Tag: Autonomous Science

Weekly Technology Brief #3

Intelligence Is Escaping the Data Center, WHAAAAT?

22 October 2026

Dear readers,

There was a time, not very long ago, when talking about Artificial Intelligence mostly meant talking about increasingly large models running inside increasingly large data centers. That description is rapidly becoming rather inadequate.

This week, AI appeared inside proactive work agents, tiny edge computers, scientific research systems, robots and even the electrical architecture feeding AI servers themselves.

Oh! And, another interesting thing happened: The word model seemed slightly less important, while other words such as system, platform, workflow, toolchain, power and deployment appeared rather more often.

An accident, perhaps? Me thinks not!

The next stage of AI may be less about putting a bigger brain in one enormous computer and more about getting different kinds of intelligence into the places where useful work actually happens. When you think about it long enough it does sound logical but I think it was not obvious at all that this was going to become the trend so soon.

1. Foundation Models Are Beginning to Escape Language

Mistral released Timestral on 15 October, its first family of foundation models specifically designed for time-series forecasting. The company says the largest model contains only 209 million parameters and performs strongly on several public forecasting benchmarks while handling multivariate and probabilistic predictions.

Now, I am deliberately not going to get too excited about benchmark claims made by the company releasing the model. We have been down that road before. But still, the direction is interesting.

The phrase “foundation model” became almost synonymous with enormous language models because text was where the modern AI boom first became visible.

But as you may have noticed, the world is not composed entirely of text. For example, factories generate sensor streams, hospitals generate physiological measurements, electric grids generate demand patterns, financial systems generate sequences of transactions, and, perhaps rather more interestingly, robots generate continuous streams of physical observations (I just find this whole robotics stuff super exciting).

Anyway, in all these cases, time matters.

If the foundation-model idea transfers effectively into time-series data, we could end up with reusable pretrained models for forecasting and anomaly detection across many industries. And perhaps more importantly, Timestral is another reminder that bigger is not always the same as better.

A 209-million-parameter model optimized for the right problem may be considerably more useful than an enormous general-purpose model being asked to do something it was never really designed for.

Call me old-fashioned but I think specialization still matters!

2. Anthropic Wants an Agent That Notices Work Before You Ask

On 21 October, Anthropic introduced Claude Duet, a research-preview application designed to work proactively: Identifying things requiring attention, continuing tasks while the user is away and checking before acting on the user’s behalf. Anthropic itself warns that this early version can misread conversations or incorrectly judge what is urgent. Oh boy, have we not seen that before?

There is something interesting here through:

We have spent years training people to ask AI systems questions. Now the emerging model is:

The AI notices something → decides it matters → brings it to you → potentially acts

That is a very different relationship: A reactive assistant waits, but a proactive agent observes.

And you know about observation, right?

Observation immediately introduces questions of access, privacy, judgment and authority.

So, here we go again: What information is the agent allowed to watch? How does it decide what deserves attention? When can it continue without asking? When must it stop? And, perhaps most importantly, how easy is it for the human (hopefully one in the loop) to understand why it decided to do something?

Anthropic’s decision to emphasize confirmation before action is therefore not merely a user-interface detail. It reflects a deeper truth about agentic AI.

As autonomy increases, permission architecture becomes part of intelligence architecture. There is little value in building an extraordinarily capable agent if nobody sensible is prepared to give it access to anything important.

Hmmm. Let’s watch this space closely!

3. Edge AI Is Starting to Acquire Its Own Computing Architecture

On 20 October, Synaptics introduced its Astra SR200 family, describing it as a new class of microcontroller capable of running multimodal AI locally. At the same time, Synaptics and Google Research expanded the Coralboard ecosystem, combining embedded AI hardware with an open-source compiler and developer platform aimed at low-power devices.

This one deserves attention because the cloud is not necessarily the natural home for every AI workload. Ater all, sending everything to a distant data center creates latency, uses network bandwidth, and, of course, can introduce privacy issues.

And on top of that, a robot, camera, industrial machine or wearable device may occasionally have the inconvenient requirement of continuing to work when the internet does not. It is, therefore, inevitable that some intelligence eventually moves closer to the device.

But once again, the interesting part is not simply the processor. Nope, it is the toolchain.

Hardware becomes much more valuable if developers can take models they already understand and deploy them without rewriting half their software every time they change chips.

This suggests another possible bottleneck:

Model availability → Compute → Memory → Power → Developer tools

Solving the hardware without solving the software does not create an ecosystem. And if edge AI really does spread through cameras, appliances, wearables, industrial equipment and robots, we may eventually stop thinking of “AI hardware” as a special category at all. I mean, it will simply be hardware!

Think of it this way: If it is just hardware, nobody will get particularly excited about it anymore, just as today we don’t get overjoyed because a washing machine contains a microcontroller😊.

Maybe that transition itself is a sign of success!

4. Robotics Is Beginning to Build a Standard Development Stack

NVIDIA used its GTC Berlin event on 21 October to expand the Isaac robotics platform, adding a new runtime, robot-learning tools, datasets, simulation workflows and a reference humanoid platform developed with Agile Robots. The Agile ONE SN system combines NVIDIA’s GR00T robotics software with Jetson Thor hardware and is intended as an integrated platform for developing and validating humanoid capabilities.

Here again, the robot itself may not be the most interesting part. There are already plenty of humanoid robots available for impressive videos (walking, folding things, blah, blah).

The harder problem is turning robotics into a repeatable engineering discipline.

But, for that to happen, developers need simulation, training data, clever control systems, safety mechanisms, hardware interfaces, evaluation methods, and, oh, a little thing called deployment tools.

And not to forget the need for finding out ways of transferring something learned in simulation into an unpredictable physical environment. Easy, heh?

NVIDIA is effectively trying to provide much of that surrounding infrastructure. That matters because industries tend to accelerate when developers stop rebuilding the plumbing every time that they start a project. At least, that is my impression.

All said, the important scoreboard for robotics remains stubbornly boring: How reliably can the robot complete useful tasks, how often does a human intervene, how safely does it fail …

Admittedly, those questions are less exciting than humanoid acrobatics but they are also much more important.

5. Thousands of AI Agents Are Being Pointed at Biological Data

On 21 October, OpenAI, PerturbAI and Scripps Research described an experiment using thousands of AI agents in parallel to analyze data from millions of brain cells, connect observations with scientific literature and generate hypotheses for researchers to investigate. PerturbAI reports that the system analyzed its whole-brain dataset and synthesized findings in roughly 32 minutes.

This sounds remarkable. It also requires a giant warning label! The result was research leads, not validated biological discoveries. Acknowledging the difference is critical.

No matter how quickly thousands of agents produce hypotheses, biology ultimately has the annoying habit of demanding experiments.

Having said that, the architecture is fascinating: Rather than asking one enormous AI model one big question, the problem is divided among many agents working simultaneously on different pieces of evidence before their conclusions are combined, which begins to resemble distributed scientific labor:

Experiment → Huge Dataset → Many AI Investigators → Candidate Connections → Human Review → New Experiment

If this works reliably, the productivity gain could be substantial because modern biological experiments can generate more data than human researchers can realistically inspect in detail.

But the next bottleneck immediately appears. Where? Hmm. If machines can generate hypotheses much faster, can laboratories test them fast enough?

We may end up with an abundance of plausible scientific ideas and a shortage of experimental capacity. So, age of abundance isn’t the same as age of no problems or, in this case, no bottlenecks. I mean, build this one, and that one becomes the next bottleneck😊.

6. AI’s Electricity Problem Has Reached the Inside of the Server Rack

Perhaps the least glamorous announcement of the week was also one of the most revealing.

Infineon and NVIDIA said on 21 October that they are expanding work on 800-volt direct-current power architectures for high-density AI data centers, using gallium nitride power devices for critical conversion stages inside the rack. Infineon says higher-voltage distribution is being introduced as AI racks approach increasingly extreme power densities.

Think about what this tells us:

A few years ago, most public discussion about AI infrastructure revolved around GPUs, until memory became important. Then it was networking’s turn. Then we talked about data center capacity, and then electricity generation.

And now we are discussing the voltage at which electricity should be distributed inside the rack.

I suppose this is what technological scale looks like. Once a system becomes large enough, previously boring engineering details become strategic constraints. Maybe I should have picked a different career. Who would have thought voltage would one day becoming a thing.

Anyway, higher voltage can reduce distribution losses, but it creates another set of requirements around power conversion, safety, component design and thermal management (and maybe a few more that I can’t think of right now). Infineon argues that gallium nitride is particularly useful because it can switch rapidly while reducing conversion losses and component size.

Whether 800 VDC becomes the dominant architecture remains to be seen. While we wait to see where this goes, we note that the broader point is, probably, harder to dispute:

AI is now sufficiently physically demanding that its future depends increasingly on electrical engineering. One could even say that software has finally discovered electricity!

Welcome to the real world folks…

The Connection: Intelligence Is Becoming Embedded

Once again, the highlighted developments this week look very different.

  • A forecasting model,
  • A proactive office agent,
  • An embedded processor,
  • A humanoid development platform,
  • Thousands of scientific agents, and
  • A new electrical architecture for AI racks.

Yet, surprise, surprise, they share an important theme: AI is moving from a single centralized activity toward something embedded inside other systems.

“What system?”, I hear you ask. Well, it seems to me that it is going:

  • Inside business processes,
  • Inside physical devices,
  • Inside scientific workflows,
  • Inside robots, and
  • Inside industrial infrastructure.

Since it is clear by now that I like linking keywords with arrows to make a high-level observation, let’s say the evolution we are witnessing might look something like this:

Model → Platform → Workflow → Device → Physical System

And, as it may be equally probably clear by now, at every transition, a new engineering problem appears (I said it in my introduction to these reports that I am not giving investment advice to my readers but hey, a bottleneck is not just a bottleneck, right? Remember where you read this first😊).

Anyway, more capable agents require better controls, just as Edge AI requires new chips and toolchains.

And physical AI requires simulation and safety, just as AI science requires experimental validation.

And apparently, dense AI computing requires entirely new approaches to power delivery. That is obvious when I read it a few times more but I did not see that coming, to be quite frank (I hope you are learning while reading these reports as I am while reading and preparing them).

This is why I continue to believe that asking only which model is currently “best” tells us progressively less about where technology is actually going. That ship has sailed.

The model may provide the intelligence, but the surrounding system determines whether that intelligence becomes useful.

What I Am Watching Next

I want to see whether proactive agents can operate for long periods without requiring so much supervision that the supposed productivity gain disappears.

I am also keen on learning whether edge-AI platforms attract real developer ecosystems to become a real ‘meaty’ thing.

And in scientific AI, I want to see the one thing that matters most: Whether machine-generated hypotheses survive experimental testing, or just end up producing more noise that is hard to look at.

In essence, the big question remains which “demonstrated” technology survives contact with reality!

Sources & Further Reading