Here’s what’s on our plate today:

• 🧪 Formula 1's newest engineer has never felt the wind.

• 📰 The AI jobs debate gets messier; Madonna takes on AI; Tidal cracks down on AI music.

• 🧠 Brain Snack: Use surrogate models to narrow the field, verify the edges against reality.

• 🗳️ Poll: Where does AI leave the F1 wind tunnel?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Surrogate models take the wheel. Four of ten 2026 F1 teams are using AI surrogate models to design cars for rules no one has raced under before, replacing slow, expensive physical simulation with pattern-matching trained on past data.

• Speed is real; limits are structural. Neural Concept's platform compresses hour-long CFD runs to 20 seconds. IBM and Dallara matched full CFD results in 10 seconds. But surrogates only know the territory their training data covered, and 2026's rules are almost entirely unmapped.

• The pattern extends beyond racing. NOAA now runs operational AI weather models built on GraphCast. The EMA has qualified AI-generated patient twins for clinical trials. Every field is quietly substituting imitation for reality.

• The stakes are epistemic. When a model is right 99 times, you stop checking the hundredth. The hundredth is always the one drawn from somewhere the model has never been.

Formula 1's newest engineer has never felt the wind

Since 1946, Formula 1 has been one of the clearest expressions of human ingenuity and engineering. Every season brings machines that are faster, lighter, and more efficient than the ones they replace, the product of thousands of hours spent refining every curve, surface, and component. And the innovations developed in pursuit of speed have often found their way beyond the racetrack, influencing the broader automotive industry for decades.

Now, AI is beginning to play a role in that process. In an office in Switzerland, a neural network is helping determine what a Formula 1 car should look like for a season that has not yet begun. Unlike the engineers who spent years studying aerodynamics in wind tunnels and on race circuits, the model has never watched air flow over a wing at 200 miles per hour or seen a prototype respond to the forces of a high-speed corner. It has never experienced the physical world it is modeling. Yet four of the 11 teams on the 2026 grid will arrive at the opening race with aerodynamic shapes that this system, or one very much like it, helped select.

For nearly a century, making a race car faster meant building something and testing it against reality. Engineers fabricated prototypes, conducted wind-tunnel experiments, collected measurements, and returned to the drawing board. Reality was the final authority, and proving an idea meant paying the considerable cost of exposing it to the physical world. Today, much of that work happens before anything is built. AI models have become accurate enough to predict how a design will perform, allowing engineers to explore thousands of possibilities in software before committing to a single prototype. Reality still has the final say, but it is no longer the first place engineers turn for answers.

The world before the shortcut

The impact of this quiet shift cannot be measured solely in terms of what AI models design or how their designs perform in the real world. Their true impact extends beyond the design into the cost of designing the championship-winning car.

For decades, finding speed has depended on a painstaking process known as computational fluid dynamics, or CFD. The software simulates how air flows around a car by solving fluid-dynamics equations with extraordinary detail. Accuracy comes at a price: a single high-fidelity simulation can take hours to complete, even on powerful computing clusters.

Physical testing is even more demanding. Wind tunnel sessions cost significantly more than computer simulations, and under Formula 1's financial and technical regulations, teams are tightly restricted in how much wind tunnel time and aerodynamic testing they can use. Every simulation run and every hour in the tunnel must be carefully spent because both are limited resources.

For engineers, the challenge has never been a shortage of ideas. They have always had more concepts than they could afford to test. The real bottleneck has been the cost of asking nature a question and waiting long enough for a reliable answer. Every design decision required spending precious computing time, valuable wind-tunnel hours, or both, making progress as much a matter of resource management as of engineering skill.

With the advent of AI, companies like Neural Concept are taking center stage in developing the next generation of Formula 1 cars.

Neural Concept began as a spin-out from a Lausanne research lab that started by helping design the world's most aerodynamic bicycle in 2018. However, within six years, its technology had spread to roughly four of the 11 Formula 1 teams, including the once-dominant Williams outfit, which was trying to claw its way back to the front. The idea behind the attempt was deceptively simple. Feed a neural network enough examples of designs and the airflow results that accompany them, and it stops needing to solve the physics. It learns the relationship directly.

Neural Concept's chief executive estimates that a full CFD run, normally taking an hour, can be reduced to as little as 20 seconds through its system. The model has not learned air. It has learned what answers the air tends to give.

These imitators have a name that rarely escapes engineering circles: surrogate models, and their label is honest about what they do. They stand in for something more expensive—a simulation, an experiment, a measurement—and the entire bet is that a good enough stand-in beats a slow original.

And this bet has now gone beyond discussions, drawing the attention of investors. In December 2025, Neural Concept closed a $100M funding round led by Goldman Sachs' growth equity arm, with customers including General Motors, Renault, and Safran.

Why this season, more than any other season

Formula 1 has always embraced new technology. Throughout its history, the sport has served as a proving ground for innovations that eventually found their way into the wider automotive industry. However, what makes the current moment different is that the sport is preparing for its biggest technical reset in more than a decade.

The 2026 regulations introduce redesigned hybrid power units alongside active aerodynamics, allowing the front and rear wings to change shape during a lap to reduce drag on the straights and generate more downforce through corners. These changes fundamentally alter how a Formula 1 car behaves, forcing engineers to rethink designs that have evolved over many seasons.

The challenge is that no team has ever built a Formula 1 car for these rules. There are no historical race results to study, no years of accumulated experience to draw from, and no established design philosophy that has already proved successful. Every team is starting from the same blank sheet of paper, facing an enormous design space with countless possible solutions and very few clues about which ones will ultimately be the fastest.

That is precisely the condition under which an industry reaches for a tool that can explore thousands of possibilities cheaply. When you cannot afford to be wrong slowly, you become willing to be approximately right quickly. The Racing Bulls team said as much when it deployed Neural Concept ahead of the rule change, with team principal Laurent Mekies framing it as a way to explore more design variants in pursuit of an edge "where it matters most." The overhaul did not invent the surrogate model. It simply removed the last reason not to trust one.

What the substitution actually trades

Here is where the thinking gets interesting and slightly uneasy. The speedup is real and frequently enormous. In a neighboring corner of motorsport, IBM and the constructor Dallara built a surrogate to evaluate a race car's rear diffuser and found it reached the same conclusion as full CFD analysis in roughly 10 seconds rather than several hours, identifying the same optimal design with comparable error margins. Both methods picked the better design. The machine just got there first, by a margin that makes a multi-day study feel like a coffee break.

But notice what changed in that comparison, and what did not. The surrogate did not verify the diffuser against the air. It verified it against its own learned impression of what air does, an impression assembled entirely from past simulations. When the new design resembles things the model has seen before, that impression is excellent. The trouble is structural rather than incidental: a surrogate knows only the territory its training data covered, and a shape far enough outside that territory is a shape the model is guessing about, even when it answers with the same confidence it had when it was right.

The engineers building these systems are careful about this, treating the models as a way to narrow the field before committing to high-fidelity simulation rather than as a replacement for it. IBM frames its own work with Dallara as speeding up the workflow without replacing the underlying physics, with generalization to new conditions named explicitly as the open question. The 2026 cars, though, are being built for rules so new that the familiar territory barely exists. Yet the very novelty that makes the surrogate so useful, namely the inability to test these ideas any other way, is also what pushes the design toward the edges of what the model actually knows.

When the imitation becomes the world

Step back from racing, and the pattern repeats wherever reality is expensive to consult. Weather agencies have started replacing parts of their physics-based forecasting with AI emulators; NOAA now runs operational models built on a fine-tuned version of Google DeepMind's GraphCast, a model that produces 10-day forecasts in under a minute, where supercomputer methods needed hours. In drug development, regulators are beginning to accept AI-generated ‘digital twins’ of patients as stand-ins for real control groups, a substitution the European Medicines Agency has formally qualified for certain trials. The same logic governs every case: the original is slow and costly, the imitation is fast and cheap, and the institution decides the imitation is close enough to act on.

What unites these examples is not the technology. It is a decision, made quietly and separately in each domain, about how much imitation a field is willing to treat as truth. Nobody has set that threshold deliberately. It has been set by economics, by the gap between what reality costs and what a model costs, and that gap only widens as the models improve.

The danger is not that surrogate models are dishonest. They are doing exactly what they were trained to do. The danger is subtler: an imitation that is right 99 times teaches you to stop checking the hundredth, and the hundredth is always the one drawn from a place the model has never been.

So the 2026 cars will line up on the grid, shaped in part by a judgment no wind ever rendered, and most of them will be fine, possibly better than they would have been the old way. The question the season cannot answer is the one worth holding onto. When a machine that has never felt the wind tells you what the wind will do, and it is usually correct, at what point do you forget there was ever a difference between the answer and the air?

Brain Snack (for Builders)

Wednesday Poll

Quick Bits, No Fluff

• The AI jobs debate gets messier: New data is complicating the AI-and-jobs picture, with conflicting signals on whether AI is destroying roles, creating them, or just reshuffling who does what.

• Madonna takes on AI: Madonna is wading into the AI music fight, joining the growing roster of major artists pushing back on how their work and likeness get used to train models.

• Tidal cracks down on AI music: Tidal rolled out a new policy to detect, label, and demonetize AI-generated music, one of the clearest moves yet by a streaming platform to draw a line.

Meme Of The Day

The Toolkit

• Together AI: Cloud platform for running and fine-tuning open-source AI models at scale.

• VEED: Browser-based video editor with AI subtitles, voice cloning, and one-click translation.

• Writer: Enterprise AI writing platform with custom models trained on your brand voice.

Rate This Edition