Welcome back! This week, we take a look at new models and product releases, especially Gemini 3.1 Pro; new billion-dollar investment rounds; new hardware that spits tokens at 16,000 per second; the curious case of Korean stocks, crucial to AI but “ignored” by the rest of the world; and the latest example of a company that might try to use AI to bring people from the dead.

Enjoy!

You may wonder, how does a new AI Lab get $1 billion in funding in 2026? Isn’t the game already decided between OpenAI, Anthropic, DeepMind, etc.?

It turns out, some researchers think not (and I agree). The man behind this new AI Lab, Ineffable Intelligence, is none other than David Silver, one of the key figures behind Google DeepMind’s success. Put potentially, and I’m only speculating here, the reason he got so much money has more to do with the approach he’ll follow.

Nothing reveals more about what he’s trying to do than his 'Welcome to the Era of Experience’ paper, which he released last year with Rich Sutton, one of the founding fathers of AI (creator of the Reinforcement Learning paradigm back in the 1980s, which is all the rage today).

And what does this paper suggest?

Currently, AIs are episodic engines, models that live in the current interaction, but do not carry over to future interactions what they learned from this one.

Careful, current models can bring context from past conversations to new ones, but that’s not the same thing; it’s as if you carried over a diary of past interactions to consult on every new interaction you have.

Instead, real experience is about learning from past experiences; it’s not about keeping a cheating diary, but about having your brain learn from that feedback and perform better on new ones without having to recall every single instance from the past.

Which is to say, what Silver seems to be proposing is an experience feedback loop: act → receive feedback → optimize → act again. This is a fundamentally different behavior from what ChatGPT or Claude represent today, which is more like act → copy into context → act. The latter also involves learning between interactions, but this learning is ephemeral and forgotten once it is out of context.

Crucially, this ability of models to learn from experience might allow them to be smaller without having to retrieve solutions from memory (one of the reasons larger models are better is that they store and can recall more stuff).

In a sense, it’s more about having an AI model that doens’t have to know every single fact in human history, but can suggest possible good solutions, test them, measure feedback, optimize, and repeat, similar to how humans work.

Markets are not rational. We both know that. But one of the prime examples of this irrationality is the Korean discount: companies in the Kospi index, such as Samsung and SK Hynix, trade at a considerable discount relative to similar companies in other markets.

Look at the four companies below, two Korean, two from the US:

These four companies do basically the exact same thing: memory and storage (Western Digital only does storage).

Samsung and SK Hynix are not only more profitable than the other two but also have larger market shares. Yet the other two are in the US index, which, as you can see in the thumbnail above, has an average multiple 2.5x higher than the Korean index's forward price-to-earnings.

The main reason for this is liquidity, and ironically, a self-inflicted wound from Korean regulators, who make it quite hard to invest in Korean stocks.

So with these stocks, we could havea combination of two things that could send them on their way up:

The first one is self-explanatory, but the second one requires some points to mention. Currently, the sentiment around memory stocks is that their business is cyclical; good in healthy times, bad in downturns.

When that happens, you can’t just apply a multiple to earnings using methods like DCF (Discounted Cash Flow) because that requires looking into the future and assuming future demand is predictable.

Instead, investors value these stocks relative to their book value (total assets - total liabilities), i.e., what the company actually owns net of liabilities, since current earnings are not indicative of future ones.

Naturally, multiples over book value are smaller. But if demand indeed becomes predictable, or still cyclical but from a much larger demand base, we have two companies that:

This is all, of course, a hypothesis; nobody actually knows or can predict how markets will behave, but what nobody can explain to me is why Micron has a higher multiple (and valuation) than Hynix despite the latter being better literally everywhere.

A new start-up, Taalas, has come out of stealth with what is quite possibly one of the most impressive showings I’ve seen. This new hardware company promises (and delivers) up to 16,000 tokens per second, per user.

In fact, you don’t have to take their word or mine; you can literally try it for yourself here. It’s not a joke, it works.

The idea is to create custom silicon for each model, literally etching the model’s architecture into the chip (we’ve already seen something similar being proposed by Etched and their Sohu chip). This, combined with no off-chip memory, means that the data never actually leaves the compute chip, making it much faster to process

Additionally, this requires much less power (about 10 times) and can be built much more cheaply (20 times), or so they claim. Specifically, they mention each Taalas server requiring just 2.5 kW (approximately the same power as your blow dryer), while an NVIDIA Blackwell Ultra requires ~145kW, 58 times more, meaning we might have found a very cost-effective and power-efficient way of deploying AI.

And it doesn’t seem far-fetched considering that current GPUs and other top AI systems require a lot of power to move data in and out of the compute chip.

Intuitively, you can think of this as doing the same exercise but more cleanly, with no wasted power at all, as the chip is purposely designed for that model and that model only.

The results are incredible; I’ve never seen a model run this fast. Of course, we must take this with a pinch of salt, as we need more data, and most importantly, more proof that this can be scaled.

One big concern is custom silicon: the model frontier changes rapidly, to the point that by the time your hardware is out, your model might be obsolete. They mention LoRA adapters in the blog, suggesting they are designing the platform to support model adaptation, but this remains in question.

All very early to draw more conclusions, but I can’t deny this looks promising and, if it scales, may put into serious question how much money we’re spending on chips.

The start-up Zyphra has announced ZUNA, an AI model designed to work with EEG brain-signal recordings (scalp sensors). It’s described as an early step toward “thought-to-text,” but the main focus right now is making EEG data more usable.

ZUNA is meant to handle real-world EEG issues: it can clean noisy signals, fill in missing channels, and estimate signals at different sensor positions if you provide the sensor coordinates. Zyphra argues this helps because EEG datasets are small and inconsistent, so a broadly trained model could transfer better across devices and studies.

In simple terms, it’s an AI that takes in EEG data, cleans it, and returns a better datapoint. But how is that remotely useful? Well, actually, it’s very useful.

If we recall last week’s discussion on Google’s new bioacoustic model, Perch 2.0, we touched on the idea that the model, despite having been trained mostly on bird sounds, was still capable of identifying whale sounds and even cluster different whale species despite being a new species to them.

The reason behind that “superpower” was that Perch is a bioacoustic foundation model, one that has seen vast amounts of bioacoustic data, to the point that even sounds from “unknown” species are familiar. That’s why it’s called a foundation model; it serves as a foundation for several downstream applications in the field of expertise.

EEGs are important across several fields:

Therefore, having a model that can serve as the “foundation” for all these downstream tasks is extremely valuable. And we’ve just taken the first step in that direction.

I don’t normally say this, but The Information actually published something that made me think (they are usually just good at reporting what they hear). This semi-opinionated column on world models is actually very good.

World models are one of the hottest, if not the hottest, avenues of research in AI today, with examples like Google’s Project Genie taking the industry and media by storm.

And this article by The Information tackles the usually-ignored problem with world model sequentiality: the notion that world models can’t be parallelized as well as normal models.

As you probably know, AI runs mainly on GPUs and other ‘accelerator’ hardware. Here, accelerator is the marketing term for a parallelizer, a hardware that speeds up computation by parallelizing it; taking a workload that would take ‘x’ time to execute, dividing it into ‘n’ parts, and thus taking roughly ‘x/n’ time to process it by executing all ‘n’ parts in parallel (there’s always a portion of the workload that can’t be fully paralellized).

The problem is that the workload must be parallelizable in the first place. In layman’s terms, each part has to be unrelated.

This works great with standard Large Language Models (LLMs), as they can serve responses to different users because their responses are completely unrelated. As the bulk of the workload, the actual model, is common across all users, you can use a single model instance to serve them.

With world models is not that easy. World models are AI models that “carry a state”, meaning the AI model has to take in feedback from the environment (the “world” it’s located in), see how the model’s actions impact the environment, and then use both the model’s past actions and the new environment state to make the next action.

It’s no longer just what the model did in the past, but also how it has affected a changing variable, the environment around it.

A perfect example of this is humanoid robots. If tasked with cleaning the dishes, the robot may clean a plate, changing its state from dirty to clean, and it must learn to recognize when the plate is clean enough.

But a world model also needs to be able to identify that the plate is slipping in the robot’s hand and that it’s ceramic, which will likely break if it hits the counter.

Hence, a world model has to factor all this in to make its next move (instead of continuing to clean, stop when it’s clean enough, and also tighten the grip because its robotic hand sensors are detecting the plate slipping).

All that information is unique to that particular user and environment, so it’s not parallelizable across users.

This large payload of “user-only” data increases the workload's overall sequential nature (i.e., subsequent users have to wait until this user’s interactions finish). In practice, that is not feasible at all, so major companies have to take the hard measure of assigning GPUs to individual users.

This means each $40k GPU, designed to parallelize across potentially hundreds of users, is being used on a sequential, batch 1 workload, an absolutely terrifying waste of money.

In practice, this means world model labs have to charge users huge sums (as Claude’s fast mode charges up to $150/million tokens because it serves GPUs with tiny batches), making this a game for the uber-rich.

To me, this means two things:

World Labs, an AI start-up founded by Fei-Fei Li, an AI pioneer who teaches at Stanford, has announced a huge $1 billion funding round to continue pursuing the creation of spatial intelligence.

In their view, current AIs lack spatial intelligence, the ability to visualize, manipulate, and navigate the 3D world, even the video ones. Thus, the idea is to train AI models to learn such a core feature of intelligence.

Their core product today is Marble, which lets you create 3D worlds from text, images, or video and immerse yourself in them. Interestingly, you can test Marble for free here; it’s an interesting experience because you can give it images of places close to your heart and immerse yourself in them.

It kind of feels similar to what Google is doing with Project Genie, but without Genie's character interactivity, and is more focused on generating high-fidelity scenes.

To me, this feels kind of a lesser version of Genie, in that it pursues the same idea, being able to create 3D worlds to train AIs in them, but without the quality and interactivity that make Genie what perhaps is the most impressive AI model in the world.

That said, World Labs’ research has a place in AI, as it’s much more likely that these models learn good spatial intelligence than models like ChatGPT, which were conceived in the realm of text.