ARTIFICIAL INTELLIGENCE, REAL MISERY

How Fields of Research Science Are Already Answering The Questions Architecture Should Be Asking.

In This Post:

• Make Biology Design-Able

• Hey Chatgpt, Make Me A Protein!

• Ok, Now Make Mice Glow!

• Ai Boosts Everyone's Productivity, Job Dissatisfaction

• Artificial Intelligence, Real Misery

• Ai, On Our Own Terms

• ---

Of all the questions around AI that we find contentious, the idea of AI being 'creative' stands out.  Can AI 'design'?  It seems to be able to create.  It seems to be able to create novel things, as well.  But we bristle at the idea that it is a designer.  In that sense, it's much like Mother Nature.  No one could argue that the designs which we find in the wild can be elegant, efficient, inspiring, and all the other things we human designers hope our designs can be.  But it would be awkward, if not controversial, to say that Mother Nature is designing.

Mother Nature creates via Evolution: an elegant process, that turns random accidents into productive changes along extremely long time scales. Evolution is creative in the sense that we are all creations resulting from that process. Evolution creates newer, better solutions to persisting challenges, as well as identifying novel solutions when circumstances change. But we don't really think of it as design because it lacks intentionality, and authorship.

Design requires an actor or actress as instigator.  And that actor has to have some sort of animus - an idea, or plan of the outcome she wishes to see, even if its form remains elusive.    Good design, like evolution, is a stewarded process.  But if there is no animus, then there is no design - it is merely a child scribbling on walls.  And there can be no animus without a human actor at its root.

Design and evolution share many qualities, if you think about it.  But it would be wrong to consider evolution an act of design, because it is undirected, and authorless.  But what if it weren't?

MAKE BIOLOGY DESIGN-ABLE

I've been taking a break from writing (or, more accurately, publishing) since I read a paper a few months back by a company named Evolutionary Scale who claimed to have trained an AI to design, well, life, in a sense. Their goal is to 'make biology programmable.'  I didn't write about it, at the time, because honestly, it just felt a little overwhelming.  And it took me a few months to digest the implications of such a claim.  Per their press release, they got an AI model to simulate 500 million years of evolution and come up with an entirely new protein.  If that doesn't sound impressive, it's probably because you don't know enough about proteins.  But don't worry, there's more on that below.

I continued to gnaw on the same issue: if AI were genuinely capable of designing life (or, at least, a meaningful part of it) then all other conversations about AI's design capabilities really ought to be put on hold.  I've never seen a building that came even close to a cell in its complexity, efficiency or elegance.

I probably would have stewed on it a while longer, but a few days ago, another paper dropped about the role of AI in research science.  Artificial Intelligence, Scientific Discovery, and Product Innovation (Toner-Rodgers, 2024) detailed an experiment that saw how 1,000 scientists, newly armed with an AI companion, massively expanded their productivity while eviscerating their job satisfaction.  This more recent paper threw enough cold water on the 'AI in Science' thing that it felt manageable all of a sudden.  Even if AI does achieve God-like design capabilities, integrating it into our very human world is going to be a bumpy road, with plenty of opportunities for panic, and an equal number for course correction.

It's also fortunate that research scientists are going through this process before architects - we will have a chance, hopefully, to learn from their mistakes.  

HEY CHATGPT, MAKE ME A PROTEIN!

So did Evolutionary Scale live up to their bold claims?  They didn't make a Frankenstein monster, if that's what you're thinking.   They made an AI, which then designed a completely new protein.  AI's designing proteins, or at least helping to, isn't altogether new.  Demis Hassabis and John Jumper, of Google's Deep Mind, shared this year's Nobel Prize in Chemistry with Professor David Baker, for their collective work on computational protein design.  

Why does that merit a Nobel?  Any protein is an elaborate 3D 'folding' problem.  The way a protein folds into shape is critical, because its eventual shape determines how it interacts with other parts of the cell.  How it folds is governed by the sequence of amino acids, the physical and chemical properties of those amino acids, and environmental factors like Ph, temperature, etc.

It’s computationally complex, and to predict how any sequence of amino acids will fold into a protein, researchers have developed a variety of 'protein language models' which work exactly like ChatGPT, but instead of predicting the next word in a sentence, they predict the structure of a protein.  In both cases, machines are making use of understood rules of 'grammar' and lots of example data.  Large Language Models use words, Protein Language Models use proteins.  

All of that is pretty cool, and apparently enough to win you a Nobel Prize.  But it's also yesterday's news.  The problem with protein models like AlphaFold is that they tell you little about what a protein actually does.  It’s like having a large language model trained on pure text and no speech.  It could arrange all the sentences and words correctly, but wouldn't be able to match them with sounds.  'Frankenstein' wouldn't necessarily correspond to ˈfraŋ-kən-ˌstīn' unless that LLM had some kind of speech model where it could pair letters with sounds.

Evolutionary Scale seems to have solved that problem back in June with their model ESM3, which simultaneously reasons over the sequence, structure, and function of proteins.  It's trained on human language, like ChatGPT or Claude, but it's also trained on 2.8 billion naturally occurring proteins sampled from various organisms and biomes, so it can speak to, and reason between, both human language and the 'language' of proteins.  It used this ability to create a completely new protein called ‘esmGFP’, which is complete mad scientist stuff.

OK, NOW MAKE MICE GLOW!

There's a protein called Green Fluorescent Protein ("GFP") which is naturally found in certain kinds of jellyfish.  It does exactly what you think.  It causes some jellyfish to glow fluorescent green.  For some time now, scientists have been able to isolate this protein and put it in other things, like mouse cells.

Real Lab Photo - not Midjourney

Why would anyone do this?  I think the question is 'why would anyone not do this?'  If I could put this protein in my dog, I would do it today, and then read in bed tonight by his gentle glow.

Definitely Midjourney - my dog does not do this

But seriously, you can do useful things like put them in cancer cells and then track where the cancer lights up.  But that’s also old news - they’ve been transplanting GFPs into all sorts of things for years.

What Evolutionary Scale did recently was a quantum leap forward: a few months ago, Evolutionary Scale announced that their protein/language model 'ESM3' had designed an entirely new glowing protein.  They basically asked ESM3 to create 100 designs for proteins that might glow.  Because ESM3 can reason over structure and function, it was able to say to itself 'hey, self, if I wanted a protein to glow, what kind of amino acids would I put in there?'  In other words, 'how would I design it?'  

This is, of course, what architects do.  We take an imagined end result, and then try to reverse engineer the process and materials to get there, through a rigorous process that’s informed by lots of precedents and experience.  That imagined end result is not complete, in the sense that we're not entirely sure how the design is going to come together.  But we have a functional result in mind, and our path towards it is governed by our vast experience in buildings (although probably not 2.8 billion of them).

In ESM3's initial effort, only yielded a handful of weakly glowing proteins.  But Evolutionary Scale scientists took those more successful attempts back to ESM3, and told it 'these are on the right track, make more variations of these.'  In the subsequent effort, ESM3 produced a new GFP that shined as brightly as the all-natural one.  The new GFP, which they named esmGFP, does not exist in nature, is only 58% similar to the closest known fluorescent protein.  In genetic terms, that's a lot.  esmGFP is not a riff on some protein we already know - it is, in evolutionary terms, a novel creation.

Evolution is a terribly slow process.  Mother Nature never had a button to push and say 'okay, take this horse and give it a longer neck' and thereby create a giraffe.  I can only imagine the tens of thousands of generations of horses that had to look longingly up at the delicious tree leaves, waiting for their necks to grow.  And evolution is undirected, in the sense that it relies on mutations that must occur naturally, and do so randomly.

ESM3 allows for intentionality in evolution. 

You can instruct the model to, say, 'make a protein that smells like ham' and, after a few months of compute time, that protein comes out.  In the case of esmGFP, Evolutionary Scale estimated that ESM3's process in creating esmGFP was the equivalent of 500 million years of evolution, accomplished in just a few months.  The chances of esmGFP occurring via random mutation, was estimated at a mind-boggling 20²²⁹ X 4096²⁹⁹

More exciting than the technology itself, is how Evolutionary Scale has deployed it.  Evolutionary Scale has committed to open-sourcing their work.  They actually have an open-source option now.  You can download it here.  Go make a dragon, I won't tell anyone.  

In all seriousness, that is a profound move, because it means that every scientist - professional and amateur, globally, can get to work on designing custom proteins to do all sorts of things.  How about a carbon-eating, self-healing sheet product to replace drywall?  Install it once, and never touch it again.  Any scuffs, gouges, etc. are healed via the same mechanisms through which your skin heals after a cut.  We've dreamed about such sci-fi materials before.  But now the world has wide access to a powerful tool to actually design them.

It's a palatable, and exciting vision of the near future.  So let's come back down to Earth:

AI BOOSTS EVERYONE'S PRODUCTIVITY, JOB DISSATISFACTION

More recently, Aaron Toner-Rodgers, an MIT researcher conducted an experiment which gave an AI Tool (‘The Tool’) designed for new material discovery to 1,018 materials research scientists working in a large U.S. company.  This company develops materials for healthcare, optics, and industrial manufacturing, cutting across the fields of chemistry, physics, and engineering.

The results were striking, and yet predictable.  The AI-augmented scientists were able to create 44% more new materials and file 39% more patents. 

Overall new product development rose by 17%, suggesting that the AI not only had a knack for creating new materials in the abstract, but an intuition about which materials had practical applications ripe for commercialization.

The only problem?  A whopping 82% reported decreased satisfaction with their jobs, compared to their life before AI.

ARTIFICIAL INTELLIGENCE, REAL MISERY

The thing is, The Tool was able to automate the majority of the "idea generation" tasks, forcing the scientists into the role of evaluating the AI's suggestions, rather than dreaming up new materials.  Prior to the introduction of The Tool, the scientists had spent roughly half their time conceptualizing potential new materials, which does sound like a lot of fun ('hey, what if we made self-healing drywall!).  

After the introduction of The Tool, time spent on this dreaming fell to 16%.  The time spent in assessing candidate materials consequently rose to 74%.

Historically, material discovery is a lot of trial and error.  Materials scientists have to wonder 'what if' and tinker about with their chemistry sets.  The Tool was able to virtually simulate a lot of this early experimentation on its own.  Human scientists then just had to provide the second-order analysis - to build the proposed new materials, test them, and apply critical judgment.

That's the lede.  The more interesting outcomes lay further down in the 77 page paper, specifically, in how the effects of AI were unevenly distributed through the sample group of 1.018 scientists.

The most productive scientists benefited way, way more from the introduction of AI.  The scientists in the top decile of pre-treatment productivity showed an 81% increase in their own productivity.  Toner-Rodgers measured 'pre-treatment productivity' by documenting how many new materials and patents each scientist had to their name, before the introduction of The Tool.¹  

He confirmed the validity of his 'pre-treatment productivity' metric by observing a positive correlation with the ranking of scientists' alma maters and their tenure in the lab.  Translation: if you went to a leading school, and have been doing this a long time, you're much more likely to have discovered more materials.

It would seem that the more productive scientists were able to deploy some kind of superior intuition against the new candidate materials being proposed by The Tool.  This allowed them to rapidly and effectively filter the AI's suggestions into those that showed great promise, and those that showed less.  The less productive scientists had no such intuition to deploy - so they essentially had to accept all of The Tool's suggestions, and experiment with them randomly to see which ones would pan out.

That raises two unsettling problems:

First, if The Tool dramatically improves the outcomes for those scientists that are already super productive, and has no effect for the less productive scientists, it dramatically increases the skills gap and the value gap between the two groups of scientists.  As AI technology continues to improve, this gap only widens.  It's easy to imagine the end result: you have 1,018 scientists, but 95% of your results are achieved by only the top 100.  At some point, you fire the other 918 because they are just taking up space.  Then you have 918 mad scientists, out on the street, each one on their way to becoming a Marvel supervillain.

Second, and although It's never stated outright, it seems clear that at least part, if not most, of a scientist's 'pre-treatment productivity' had to do with age.  The more tenure a scientist had, the more effective and efficient they were at evaluating The Tool's suggestions.  That's not a surprise, and hardly feels like an injustice.  Judgment comes from experience.  Senior people have more judgment, and more intuition, about their subject matter than junior people.  It follows that they would be better equipped to evaluate the judgment of an AI.  But experience comes from time, and mistakes.  If the participation of those younger, 'less productive' scientists is marginalized, they never get a chance to grow into the judgment that senior scientists clearly possess.

This problem has already materialized in medical surgery.  Historically, we train young surgeons in a 'see one, do one, teach one' model.  The senior surgeon literally can't do what they do without constant, hand-on assistance from a resident (e.g. holding a patient's chest cavity open).  With the rise of robotic-assisted surgery, all surgeons - young and old - must learn to use the robot.  But the robot allows one surgeon to do the surgery, so the resident is often left to just watch, rather than participating in the doing.

AI, ON OUR OWN TERMS

The prospect of AI being able to design life is pretty exciting, and a teaser of the kind of scientific discoveries we can look forward to. New Cancer drugs? Check. Self-healing building materials? Check. Infinite, biodegradable alternatives to plastic? Check. 

But the Toner-Rodgers paper highlights the difficulty we'll have in integrating such awesome power into our everyday existence.  We will have to draw some hard lines around what we will, and won’t accept:

• We cannot accept a measure of productivity that does not account for 82% of people being sad, angry, bored and idle.  That mistake, is, arguably, at the core of a lot of our contemporary problems.

• Nor can we cannot accept a technological revolution that permanently severs the chain of human knowledge. If junior people are blocked from being able to dream, experiment, make mistakes, and grow, they will never be able to cultivate the kind of intuition that proved to be such an advantage in Toner-Rodgers experiment.  As those more senior scientists start to grow old, retire, and die, there won't be anyone left to apply judgment to the suggestions that AI is making.

• We can't accept any form of revolution that results in a large population of young, unemployable, angry people in any society.  That hardly ever works out well for anybody, and it may shortly happen with every country at the same time, if we're not careful.

And if we accept that 10% of the scientists, augmented with AI, can create all the new materials humanity needs, we have to have a plan for what to do with the other 90% of scientists.  What does a society do with someone who is highly, technically skilled, and has multiple advanced degrees, but is junior enough that they don't create sufficient economic value to merit their own employment in the age of AI?

Extrapolating outward to the scale of society, you can imagine one of the scenarios that has always kept me up at night: it's not that AI would completely eliminate human roles in science, or architecture, or anything else.  It's that it would eliminate a lot or most human roles. 

A society where 10% of the people create all the economic value and everyone else literally has nothing to do would be catastrophic, on several levels.  The idle 90% would inevitably grow depressed, and angry.  The 10% would grow resentful, considering that they're the only ones that still have to go to work every morning.  It seems likely that at some point the 10% would grow totalitarian, insisting that because they do all the work, they should make all the decisions for everyone else.

---

I mentioned in the introduction that it was fortunate that research scientists were confronting this conundrum before architects did, because we'd be able to learn from their mistakes.  That is only true if we actually learn from their mistakes.  If we don't actively learn from how AI is already penetrating other fields, the manner in which it penetrates architecture will likely be predetermined.  In other words, we have to adapt before the wave hits.  It would be foolish to persist with the naive idea that the AI wave will somehow pass over architecture, because what architects do is design and design can't be done by machines.  Machines are designing novel forms of life, for f*&#;s sake.  It would be equally foolish to assume that we can contain AI's influence to making cool images on Midjourney, or otherwise dictate how the wave will come in.  We can, however, design our response to it.