Welcome back! This week, we have a lot to discuss. The last few days were probably the most active for top model releases from the US in perhaps a year, with two Labs that had fallen from grace somehow making a comeback.

We also discuss really cool tech from Google in SensorFM, a robotics hand that is the stuff of nightmares, a Databricks research showing that AI is much more than the model itself, and more.

Enjoy!

In one of the most heated weeks so far, with multiple incredible model releases, my favorite news has been a new paper from Google and its model SensorFM.

This model has been trained on data from 5 million people and has learned a general-purpose representation of human physiology that transfers across 35 health prediction tasks.

Ok, but what does that even mean? Well, it means I’m excited now because it opens the path to personalized health. Let me explain why.

Currently, billions of wearables track sensory signals from the wrists of millions of people. However, these devices are extremely noisy, and translating that sensor data into meaningful health insights is debatable at best.

Each company measures differently, each device registers the data differently, and, importantly, each human is different. This means that, in reality, the signal-to-noise ratio is very low.

But what if we trained an AI to process the data across millions of humans? That is what Google did, training the AI on one trillion minutes of data, or roughly 2 million years, and the results are pretty incredible.

But first: how do we train a model with that data?

If you read my last newsletter, which I recommend you do, I explained that an AI can only learn what can be measured. And by "measure," I mean models learn by making predictions and comparing them to a ground truth (what they should actually have predicted).

For these one trillion minutes, we don’t have a ground truth (officially known as ‘label’); we don’t know what each of these one trillion minutes is showing. Depression? Anxiety?

Ideally, you would want a dataset that describes the following: if sensor data looks like this, the human has heart problems. If the data looks like this, the human is fine. Eventually, the AI makes the connection and learns to associate certain data patterns with certain outcomes.

But if there’s no way to measure the outcome, how are they supposed to learn? And the answer is unsupervised learning, which I did not include in an already 8,000-long piece, but it’s essentially a rare type of training where the data itself is the signal.

Essentially, we’re telling the AI: “We can’t really associate the patterns you’re going to find with particular outcomes. However, we still want you to find them.”

But why would we want this? For things like clustering. Although the AI doesn’t fully understand the implications of each pattern, it can still classify them.

If we give an AI a huge unlabeled dataset of flowers, the model might know what each flower actually is—it might not even know what a flower is—but it learns to classify them nonetheless; it will still learn to separate roses from carnations even if it doesn’t know what those are.

Here, they’ve done the same with wearable sensor data.

Okay, got it. At this point, we have a great-but-useless model. We have a global representation of sensor data, a model that receives sensorial input and captures key patterns. However, we have no way to decode those patterns, to associate them with actual outcomes.

For that, they add extra training phases, but ones focused on selected, well-labeled data from real humans with real metabolic, mental, and sleep-related signals, basically saying: “this is the sensorial data for this human, and here’s how they actually feel”.

With that, we can now associate the pattern with the outcome!

Interestingly, for these new training phases, they had AIs design them, using the idea of an ‘LLM classroom’. They would give the AI the training constraints, and the groups of LLMs would design the experiments, leading to training regimes that beat human-designed experiments by a long shot.

And the results are pretty good, I have to say.

For starters, this universal representation, this one-size-fits-all model, can capture dependencies across 35 health domains and shows great prediction accuracy potential across many of them:

Importantly, the AI's potential struggles in some areas are not the model’s fault per se; they may simply reflect what many already believe to be true: wearable devices produce very noisy data.

But perhaps a more fascinating outcome was testing whether SensorFM could serve as a good context engineer for an LLM acting as a health agent.

In other words, the Gemini agent would receive predictions from the SensorFM about the user in particular based on sensor data; things like age, predicted BMI, or anxiety scores (see below for an example), and the Gemini agent not only provided much more meaningful recommendations, but using the actual ground truth data from the user (actual age, BMI, or insulitn resistance) did not improve recommendation quality relative to using SensorFM.

I’m a staunch believer that healthcare is one of the key domains where AI will change our world the most.

The option of offering personalized health recommendations to every human at scale, something our current system can’t for the life of it offer (healthcare systems around the world are completely broken), is something society can’t possibly allow not to happen.

1X, a US robotics company, has announced its new robotic hand. And let me tell you, it’s genuinely incredible. This new hand has 25 degrees of freedom (meaning it can independently control 25 different joint movements), making it incredibly dexterous and gentle.

As shown in this video, the robot can pick up glass and grapes without breaking them, play video games, and more.

They now claim Neo (the robot) can perform any task with its hands that a human can, and that the robot is also strong (enough to pick up weights).

Moreover, the hand’s skin serves as the sensor channel, allowing the robot to measure force and even detect when an object is slipping. Truly alien technology.

I’m happy to see Western robotics labs putting focus on the hardware, not just on the “brains.” This is the most common approach in China, with examples like Unitree, because many people believe (and so do I) that hardware is way harder than software in robotics.

Of course, the big question is how long it will take to transition from marketing art to actual usable products. My opinion is that we’re yet to hit robotics ‘ChatGPT moment’ and that we must remain optimistic but realistic about timelines.

I know we were recently talking about a robot coming this fall, but I won’t believe it until I see it.

Databricks has published some results on agents that may surprise some people.

As shown in the graph above, the choice of harness (e.g., Claude Code from Anthropic, a third-party option called Pi, and Codex, OpenAI’s harness) not only significantly impacts performance on the same underlying models but can also considerably elevate the performance of inferior models.

The best example is GLM-5.2 using the Pi harness, which beats Opus 4.8 using Anthropic's own harness, Claude Code. This result might be surprising, but the interesting insight for me is that it's perfectly reasonable that third-party harnesses beat the harnesses of the model Labs because the latter group has misaligned incentives; they want to build a good harness for you but, at the same time, they want their harnesses to push models into consuming as many tokens as possible, so it's reasonable to assume that if you're running your AIs on harnesses built by the same Lab that created the model, you're going to pay more.

This is palpable in the image below, where the third-party harness running OpenAI and Anthropic models required between 2 and 4 times fewer tokens than those same models running in the "official" harnesses.

In short, you have to be really careful about what system you choose, not just the underlying models, as it can really impact your wallet.

An interesting point is that the Databricks team proves something most people struggle to understand: pricier models aren't necessarily more expensive overall, because larger models (especially the new ones like Fable or GPT-5.6) require far fewer tokens.

For example, they show that Sonnet 5, despite being 1.7x cheaper per token than Opus 4.8, was more expensive overall because it required 1.9x as many tokens for the task.

This should remind companies to decide which model to use based not just on token price, but also to test models for overall expenditure (token price x number of tokens used) before deciding.

And to be clear, the results from running the Pi harness are obtained using the exact same underlying model as Codex/Claude Code, so this is a pure harness-cost difference.

To me, the immediate question an investor must ask is: if harnesses play such a vital role in both performance and cost, and third parties can beat Labs at their own harness game, where's the moat?

I have my thoughts on what I believe is the real moat in AI. To me, the answer is something like what Anthropic is trying to do with Claude Tag, a product that lets users in your company interact with Claude via Slack. People are taking it as just another AI product these days, but that's, in fact, incredibly wrong.

This product is a Trojan Horse that will "marry" you to Anthropic forever. The reason is that Claude Tag is not just “another harness”. It’s a system comprising the harness and a model trained on Slack's synthetic environments.

The point here is that the underlying model is purpose-built for Slack, so it offers performance that is basically unmatched. Furthermore, it is built to become entrenched in your organization, parsing, processing, and learning from your data.

As I said once, there’s a trade-off between privacy and personalization. If you want a truly personalized AI, you need it to see your secrets. That is why it would be hilariously naive to build your personalized AI agents on third-party software.

For these companies, it’s a combination of personalized AI training in your environments, a strong harness, and, especially, personalization by knowing everything about you, that creates the moat.

It’s not the AI per se; the real moat is the entire product.

This interesting exhibit from Goldman Sachs shows the price changes across the semiconductor chain, which can explain vendor pricing power.

As you can see, the price hikes in memory and fast storage (DRAM and NAND, respectively) are something to behold, with GS seeing considerable supply tightness at least throughout H1 2027, although the companies in those sectors, mostly the Big 3 (SK Hynix, Samsung, and Micron), believe the supply tightness could remain very strong through 2028.

Beyond memory, you could argue almost everything is tight.

A note of caution: this can be easily misinterpreted as a “bottleneck ranking” to suggest which companies have the highest pricing power.

And although supply tightness and pricing power are indeed deeply correlated, some companies may simply decide not to raise prices even if they could.

For example, companies like TSMC and ASML are known to be particularly unwilling to raise prices, even though they could easily do so (they are basically monopolies in their respective markets).

In fact, I maintain that the biggest bottleneck in semis is not memory, but advanced packaging, the part of the process that packs the compute and memory chips into a single package, giving you the actual accelerator (e.g., GPUs, TPUs, etc.), which is largely dominated by TSMC.

OpenAI has finally released its newest models, Sol, the strongest, Terra, and Luna, the smaller one.

The most remarkable thing seems to be its much-improved cost-per-intelligence, being considerably cheaper than previous OpenAI models and especially compared to Anthropic’s models (take the performance scores with a pinch of salt). And as we’ll see below in the Grok 4.5 article, OpenAI’s models are now all three on the “Pareto frontier,” meaning they offer the best performance per cost amongst all models, at their respective sizes.

But without a doubt, the most impressive part of the release was that, according to OpenAI, Luna was trained solely by Sol. In layman’s terms, the smaller model was built by the larger model, with zero human participation and in a zero-shot fashion.

In fact, they shared a snippet of the prompt they gave the larger model, trying to convey the idea that RSI (Recursive Self-Improvement) or AIs helping or autonomously creating better AIs, is starting to happen.

The real breakthrough will happen once AIs are capable of training better AIs, not smaller versions of themselves, because in reality, although still impressive, this is just an AI model running the training run of another model; the tests, experiments, and all.

I insist it’s really cool, but in reality, it’s just an AI running a bunch of scripts.

Considering Fable’s guardrails and all, these are quite possibly the best models you can use today. Personally, after several days of using them, I would argue that GPT-5.6 Terra in high mode is the best bang for your buck.

However, overall, I haven’t felt much of a change. This could be the first indication that I might be reaching my personal ceiling of capabilities, meaning, yes, for some stuff like coding, these models feel genuinely better, but for many tasks I use AIs for, like discussing abstract stuff about papers, critiquing my own opinions, and the lot, I barely notice a difference with the previous generation.

Is that a sign that these models are becoming so good that we can’t notice improvements anymore, or are they hitting a wall?

I fear it might be a mixture of both. I don’t mind these AIs becoming incredibly smart, but what worries me is that, for most economically valuable tasks, incremental improvements won’t improve results and will simply make your bill larger.

Could it be that AIs are becoming too good for their own good, meaning they are too expensive relative to what they offer, because what they offer is simply overqualified for most white-collar work?

I recently listened to a podcast where the guest basically discredited the idea of AIs destroying jobs not because they could, but because he argued most jobs were “made up” and there was nothing of economic value to disrupt.

And while I’m not sure I would put it that way, after almost a decade as a consultant for corporations, I have come across my fair share of jobs that literally produced zero value despite those people earning 6 figures or more.

It’s going to be interesting to see how this industry handles the slight possibility that most economic activity is non-disruptible. A potential outcome is that AIs open the door to new types of jobs and economically valuable activities that do not exist today.

Who knows, but I remain skeptical about how much it can disrupt what exists today.

Notion, a note-taking app, has announced an agent-orchestrator product called ShipOS, a way to organize your agents from multiple sources (Codex, Claude Code, etc.) into a single Kanban board where you can assess progress and make better decisions.

Basically, an agent manager software, something very similar to what Linear already offers.

The reason I’m pointing this release out to you is that I recommend you watch the presentation video, as I believe it clearly shows what the future of managing agents might look like.

Listen, I use coding agents daily, and I continuously rotate from “this is the best thing ever” to “I really, really hate them” because there’s too much information being thrown at you, agents lose the plot… and many other concerns that tell me the current form factor is clearly suboptimal.

Assuming looking at the code is almost impossible considering the rate at which these models generate code, we need to find an abstraction layer that allows us to see the things that have to be seen, even some code snippets in particular, and that way make good decisions.

Right now, I feel like most of the time I’m just telling it to do whatever it suggests, because I really don’t have the time to judge every decision it makes among the thousands of agents' decisions made each day.

Sadly, I know what I’m doing isn't how software should be built; you need more control, you need to “own” your work, but I don’t think the current status quo lets you do that without scrutinizing every single thing these agents throw back at you, which is impossible.

Whether it’s Linear or Notion, these companies have a golden opportunity to own that abstraction layer, but I don’t think anyone really knows for sure what it will look like in six months from now. But this is a start.

As reported by Reuters and Meta, Meta has released Muse Spark 1.1, a multimodal AI model designed for coding, debugging, and complex agentic tasks involving external tools and multiple steps. It is now available to US developers through the new Meta Model API in public preview.

Besides the fact that the model looks very competitive in terms of raw performance, the real highlight is performance per cost: API pricing starts at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in introductory credits.

Not only is this a highlight in itself, given it’s the first time Meta will serve its models via APIs, but the prices are also considerably lower than at the frontier. Muse Spark’s prices are even lower than the API price of the best Chinese model, GLM-5.2.

In fact, Meta’s prices are so outrageously low relative to peers that they are the cheapest option available right now, including some Chinese Labs like Zhipu and its GLM-5.2, with only DeepSeek offering much more competitive pricing.

Also, much as we saw with the Databricks evals, a lower token price does not necessarily mean a cheaper overall task cost.

As you can see above, while GPT-5.6 Luna is 27% more expensive per token, the overall cost per task on AA is lower because it uses fewer tokens. I suspect enterprises will soon transition from measuring costs at the token level to the task level (i.e., what's the cheapest model for my task?) instead of measuring this abstract idea of a ‘token’.

Grok 4.5 was launched on July 8 as SpaceXAI’s latest model for coding, autonomous-agent tasks, and technical knowledge work.

The company says it was trained alongside Cursor, the team of talented engineers SpaceX paid $60 billion for a couple of months ago. Training used tens of thousands of NVIDIA GB300 GPUs, which means it’s a large, large model.

SpaceXAI reports that Grok 4.5 scored 62% on DeepSWE 1.0, 29% on SWE Marathon, 83.3% on Terminal Bench 2.1, and 64.7% on SWE Bench Pro, signaling very strong performance on coding and agents.

The model is served at about 80 tokens per second, which is reasonably fast by today’s inference standards. It’s also considerably token-efficient.

On SWE Bench Pro, it generated an average of 15,954 output tokens per task, compared with 67,020 for Anthropic’s Opus 4.8 in SpaceXAI’s comparison. The company says this makes Grok 4.5 roughly twice as token-efficient as comparable models overall.

They don’t seem to be lying, considering that Artificial Analysis places this model in the green quadrant alongside Meta’s Muse Spark and OpenAI’s GPT-5.6 Luna as the most cost-efficient models out there right now.

API pricing is $2 per million input tokens and $6 per million output tokens, making it very competitive, as you can see in the table in the Muse Spark news above.

Very solid release that suggests SpaceXAI is back on track after several months of falling very behind. It’s clear that the Cursor acquisition has benefited them a lot.

Overall, based on this release and that of Meta’s, it seems we are amidst a new price war on token prices.

However, this time the price pressures are not coming from China but from within the US, specifically from Meta and SpaceXAI's new models.

As discussed above, Meta's latest model, Muse Spark 1.1, is comparable in capability to Opus 4.8 and GPT-5.5 (though not quite at the Fable/GPT-5.6 level), but its API prices undercut even China's GLM-5.2. Grok 4.5 is in the same performance ballpark, only slightly pricier per token but offering slightly higher performance.

Most fascinatingly, it seems the tables have turned in terms of performance per task relative to China. OpenAI is now very comfortably leading the Pareto frontier with its three new models (Luna, Terra, and Sol), and the green quadrant is now dominated by three US models. How the turn tables.