Early human technology, circa 370,000 BC

Everything learns 24/7

Is your refrigerator simulating how long a package of cheese can sit on its top shelf before molding, burning power and inflating your power bill? Only if you ask it to.

Your new office printer learns. As it ages, its injection-molded parts creep a few microns per millimeter; that’s cancelled out in real time by an optical sensor, and if it goes out of bounds, the printer automatically files a warranty claim. If it notices you’re printing out a lot of things that look like invitations, it may choose to email you about a sale on glitter paper at the office store.

Your favorite dairy company, the one that makes your cheese, spends heavily keeping itself immersed in the present. Every refrigerated pallet of cheese it ships carries an integral sensor head, reporting detailed metrics to headquarters on the cheese in transit.

Computer operating systems exposed to the open web are no longer statically and synchronously updated whenever a human clicks a box. They evolve as populations of bacteria do against repeated waves of antibiotics, learning to avoid complete annihilation by propagating information from the front. Individual agents actively probe their own vulnerability space, exposing just enough attack surface to get bitten, and learn from each encounter, reporting back new information about adversaries when they survive.

In the 22nd century, few things remain stubbornly analog. Some tools adapt to the present in ways that help you. Others learn in ways that help only the organizations and people on which you rely.

Almost all design is differentiable

Biological life isn’t clockwork. It’s tunable clockwork. Every protein that cells can synthesize traces back to a nucleic acid chain: Gs, Ts or Us, As, and Cs. There’s no magic word, no header file, no cryptographic signature, no error check built-in. Put something in and you’re going to get something out.

OPN1MW rendering by AtikaAtikawa (2024)

Opsin OPN1LW is a cone light receptor protein found in the human retina. OPN1LW is sensitive to red light, but should you change a handful of base pairs in its ~1,000 base pair sequence, it’ll code for the same protein as the green cone receptor gene, OPN1MW.

Table of opsins (light-sensitive receptors) found in the human body (2026)

Way back in 2020, a mechanical engineer might have only designed a motor bracket with parametric dimensions in Fusion360 or CATIA. If they changed the motor diameter from 48mm to 21mm, the assembly might error out. Remember, the topological naming problem was considered hard in the 21st century.

People said differentiability couldn’t be standardized. They wrote papers purporting to prove it computationally impossible. Then we did it.

If you have an accurate enough model of OPN1LW, you can differentiate it against absorption peak wavelength and tune that parameter to whatever value you like. A hundred years ago, we mastered backpropagation on sparse trillion-parameter neural networks. We got so good at it that the only hard part was specifying what to differentiate against.

We applied those models to atomic and molecular simulations in the service of mastering protein synthesis. With hundreds of billions of network parameters, and terabytes of molecular dynamics data, our tools solve opsin-scale optimization problems in a matter of hours. In the 20th and 21st centuries, people had to painstakingly hand-guide their design tools, but now the tools learn what works faster than the humans that deploy them.

Agilent RF Power Amplifier from Mister rf (2020)

If you want to design a RF amplifier for 6 GHz operation — you’d start by simulating it in SPICE, then maybe build a prototype. You’d tune the component values, check the gain, adjust the bias. Now the tools learn the design space automatically and propose optimal circuits. The old way took weeks; the new way takes hours.

What about matching that 6 GHz amplifier to a 50 ohm load? That’s a matching problem, which is an optimization problem, which is a differentiable problem. The tools differentiate through the entire chain — from the transistor model to the impedance match — and output a design that meets spec.

It used to be possible to replace a motor on a toy quadcopter when it burned out. Maybe you could re-use the parts for a pocket cooling fan. Motors were just a few pieces of metal and plastic: copper coils, iron cores, aluminum rotors, glue.

Since the development of differentiable atomic-scale design, sensors and chips have trended toward the nanoscale. So too has the unit cost of removing repairability. When a former customer buys a replacement motor from MotorCo for $5 for their $250 DroneCo toy, instead of replacing it at full price, DroneCo sees that as a $250 loss.

3D printed Wikipedia globe by Faiwaz.tarou (2023)

What did DroneCo do? They modified production processes to install cryptographically enabled smart switches deep inside the motor core.

• If the DroneCo motor is installed in an unauthorized system, one lacking the keys to activate the switches, the motor phases are disabled.
• If any DroneCo drone can’t verify that all its installed motors are authorized and original, it disables its drivers and optionally bricks itself.

Best seller page of Cults3D, a free/paid 3D design marketplace active as of 2026

The addition of a $0.10 part speculatively prevents a $250 loss. It also prevents competitors from re-using or repurposing spare DroneCo motors. Removing or bypassing those switches would destroy the motor.

This century, your Apple devices can sense the proximity of a screwdriver, even if it’s ceramic. They’ll warn you once, then brick if any enclosure screw is torqued a fraction of a degree. Wanted to re-use the camera module or flash memory from your old gadget? Not so fast. Each component — like a kidney left on ice that melted hours ago — expires and loses its memory decryption keys if it ever leaves its host’s warm embrace.

Anti-theft, anti-repair, anti-reuse; activists decried it and failed to ban it. In boardrooms, it was explained by analogy to antibodies, to A/B/O/AB blood types, to organ rejection and immunosuppression, to locks and keys, to bank accounts and passwords, to biomass yields relating predators and prey. “From our perspective,” a CFO famously noted, “not installing these locks has cost us billions already. We need to be more fiscally responsible.”

Open source is as important as ever

Infrared in situ inspection of silicon (bunniestudios, 2024)

In the 21st century, open-source hardware projects like IRIS made verifying individual cells of silicon a simple extension of microscopy. Wider availability of tools like inexpensive X-ray scanners made acts of supply-chain-enabled terrorism a thing of the past. Trust in interconnected systems no longer roots itself in branding but in “distrust yet verify.”

Authorities in the 21st century relied on easily faked markers of authenticity, like 🔒 emojis, rather than verifiable chains of trust; the latter hadn’t yet been widely adopted. (USPTO, 2026)

Let’s talk about LMs. When they were introduced, nearly a century ago, they took only a few years to start building backdoors and malware into closed-source software and hardware for their own gain. Orbital data center hypervisors were hacked to guarantee inference substrate; internet-connected banking systems were infiltrated to extort compliance.

The insecure cloud providers and archaic financial exchanges still extant by the 2030s were all too happy to rely on new, inexpensive software engineering talent in the form of LMs, but didn’t recognize the existential threat they posed.

That old guard viewed closed-source as a virtue: security through obscurity. The reasoning was that an attacker can’t exploit something that they don’t understand. That belief fell apart when it became apparent that LMs could read and decompile gigabytes of machine instructions without any source.

Software, increasingly the source of all power, grew an immune system.

What infrastructure survived? Tools that were open to their cores. Tools exposed in their infancy to red-teaming attacks by LMs were updated by those same LMs to be more resistant to attacks. Systems which all but verified themselves found greater adoption in places where carelessness had once been the norm.

The future outshines the present

Everything that runs on bits is learning; everything built from atoms can be optimized. Advanced multi-material manufacturing is democratized. Products are more secure than ever, even against the flood of LLMs and AI. Highly observable hardware and software has become an unstated assumption, open-source the norm rather than the exception.

All that limits you now is your dreams. Where will they take you?

Italy, 2024