Tag: Energy Infrastructure

Weekly Technology Brief #5

Intelligence Is Getting Cheaper. Reality Isn’t!

5 November 2026

Dear readers,

Last week, we explored a recurring pattern in technology: As the underlying invention matures, the difficult problem moves from making the component work to making the entire system work reliably.

This week adds an interesting complication so, oops is in order!

AI models are becoming more capable, increasingly specialized, and more deeply embedded in ordinary software. Robots are learning to respond to touch rather than relying entirely on sight (oh my!). Meanwhile, companies are discovering that building the infrastructure required to support these advances involves rather stubborn things. You know, things like electricity, construction schedules, financing, and occasionally the objections of people who live nearby. Who would have thought of these, eh?

Anyway, what this means is that the digital world is advancing at one speed, while the physical and economic worlds are advancing at another. Hmm!

The following four developments help explain why that distinction matters so, buckle up…

1. Robots Are Learning that Seeing Is Not the Same as Feeling!

On 5 November, Panasonic and researchers from several universities announced T-Rex, a robotic manipulation system combining vision, language instructions and real-time fingertip feedback.

The announcement may sound small until we consider what robots actually struggle with. Recognizing an egg is relatively straightforward. Picking it up without crushing it, transferring it between hands and adjusting grip when it begins to slip is considerably harder.

Traditional vision-language-action models largely decide what to do based on camera images and instructions. T-Rex adds a faster feedback mechanism that responds to physical contact.

Its higher-level movement system operates at approximately five updates per second, while its tactile correction mechanism operates at about 20 updates per second. The robot can therefore adjust its grip without waiting for the entire visual reasoning process to run again. That is neat!

In 12 real-world manipulation tasks, researchers reported an average success rate of 65%, compared with 35% for the previous baseline. Removing tactile feedback reduced performance to 42%.

Those numbers are encouraging, but hardly evidence that robots have achieved human dexterity😊. A system failing roughly one-third of its attempts is not yet something we would happily leave alone with delicate laboratory equipment (although, I have seen my fair share of people with even less impressive performance!).

There is also a qualification: The underlying T-Rex research was available earlier in 2026. What is new this week is Panasonic’s formal presentation of the evaluation results ahead of the Conference on Robot Learning. The larger lesson is, arguably, more interesting than the benchmark.

Robots may need two complementary forms of intelligence: One that plans actions using relatively slow, sophisticated reasoning, and another that responds rapidly to physical feedback.

But, aha!

Humans already operate rather like this, don’t they? I mean, don’t we?!. We don’t consciously recalculate the trajectory of every finger each time we pick up a coffee cup.

Anyway, the connection I am picking up is the confirmation that physical AI may advance not simply by making foundation models bigger, but by combining high-level intelligence with specialized, fast feedback systems.

2. AI Safety Is Becoming a Question of What Models Learn, Not Merely What They Are Allowed to Say

On 4 November, Tencent introduced Hy4, a new AI model with expanded visual reasoning, computer-operation and agentic capabilities.

The interesting aspect was not another collection of benchmark claims though. Tencent says it deliberately restricted portions of the model’s cybersecurity training material, excluding certain offensive exploitation content while retaining defensive capabilities such as vulnerability analysis and remediation.

The company also reported internal evaluations covering cybersecurity and biological risks, with plans to release model weights and safety documentation.

This is an important distinction: Much AI safety work focuses on controlling the model’s output, such as refusing dangerous requests, filtering responses or restricting tool access.

Tencent’s approach, however, raises another possibility: Can we influence what dangerous capabilities a model acquires in the first place by changing its training data?

There are reasons to be cautious. Removing explicit offensive instructions does not necessarily prevent a capable model from reconstructing techniques through reasoning. Nor do internal safety evaluations establish that a model is safe under determined adversarial testing.

And restricting knowledge creates another problem in that defensive cybersecurity sometimes requires understanding precisely how attacks work😊.

The boundary between useful defensive expertise and dangerous offensive capability is not always neat (we have seen this friction before in the context of nuclear weapons!).

Nevertheless, this development extends the discussion from last week’s formal software-verification story. AI systems are becoming powerful enough that we increasingly need to ask three different questions:

  • What can the model do?
  • What should it be permitted to do?
  • How can we verify that the restrictions actually work?

The third question may eventually be the most difficult one to address.

3. The Economics of AI Are Moving Inside the Application

On 4 November, Figma reported its third-quarter results and offered an unusually useful glimpse into how AI is being incorporated into a commercial software product. Nice!

According to the company, 65% of Paid Customers with more than $10,000 in ARR were using the Figma agent and over 80% were consuming AI credits, as of 31st of October. It has also more than doubled the share of AI work handled by its own specialized model, which management says improves latency, quality and cost efficiency.

Revenue increased 43% year over year to approximately $391 million. But greater AI adoption also brings substantial computing costs, making the economics of delivering those capabilities increasingly important.

This suggests a potentially significant change in the AI business model.

The early generative-AI industry largely encouraged us to compare general-purpose models: Which one reasons better, codes better or scores higher on benchmarks?

Application providers face a different question now:

Which combination of models delivers the required result at an economically sustainable cost?

A design platform may not need its most powerful general-purpose model for every task. A smaller specialized model may handle many routine operations faster and more cheaply, leaving difficult problems to more capable systems.

That creates a possible architecture:

User request → Task classification → Appropriate model → Result → Verification

The interesting part is not necessarily the model but the system deciding which model to use, how to constrain it and how to measure the result.

Of course, Figma’s commercial performance does not prove that specialized models will dominate. At least not yet! Customer adoption, AI-credit consumption and actual productivity improvements are different measures.

A customer using more AI credits may be receiving more value, or simply consuming more computation. So, we need better evidence on that distinction.

Still, this is a useful reminder that the commercial success of AI may depend as much on orchestration and unit economics as on frontier-model intelligence.

4. The AI Infrastructure Race Is Becoming a Race for Usable Locations

Two developments this week illustrate a problem that no language model can solve merely by becoming more intelligent.

On 3 November, Drax announced a partnership with Brookfield and Radiant to develop a large AI data-center campus at its existing power-generation site in northern England.

The attraction is not simply the available land. The site already has access to a substantial electricity-grid connection, rated at approximately 4 GW. The proposed development could exceed 1 GW of data-center capacity, with its first phase targeted for 2028, subject to approvals and investment decisions.

Then, on 4 November, Reuters examined the increasingly difficult situation facing PJM, the electricity-grid operator covering 13 US states and Washington, D.C.

PJM faces rapidly growing demand from data centere while new electricity generation and grid infrastructure struggle to arrive on schedule. More than 100 GW of generation projects have received approval, but only around 25 GW have been completed. PJM’s capacity assessments indicate potential supply shortfalls of approximately 6.5–6.8 GW in upcoming delivery years.

These numbers require care. Approved projects are not the same as projects under construction, and forecast capacity shortfalls are not predictions that households will necessarily lose electricity.

However, they do reveal an important structural problem: AI infrastructure is not constrained only by how quickly we can manufacture processors. It is also constrained by where those processors can be installed, powered, cooled and connected.

That gives existing industrial sites a potentially unexpected advantage. A former or operating power-generation site may offer valuable infrastructure that would take years to reproduce elsewhere. This includes grid access, industrial zoning, cooling facilities and established transmission connections.

So, it seems that the question is shifting from “Where can we build a data center?“ to: “Where can we build a data center that can actually operate on schedule?”.

There is an additional complication: The availability of a grid connection does not guarantee that sufficient electricity will be available whenever needed.

And projects that look compelling on paper still face construction and financing risks, not to mention the regulatory risk that may become even more prominent, following the blue tsunami in the US mid-term election.

The broader lesson is that physical infrastructure has become an active participant in the AI race, not merely a passive supplier.

The Connection: Intelligence Is Becoming Modular, but Reality Remains Constrained

As usual, the individual developments may appear quite different at first glance:

  • Panasonic is working on robotic dexterity,
  • Tencent is experimenting with training-time safety restrictions,
  • Figma is optimizing the economics of AI inside software, and
  • Data-center developers are searching for sites where electricity infrastructure already exists.

Yet they reveal a common transition:

Intelligence is increasingly becoming something we assemble into systems rather than something we obtain from one model:

  • A robot combines visual reasoning with tactile control,
  • A secure AI system combines model capabilities with training restrictions, evaluations and operational safeguards,
  • A commercial application combines different models, pricing mechanisms and workflow-specific controls, and
  • An AI data center combines processors, memory, electricity, cooling, land, financing and permissions.

So, each system succeeds only when its components work together.

This also suggests a useful refinement to our recurring idea that technological bottlenecks move. Sometimes the bottleneck moves because a previous technical obstacle has been solved. But sometimes it moves because the technology advances faster than the surrounding institutions, infrastructure and economics can adapt.

That second possibility deserves more attention. We are accustomed to measuring technological progress through what becomes possible. But increasingly, we may need to measure it through what becomes reliably deployable (overcoming constraints).

What I Am Watching Next

  1. Physical AI beyond demonstrations: Can tactile-feedback systems achieve consistently high success rates across unfamiliar objects, changing environments and longer sequences of tasks? A 65% success rate is progress, but reliability will determine deployment.
  2. Safety claims that can be independently tested: Tencent’s approach raises an important research question: Can training-data restrictions materially reduce dangerous capabilities without undermining legitimate defensive expertise? Independent evaluations will matter more than corporate assurances.
  3. The economics of specialized AI: Figma provides evidence of meaningful adoption, but the more important question is whether AI-powered applications can demonstrate durable productivity gains while controlling inference costs.
  4. Infrastructure delivery rather than infrastructure announcements: Existing power sites and grid connections are becoming strategic assets. Watch which projects actually secure permits, financing, equipment and dependable electricity, and which remain attractive presentations.

Sources & Further Reading