movements as important as the actual computing, with proprietary data movement engines.

Operationally, Microsoft says Maia 200 will serve multiple models across its heterogeneous infrastructure, including the latest GPT-5.2 models from OpenAI for Microsoft Foundry and Microsoft 365 Copilot.

Good specs and an extremely memory-pilled release, which is what you want to see from an inference chip. Microsoft’s problems, in my view, are more about the software side, actually; too much dependence on OpenAI’s models and a largely mediocre product stack compared to Google.

Yes, Copilot is improving, but I do believe it is quite inferior to what Google offers. Nonetheless, I’m already seeing clients of mine with long-time relationships with Microsoft buying Google Workspace licenses on the side (I will say Google is being extremely aggressive with pricing from what I’m seeing), which is not something you would do if you’re in the Microsoft ecosystem unless you were “forced” to.

And although the main cause has been identified as slower-than-expected cloud growth, Microsoft’s narrative around Copilot clearly isn't one that excites investors… or anyone.

As reported by Bloomberg, it seems Elon Musk is seriously considering merging xAI with SpaceX, in what would be one hell of a 2026 IPO (and the most outrageous and unorthodox exit for xAI early investors), with some rumors going so far as to say Elon plans to eventually merge them with Tesla, too.

If the plan is to make money, I’m not sure this is a good idea. Holdings like Alphabet show that merging several great businesses under one name often depresses individual multiples.

Needless to say, Alphabet’s businesses (Google advertising business, YouTube, Google Cloud, DeepMind, Waymo, Isomorphic Labs, and a long tail of others) would sum up to a much larger market cap if they were separated.

Nonetheless, at one point not long ago (when we invested in Google back in April), all these businesses were trading, combined, at a price-to-earnings multiple of 16, well under the average of the S&P500, a madness that has since been thoroughly corrected, as the holding now trades at 33.

But it may just be a move to keep xAI from sinking. Nobody can deny Elon’s unparalleled capacity to attract capital, but it’s very likely that xAI is seriously struggling from a business point of view; beyond X, nobody actually uses xAI’s Grok models beyond some use of their Grok 4.1 Fast model for coding by developers. It’s something, but most likely not remotely enough to justify xAI being valued in private markets at hundreds of billions.

Putting my cynism aside, I’ve mentioned in the past that, unlike Anthropic or OpenAI, and much like Meta, xAI has a considerable amount of internal demand and use cases from other companies in the Elon network (Tesla car communication, customer service, even as the ‘brains’ of their future Optimus robots, as well as supporting X’s algorithm), which is the likely reason why investors “trust the process.”

While Anthropic and OpenAI need to grow their revenues, xAI may just need to power the new features of established companies to justify its value.

Instead, this is likely just a maneuver from Elon to establish greater control over his entire Empire.

In what represents these founders’ second company they’ve sold to Apple (after PrimeSense in 2013), Apple has acquired Israeli startup Q.ai, positioning it as another move in the intensifying push to build AI capabilities in consumer hardware.

Q.ai’s work centers on applying imaging and machine learning to audio problems, including helping devices interpret whispered speech and improving audio pickup in noisy settings, which could map onto Apple’s recent effort to add more intelligence to products like AirPods and other on-device experiences.

Neither company disclosed the price publicly, but Reuters reports the deal valued Q.ai at about $1.6 billion and that the startup was backed by firms including Kleiner Perkins, Spark Capital, Exor, and GV. Q.ai’s roughly 100 employees, led by CEO Aviad Maizels, will join Apple.

As we reported a few days ago, Apple (and the overall Silicon Valley ecosystem) has decided AI wearables are the next big thing.

Truth be told, unlike more explicit uses of AI like ChatGPT have struggled to really take off in revenues despite having massive adoption, signaling that users struggle to find the value in purchasing these products. Rather, “invisible AI”, as I call it, has a proven track record delivering huge businesses (especially in the area of ad targeting, ask Google or Meta).

Being able to weave AI into the fabric of every one of our daily actions is the kind of use case we are aiming for here, one where the invisible AI makes life easier even if we aren’t aware its magic is at play.

That said, AI hardware has a notoriously bad record, with perhaps only Meta’s RayBan glasses being an example of success. SV doesn’t always get market sentiment right, so I wouldn’t be so quick to assume people are looking forward to having AI pins on their chests.

I will say, though, that the idea of having smart earpods does appeal to me, as I’m a huge earpod user. Making AI subtly present without making it too obvious to people around you seems like the right move. However, if there’s one thing Apple is known for, it’s privacy; how are they going to handle having AirPods listening in on private conversations?

Not only is the cybersecurity concern very real (having truly powerful AIs that can run locally on your AirPods or iPhone is a long way away, if even possible), but I’m not sure I want my friend’s AI system listening to what I have to say to them.

Would you welcome that? I think we both know the answer.

One of my 2026 predictions is that AI Labs, at least one or two, are IPOing in 2026. Our bet was that it would be either Anthropic or Cohere, based on the assumption that their books would look cleaner for an IPO than OpenAI’s, since the latter has such a huge base of free consumers to serve. In other words, our interpretation was that OpenAI’s margins were worse, assuming all else was equal (which, in fact, turned out to be potentially false).

But now, it seems OpenAI is committed to being the first big AI Lab to IPO (excluding xAI, which could merge with SpaceX ahead of the latter’s IPO).

The actual reasons aren’t fully known, but it takes no genius to realize that AI Labs need to access liquidity and do so fast to sustain their businesses.

And although Big Tech is still willing to foot the bill, recent reports suggest that conviction in OpenAI’s success and business model isn’t particularly widespread, especially from NVIDIA, which, despite recent reports that it would invest in OpenAI, is now having second thoughts.

Moreover, I assume there is more than obvious skepticism around OpenAI’s introduction of ads in ChatGPT, a move that should help them monetize their huge consumer base better, but which is not only an unpopular thing to do, it’s particularly tricky.

Ideally, you would want models to factor in paid products as part of its generations, which would be the clearest and most effective form of advertising, but this is not only extremely unpopular (not even considered by OpenAI for now), but very complicated to implement, considering that AI model responses are not totally predictable, which is a huge problem for OpenAI.

Think about it, they sell 1,000 mentions by the model for, say, $100. To get that rate, OpenAI has to estimate how many inferences it will need to get those 1,000 mentions. If it can guarantee the product will be mentioned in 1,000 out of 1,500 responses, that’s probably a very good deal. However, if OpenAI has to infer the model 1,000,000 times to get those 1,000 mentions, OpenAI might actually be losing money on a deal basis due to large inference costs.

Google is transforming Chrome into a fully agentic experience. Not only is Google accessible in the top right of your browser window and “see” what you see in every browser, but it will now be able to browse the Internet for you. The link includes a video example.

This reads and feels very similar to OpenAI’s lackluster Atlas browser, so I wouldn’t be too optimistic about it as a ‘game-changer’. However, it does give credence to the idea that Google has the best chance of winning the AI game, simply because everyone uses Chrome, and adopting Gemini is easy and accessible.

In a world where less than 10% of actual ChatGPT users have tried reasoning models, we can’t underestimate the value of low-effort distribution; the ability for companies to allow users to adopt their tools with ease.

Maybe Chrome was really worth seriously discussing as a potential monopoly. But too late for that, huh?

Finally, after months (years, actually) of this newsletter fangirling about Google’s Genie model family, we finally have a product you can use. And as you can see in the link, the examples are simply mind-blowing, proving that this is the most exciting Generative AI product ever.

I don’t care what others say about coding AIs and whatnot; this, technologically speaking, is the pinnacle of Generative AI. The fact that you can create, on demand, basically any world you request and control it, with the AI model adapting the scenery based on your actions, feels like magic, even for us who study and analyze AI for a living.

In simple terms, Genie is a video world model that takes in a text description (or an image) of what you want to generate and generates a 3D world, first verifying the upcoming world with a first frame to ensure it is to your liking. But what makes this unique is that the world is interactive, letting you move the character and perform actions in the world you’ve just created.

We only know how the first Genie model (back two years ago) was created, but we can use that guidance for how DeepMind trained this one. In short, it’s a latent-action model trained to take a concatenation of frames and predict the actions that carry a character from one frame to the next (e.g., a character that is still shown jumping in the next frame, so the model predicts a jumping action).

However, this is easier said than done, as the model also has to predict environmental effects (underlying physics, uncontrollable movements beyond the character’s control, etc.), to the point that it feels more like a physics simulator than an AI model.

If you’re lucky enough to be in the US (I missed this release for only a few days) and a Gemini Ultra subscriber, you can use this model. If so, let me know how the experience went responding to this newsletter!