Why the Future of AI & Computers Will Be Analog

If your taste in TV is anything like  mine, then most of your familiarity with  

what analog computing looks like probably  comes from the backdrops of something like

Columbo. Since digital took over the world,  analog has been sidelined into what seems  

like a niche interest at best. But this retro  approach to computing, much like space operas,  

is both making a comeback, and also something  that never really left in the first place.

I found this out for myself about a year  ago, when a video from Veritasium sparked  

my curiosity about analog computing. After that,  I started to read a few articles here and there,  

and I gotta say…it broke my brain a bit. What  I really wanted to know, though, was this:  

How can analog computing impact our  daily lives? And what will that look  

like? Because I definitely don’t  have room in my house for this.

I’m Matt Ferrell … welcome to Undecided. 

This video is brought to you by Surfshark and all  of my patrons on Patreon, but more on that later.

Depending on how old you are, you may remember  when it was the norm for a single computer to  

take up more square footage than your average  New York City apartment. But after the end of the  

Space Age and the advent of personal computers,  our devices have only gotten smaller and smaller.  

Some proponents of analog computing argue that  we might just be reaching our limits when it  

comes to how much further we can shrink. We’ll  get to that in a bit, though. Emphasis on bits.

Speaking of bits, this brings us to the  fundamental difference between analog  

and digital. Analog systems have an infinite  number of states. If I were to heat this room  

from 68 F to 72 F, the temperature would  pass through an infinite set of numbers,  

including 68.0000001 F and so on. Digital  systems are reliant on the number of “bits”  

or the number of transistors that are  switched either on or off. As an example,  

an 8-bit system has 2^8, or 256 states. That  means it can only represent 256 different numbers.

So, size isn’t the only aspect of the  technological zeitgeist that’s changed. Digital  

computers solve problems in a fundamentally  different way from analog ones. That’s led  

to some pretty amazing stuff in modern day…at  a cost. Immensely energy intensive computing  

is becoming increasingly popular. Just look at  cryptocurrencies and AI. According to a report  

released last year by Swedish telecommunications  company Ericsson, the information and  

communication technology sector accounted for  roughly 4% of global energy consumption in 2020.

Plus, a significant amount of digital  computing is not the kind you can take to  

go. Just among the thousands of data centers  located across the globe, the average campus  

size is approximately 100,000 square feet (or  just over 9,000 square meters). That's more  

than 2 acres of land! Data scientist  Alex de Vries estimates that a single  

interaction with a LLM is equivalent to “leaving  a low-brightness LED lightbulb on for one hour.”

But as the especially power-hungry data centers,  neural networks, and cryptocurrencies of the world  

continue to grow in scale and complexity…we still  have to reckon with the climate crisis. Energy  

efficiency isn’t just good for the planet,  it’s good for the wallet. A return to analog  

computing could be part of the solution. The  reason why is simple: you can accomplish the  

same tasks as you would on a digital setup  for a fraction of the energy. In some cases,  

analog computing is as much as 1,000 times  more efficient than its digital counterparts.

Before we get into exactly how it works and why  we’re starting to see more interest in analog  

computers again, I need to talk about another  piece of tech that can really help in your daily  

digital life and that’s today’s sponsor,  Surfshark. Surfshark is a fast, easy to  

use VPN full of incredible features that you can  install on an unlimited number of devices with one  

account. Most of the time when we talk about VPNs  we’re focused on giving yourself security as you  

travel around the world, but it can do way more  than that. Since you can make it look like your  

IP address is coming from somewhere else in the  world, it unlocks geofencing blocks on content,  

like streaming services. But … that’s not  all. Even shopping services will sometimes  

gate prices based on your location, so you can  change your location to make sure you’re getting  

the best prices. They also have add-ons to their  VPN service to unlock things like Surfshark Alert,  

which will let you know if your email or personal  details, like passwords, have been leaked online  

in a data breach. Right now they’re running a  special deal … use my code UNDECIDED to get up  

to 3 additional months for free. SurfShark  offers a 30-day money-back guarantee,  

so there’s no risk to try it out for yourself.  I’ve been using Surfshark for years and love it.  

Link is in the description below. Thanks  to Surfshark, for supporting the channel.  

And thanks to all of you, as well as my patrons,  who get early, ad-free versions of my videos. So  

back to how much more energy efficient analog  computing is from its digital counterparts.

To understand how that works, exactly, we  first need to establish what makes analog  

computing…analog. The same way you would make  a comparison with words using an analogy,  

analog computers operate using a physical  model that corresponds to the values of  

the problem being solved. And yeah,  I did just make up an analog analogy.

A classic example of analog computing is the  Monetary National Income Analogue Computer,  

or MONIAC, which sounds like a long forgotten car  brand, which economist Bill Phillips created in  

1949. MONIAC has a single purpose: to simulate  the Great British economy on a macro level.  

Within the machine, water represented money as  it literally flowed in and out of the treasury.  

Phillips determined alongside his colleague  Walter Newlyn that the computer could function  

with an approximate accuracy of ±2%. And  of the 14 or so machines that were made,  

you can still find the first churning away  at the Reserve Bank Museum in New Zealand.

It’s safe to say that the MONIAC worked  (and continues to work) well. The same goes  

for other types of analog computers, from  those on the simpler end of the spectrum,  

like the pocket-sized mechanical  calculators known as slide rules,  

to the behemoth tide-predicting  machines invented by Lord Kelvin.

In general, it was never that analog computing  didn’t do its job — quite the opposite. Pilots  

still use flight computers, a form of slide  rule, to perform calculations by hand,  

no juice necessary. But for more generalized  applications, digital devices just provide a level  

of convenience that analog couldn’t. Incredible  computing power has effectively become mundane.

To put things into perspective, an iPhone  14 contains a processor that runs somewhere  

above 3 GHz, depending on the model.  The Apollo Guidance Computer, itself a  

digital device onboard the spacecraft  that first graced the moon’s surface,  

ran at…0.043 MHz. As computer science  professor Graham Kendall once wrote,  

“the iPhone in your pocket has over 100,000 times  the processing power of the computer that landed  

man on the moon 50 years ago.” … and we use it  to look at cat videos and argue with strangers.

In any case, that ease of use is one of the  reasons why the likes of slide rules and  

abacuses were relegated to museum displays  while electronic calculators reigned king.  

So much for “ruling.” But, while digital  has a lot to offer, like anything else,  

it has its limits. And mathematician  and self-described “analog computer  

evangelist” Bernd Ulmann argues that we can’t  push those limits much further. In his words:

“Digital computers are hitting basic  physical boundaries by now. Computing  

elements cannot be shrunk much more than today,  

and there is no way to spend even more  energy on energy-hungry CPU chips today.”

It’s worth noting here that Ulmann said  this in 2021, years ahead of the explosion  

of improvements in generative AI we’ve  witnessed in just the past few months,  

like OpenAI’s text-prompt-to-video  model, Sora. Which, really disturbs  

me and I'm very excited by all at the same  time, I need to make a video about that.

But what did he mean by “physical  boundaries”? Well…digital computing  

is starting to bump up against the law.  No, not that kind…the scientific kind.  

There’s actually a few that are at play  here. We’ve already started talking about  

the relationship between digital computing  and size, so let’s continue down that track.

In a 1965 paper, Gordon Moore, co-founder of  Intel, made a prediction that would come to  

be known as “Moore’s Law.” He foresaw that  the number of transistors on an integrated  

circuit would double every year for the  next 10 years, with a negligible rise in  

cost. And 10 years later, Moore changed his  prediction to a doubling every two years.

As Intel clarifies, Moore’s Law isn’t a scientific  observation, and Moore actually isn’t too keen on  

his work being referred to as a “law.” However,  the prediction has more or less stayed true as  

Intel (and other semiconductor companies)  have hailed it as a goal to strive for:  

more and more transistors on smaller and  smaller chips, for less and less money.

Here’s the problem. What happens when we can’t  make a computer chip any smaller? According to  

Intel, despite the warnings of experts in the past  few decades, we’ve yet to hit that wall. We can  

take it straight from Moore himself, though,  that an end to the standard set by his law is  

inevitable. When asked about the longevity of his  prediction during a 2005 interview, he said this:

“The fact that materials are made of atoms is  the fundamental limitation and it's not that  

far away. You can take an electron micrograph from  some of these pictures of some of these devices,  

and you can see the individual atoms of  the layers. The gate insulator in the most  

advanced transistors is only about three molecular  layers thick…We're pushing up against some fairly  

fundamental limits, so one of these days we're  going to have to stop making things smaller."

Not to mention, the more components you cram  onto a chip, the hotter it becomes during use,  

and the more difficult it is to cool down. It’s  simply not possible to use all the transistors  

on a chip simultaneously without risking a  meltdown. This is also a critical problem  

in data centers, because it’s not only  electricity use that represents a huge  

resource sink. Larger sites that use liquid  as coolant rely on massive amounts of water  

a day — think upwards of millions of gallons.  In fact, Google’s data centers in The Dalles,  

Oregon, account for over a  quarter of the city’s water use.

Meanwhile, emerging research on new  approaches to analog computing has  

led to the development of materials that  don’t need cooling facilities at all.

Then there’s another law that stymies  the design of digital computers:  

Amdahl’s law. And you might be able to get a  sense of why it’s relevant just by looking at  

your wrist. Or your wall. Analog clocks, the kind  with faces, can easily show us more advantages of  

analog computing. When the hands move forward on  a clock, they do so in one continuous movement,  

the same way analog computing occurs in real time,  with mathematically continuous data. But when you  

look at a digital clock, you’ll notice that it  updates its display in steps. That’s because,  

unlike with analog devices, digital information  is discrete. It’s something that you count,  

rather than measure, hence the  binary format of 0s and 1s.

When a digital computer tackles a problem,  it follows an algorithm, a finite number  

of steps that eventually lead to an answer.  Presenting a problem to an analog computer is  

a completely different procedure, and this cute  diagram from the ‘60s still holds true today:

First, you take note of the physical laws  that form the context of the problem you’re  

solving. Then, you create a differential  equation that models the problem. If your  

blood just ran cold at the mention of  math, don’t worry. All you need to know  

is that differential equations model dynamic  problems, or problems that involve an element  

of change. Differential equations can be used  to simulate anything from heat flow in a cable  

to the progression of zombie apocalypses. And  analog computers are fantastic at solving them.

Once you’ve written a differential equation,  you program the analog computer by translating  

each part of the equation into a physical part of  the computer setup. And then you get your answer,  

which doesn’t even necessarily  require a monitor to display!

All of that might be tough to envision, so  here’s another analog analogy that hopefully  

is less convoluted than the labyrinth of wires  that make up a patch panel. Imagine a playground.  

Let’s say two kids want to race to the same  spot, but each one takes a different path.  

One decides to skip along the hopscotch court,  and the other rushes to the slide. Who will win?

These two areas of the playground are  like different paradigms of computing.  

You count the hopscotch spaces outlined on  the ground and move between them one by one,  

but you measure the length of a  slide, and reach its end in one  

smooth move. And between these two  methods of reaching the same goal,  

one is definitely a much quicker process than  the other…and also takes a lot less energy.

There are, of course, caveats to analog.  If you asked the children in our playground  

example to repeat their race exactly the same  way they did the first time, who do you think  

would be more accurate? Probably the one whose  careful steps were marked with neat squares,  

and whose outcomes will be the same — landing  right within that final little perimeter of  

chalk. With discrete data, you can make perfect  copies. It’s much harder to create copies with  

the more messy nature of continuous data. The  question is: do we even need 100% accurate  

calculations? Some researchers are proposing  that we don’t, at least not all the time.

That said, what does this have to do with Amdahl’s  law? Well, we can extend our existing scenario  

a little further. It takes time to remember  the rules of hopscotch and then follow them  

accordingly. But you don’t need to remember any  rules to use a slide — other than maybe “wait  

until there isn’t anybody else on it.” Comment  below with your favorite playground accidents!

In any case, because digital computers 1.  reference their memories and 2. solve problems  

algorithmically, there will always be operations  (like remembering hopscotch rules) that must be  

performed sequentially. As computer science  professor Mike Bailey puts it, “this includes  

reading data, setting up calculations, control  logic, storing results, etc.” And because you  

can’t get rid of these sequential operations,  you run into diminishing returns as you add  

more and more processors in attempts to speed up  your computing. You can’t decrease the size of  

components forever, and you can’t increase  the number of processors forever, either.

On the other hand, analog computers  don’t typically have memories they  

need to take time to access. This allows  them more flexibility to work in parallel,  

meaning they can easily break  down problems into smaller,  

more manageable chunks and divide them  between processing units without delays.

Here’s how Bernd Ulmann explains it In his 2023  textbook, Analog and Hybrid Computer Programming,  

which contributed a considerable  amount of research to this video:

“Further, without any memory there  is nothing like a critical section,  

no need to synchronize things,  no communications overhead,  

nothing of the many trifles that haunt  traditional parallel digital computers.”

So, you might be thinking: speedier, more  energy-efficient computing sounds great,  

but what does it have to do with me? Am I going to  have to learn how to write differential equations?  

Will I need to knock down a wall in my office  to make room for a retrofuturist analog rig?

Probably not. Instead, hybrid computers that  marry the best features of both digital and  

analog are what might someday be in vogue.  There’s already whisperings of Silicon Valley  

companies secretly chipping away at…analog  chips. Why? To conserve electricity … and  

cost. The idea is to combine the energy  efficiency of analog with the precision of  

digital. This is especially important for  continued development of the power-hungry  

machine learning that makes generative  AI possible. With any hope, that means  

products that are far less environmentally  and financially costly, to maintain.

And that’s exactly what Mythic, headquartered in  the U.S., is aiming for. Mythic claims that its  

Analog Matrix Processor chip can “deliver the  compute resources of a GPU at 1/10th the power  

consumption.” Basically, as opposed to storing  data in static RAM, which needs an uninterrupted  

supply of power, the analog chip stores data  in flash memory, which doesn’t need power to  

keep information intact. Rather than 1s and 0s,  the data is retained in the form of voltages.

Where could we someday see analog computing  around the house, though? U.S.-based company  

Aspinity has an answer to that. What it  calls the “world’s first fully analog  

machine learning chip,” the AML100, can act as  a low-power sensor for a bunch of applications,  

according to its website. It can detect a wake  word for use in voice-enabled wearables like  

wireless earbuds or smart watch, listen for  the sound of broken glass or smoke alarms,  

and monitor heart rates, just to name a few.

For those devices that always need to be  on, this means energy savings that are  

nothing to sneeze at (although I guess  you could program an AML 100 to detect  

sneezes). Aspinity claims that its chip  can enable a reduction in power use of 95%.

So, the potential of maximizing efficiency  through analog computing is clear,  

and the world we interact with every day is itself  analog. Why shouldn’t our devices be, too? But to  

say that analog programming appears intimidating  (and dated) is…somewhat of an understatement.

It’ll definitely need an image upgrade  to make it approachable and accessible to  

the public — though there are already models out  there that you can fiddle with yourself at home,  

if you’re brave enough. German company  Anabrid, which was founded by Ulmann in 2020,  

currently offers two: the Analog Paradigm  Model-1, and The Analog Thing (or THAT).

The Model-1 is intended for more experienced  users who are willing to assemble the machine  

themselves. Each one is produced on  demand based on the parts ordered,  

so you can tailor the modules to your needs.

THAT, on the other hand…and by THAT I mean THAT:  The Analog Thing, is sold fully assembled. You  

could also build your own from scratch — the  components and schematics are open source.

So what do you actually do  with the thing? Y’know…THAT?  

I’ll let the official wiki’s FAQ answer that:

“You can use it to predict in the natural  sciences, to control in engineering,  

to explain in educational settings, to imitate in  gaming, or you can use it for the pure joy of it.”

The THAT model, like any analog computer,  solves whatever you can express in a  

differential equation. As a reminder, that’s  basically any scenario involving change,  

from simulating air flow to solving  heat equations. You can also make music!

But as analog computing becomes more  readily available, there’s still a  

lot of work to be done. For one thing,  It’ll take effort to engineer seamless  

connectivity between analog and digital  systems, as Ulmann himself points out.

Until then, what do you think? Should we take  the word of analog evangelists as gospel? Or  

are we better off waiting for flying cars?  Jump into the comments and let me know. Be  

sure to check out my follow-up podcast, Still  To Be Determined, where we'll be discussing  

some of your feedback. Before I go, I’d  like to welcome new Supporter+ patrons  

Charles Bevitt and Tanner. Thanks so much for  your support. I’ll see you in the next one.