Compared with OpenAI’s CEO Sam Altman, Demis Hassabis, the cofounder of DeepMind, is perhaps less well know amongst the general public. However, after watching The Thinking Game recently, the 2024 documentary tracking DeepMind’s staggering progress towards Artifical General Intelligence (AGIs), Hassabis may well turn out to have been the more impactful in the course of human cultural and technological evolution. Founding the company in 2010 with the modest goal of ‘solving intelligence’, DeepMind was then acquired by Google in 2014 for an unfathonable amount of money, after being able to show remarkable success in teaching AIs to solve complex but well defined goal-orientated activities.
These complex but well defined goal-orientated activities: playing computer games, where the goals can be as simple as ‘win points’, but means of achieving this are anywhere but simple. Starting from Atari games from the late 1970s and 1980s, Alpha, DeepMind’s insatiably improving megatech, initially struggled, like any infant, to even know how to move a paddle up or down on a screen. But before too long it had learned to play an increasingly complex array of 8 bit games to greater standards than any human.
Hassabis is still fairly young, looks younger still, and appears to have lived many remarkable lives of precocious intellectual supremacy: a prize winning chess champion before the age of 10, completing secondary education a year early, and needing to wait a year before Oxford was legally permitted to admit him, he won a place at Peter Molyneux’ early 90s games powerhouse, Bullfrog, where he co-designed, and wrote the AI for, a game called Theme Park. He then turned down an offer of a million pounds to not attend university and continue coding games, because the Oxbridge degree was an essential part of his longer term vision.
Step ahead to the late 2000s. Hassabis - game player, game developer, computer scientist, and now neuroscientist - calmly and quietly explains firstly that his company aims to ‘solve intelligence’, and ‘produce an Artificial General Intellgience’, and secondly the types of modelling approach he thinks will achieve this aim.
Step ahead to 2020. Alpha has beaten the finest Go players in the world, a game recognised to be orders of magnitude harder for computers to master than Chess; it’s also almost unbeatable at StarCraft, a real time tactics game that has become an esports stalwart. Much more critically, it’s solved what may be the hardest game life has to offer: protein folding, predicting what kind of complex three dimensional structure is likely to emerge from a linear chain of amino acids, a problem that had represented a bottleneck on basic pharmaceutical and biological research for generations.
Step forward to 2024. Hassabis is a recipient of the Nobel Prize in Chemistry. AlphaFold has predicted the protein structure of almost every protein ever discovered. Rather than hoarding such knowledge, perhaps equivalent to millennia of human endeavour, DeepMind releases all predictions.
Although it may be OpenAI’s ChatGPT which showed the potential of AIs to the general public, to me the kinds of advances DeepMind have made seem much more impressive, and to offer a much stronger foundation for advancing towards a true general intelligence, based on very specific and measurable goals and aims, rather than in producing the kind of mellifluous sycophantic confabulation that LLMs tend to be (in)famous for. Real general intelligence, whether from ‘artificial’ origins or not, must be about more than the ability to endlessly produce sweet-sounding bullshit and uncanny slop.
Maybe I was too harsh on LLMs just then. My apologies. As penance, here’s a kind of exploratory reverie I had with Claude, which gives the appearance of being comparatively low in confabulation and sycophancy - to help reflect on the documentary:
What do protein folding problems entail? Why has this historically been such a challenge? What are the possible acceleration in scientific advances that should result over the next 1-10 years from Alpha’s work in this area?
The Protein Folding Problem
Protein folding is one of biology’s most fundamental puzzles: predicting how a protein’s amino acid sequence determines its three-dimensional structure. This matters because a protein’s shape dictates its function—enzymes, antibodies, structural components, and virtually every biological process depends on proteins adopting the correct 3D configuration.
What the Problem Entails
The challenge has several layers:
The fundamental question: Given a linear sequence of amino acids (the primary structure), can we predict how it will fold into its final 3D shape (tertiary structure)? Proteins can fold into incredibly complex configurations with helices, sheets, loops, and intricate binding sites.
The astronomical complexity: A typical protein might have 100-300 amino acids. Even if each amino acid could adopt just 3 conformations, that’s 3^100 possible structures—more than the number of atoms in the universe. Yet proteins fold correctly in milliseconds to seconds inside cells.
The physics: Folding is driven by countless weak interactions—hydrogen bonds, hydrophobic effects, electrostatic forces, van der Waals interactions. Simulating all these forces accurately requires immense computational power, and even then, the timescales involved (microseconds to seconds) were beyond reach for most proteins.
Why It’s Been So Hard
For decades, scientists faced multiple obstacles:
Experimental methods are slow and expensive: Determining protein structures through X-ray crystallography or cryo-electron microscopy takes months to years and doesn’t work for all proteins. Many important proteins (like membrane proteins) are extremely difficult to crystallize.
Computational approaches struggled: Traditional molecular dynamics simulations could model only microseconds of folding for small proteins. Ab initio prediction methods that tried to calculate structures from first principles were computationally intractable.
The gap between sequence and structure: By 2020, we knew sequences for hundreds of millions of proteins but had experimental structures for only about 170,000. This enormous gap limited our understanding of biology and medicine.
AlphaFold’s Breakthrough and Future Impact
AlphaFold2 (2020) and AlphaFold3 (2024) have revolutionized this field by using deep learning to predict structures with near-experimental accuracy. Here’s what this enables over the next 1-10 years:
Drug discovery acceleration: Understanding protein structures allows pharmaceutical companies to design drugs that fit precisely into binding sites. This could compress drug development timelines from 10-15 years to potentially 5-7 years for some targets, and open up “undruggable” targets that were previously too difficult to work with.
Understanding disease mechanisms: Many diseases involve misfolded proteins (Alzheimer’s, Parkinson’s, cystic fibrosis). Predicting how mutations affect folding helps us understand pathology and design therapies. We’re already seeing this with personalized medicine approaches.
Enzyme engineering: Researchers can now design novel enzymes for industrial processes—breaking down plastics, producing biofuels, synthesizing chemicals more sustainably. What once required years of trial-and-error can now be guided by structural predictions.
Synthetic biology: Designing new proteins from scratch becomes feasible. Scientists are creating novel proteins for carbon capture, biosensors, and new materials. This could lead to programmable biological systems.
Antibody design: Faster development of therapeutic antibodies and vaccines. Understanding antibody-antigen interactions at atomic resolution accelerates development of treatments for cancer, autoimmune diseases, and infectious diseases.
Basic research acceleration: Biologists can now generate hypotheses about protein function instantly rather than waiting months for experimental structures. This fundamentally speeds up the research cycle across molecular biology, genetics, and biochemistry.
Structural proteomics: AlphaFold has already predicted structures for over 200 million proteins—essentially all known proteins. This creates a comprehensive map of the protein universe, revealing evolutionary relationships and functional patterns never before visible.
The most profound impact may be how this changes the nature of biological research itself. Structure prediction becomes a routine computational step rather than a bottleneck, allowing scientists to ask bigger questions and test more hypotheses. Combined with other AI advances in protein design, gene editing, and systems biology, we’re entering an era where biology becomes more predictable and engineering-driven—potentially as transformative as the transition from alchemy to modern chemistry.
When discussing ‘sites’/‘targets’, is the lock-and-key analogy appropriate, with the challenge being that given the amino acid sequence alone neither the lock nor key structure was previously known?
Yes, the lock-and-key analogy is quite appropriate and captures the essence of the challenge beautifully!
The Lock-and-Key Model
The analogy works well for understanding molecular recognition. Proteins often have specific binding sites (the “lock”) where other molecules—drugs, substrates, other proteins, DNA, etc. (the “key”)—fit with high specificity. The shape complementarity, along with chemical properties like charge distribution and hydrophobicity, determines whether binding occurs and how strong it is.
The Double-Blind Problem
Your insight about the double challenge is exactly right. Historically, drug designers faced a frustrating situation:
Without knowing the lock’s shape, you couldn’t rationally design a key. You’d have to try thousands or millions of chemical compounds hoping something would fit—essentially trying random keys in a lock you couldn’t see. This “high-throughput screening” approach is expensive, time-consuming, and often misses good candidates.
Even when you had a target sequence, you didn’t know what 3D structure it would adopt, so you couldn’t predict where the binding site would be, what shape it would have, or which chemical properties would matter for binding.
It’s like being asked to design a key for a lock when all you have is the manufacturing instruction manual written in code, with no picture of the actual lock.
How AlphaFold Changes the Game
Now with structure prediction, researchers can:
See the lock’s shape from the amino acid sequence alone
Identify binding pockets and cavities computationally
Understand which amino acids line the binding site
Design molecules that complement that specific geometry
This transforms drug discovery from a “blind search” into a “rational design” problem.
Refining the Analogy
That said, the lock-and-key metaphor has some limitations worth noting:
Induced fit: Proteins aren’t rigid locks. They’re often flexible and change shape when the key binds—more like a glove conforming to a hand than a rigid lock. This dynamic behavior adds complexity that AlphaFold is increasingly capturing.
Multiple conformations: Proteins can adopt different shapes (conformational states), essentially changing the lock’s shape depending on conditions. Some proteins have “open” and “closed” states, or undergo large structural rearrangements.
Allosteric sites: Sometimes the best “lock” isn’t the active site at all, but a distant site that, when bound, changes the protein’s shape or function. These regulatory mechanisms add another layer of sophistication.
But fundamentally, yes—you’ve captured the core problem perfectly. AlphaFold reveals the lock’s structure from its genetic blueprint, making rational key design possible for the first time at scale.
Given the suggested decline in drug development from 10-15 to 5-7 years, what are the remaining steps that take the 4-6 years?
Great question! Even with perfect structural knowledge, drug development remains a lengthy, complex process. Here’s what fills those remaining 5-7 years:
The Remaining Timeline
Lead Optimization (1-2 years) Even after identifying a molecule that fits the binding site, you need to optimize it. The initial “hit” might bind well but have terrible properties:
Poor solubility (can’t dissolve in blood)
Rapid metabolism (body breaks it down too quickly)
Toxicity to human cells
Inability to cross cell membranes or reach the target tissue
Off-target effects (binds to unintended proteins)
Medicinal chemists iterate through hundreds of variants, tweaking the structure to improve these properties while maintaining binding affinity. Structure prediction helps but doesn’t solve pharmacokinetics (how the drug moves through the body) or toxicity.
Preclinical Testing (1-2 years) Before human trials, extensive animal studies are required:
Toxicology studies in multiple species
Pharmacokinetics (absorption, distribution, metabolism, excretion)
Safety profiling at various doses
Efficacy testing in disease models
Manufacturing process development
These can’t be fully simulated—you must observe actual biological responses in living systems.
Phase I Clinical Trials (1-2 years) First tests in humans, usually 20-100 healthy volunteers:
Is it safe at various doses?
How does the human body process it?
What are the side effects?
What’s the maximum tolerated dose?
About 70% of drugs fail here despite looking promising in animals.
Phase II Clinical Trials (1-2 years) Testing in 100-300 patients who actually have the disease:
Does it work in humans?
What’s the optimal dose?
What are the side effects in sick patients?
Early efficacy signals
About 33% of drugs that passed Phase I fail here.
Phase III Clinical Trials (2-4 years) Large-scale trials with 300-3,000+ patients:
Definitively prove efficacy compared to standard treatment or placebo
Comprehensive safety data across diverse populations
Long-term effects
Statistical proof of benefit
This is the most expensive phase—often costing hundreds of millions. Only about 25-50% of drugs entering Phase III get approved.
Regulatory Review (1-2 years) Submitting data to FDA/EMA and waiting for approval:
Reviewing thousands of pages of data
Addressing regulatory questions
Inspecting manufacturing facilities
Risk-benefit analysis
Why These Steps Can’t Be Dramatically Shortened
Biological complexity: AlphaFold tells you the lock’s shape, but not:
Whether blocking/activating that protein actually treats the disease in humans
How the drug affects other proteins and pathways (systems-level effects)
Individual patient variability in response
Long-term consequences
Safety requirements: You can’t simulate whether a drug causes liver damage, heart problems, or cancer over time. This requires observing actual patients for months or years.
Statistical rigor: Clinical trials need sufficient patient numbers and follow-up time to prove the drug works and is safe. A drug might show early promise but fail in larger populations or reveal rare but serious side effects.
Manufacturing scale-up: Making grams in a lab differs vastly from producing tons reliably and consistently. This process development happens in parallel but takes years.
Regulatory conservatism: After disasters like thalidomide, regulators rightfully demand extensive proof. The bar is high, and shortcuts have catastrophic consequences.
Where AI Might Further Accelerate Things
Beyond structure prediction, AI could potentially compress timelines by:
Predicting toxicity and drug properties computationally (reducing trial-and-error in optimization)
Identifying biomarkers for patient selection (making trials more efficient)
Optimizing trial design and patient recruitment
Predicting drug-drug interactions and side effects
But the irreducible core—proving safety and efficacy in actual human beings over time—remains. Biology is messy, patients are variable, and rare side effects only emerge in large populations.
That said, going from 10-15 years to 5-7 years would be transformative. It means patients get treatments roughly twice as fast, development costs drop significantly (making more “niche” drugs economically viable), and pharmaceutical companies can tackle more targets with the same resources. The impact compounds across the entire healthcare ecosystem.
Given this, should we expect pharmaceutical companies to grow in value faster than over the last 20-30 years?
This is a fascinating question that cuts to the heart of whether scientific breakthroughs translate into investment returns. The answer is nuanced and probably not as straightforward as you might expect.
Arguments for Faster Growth
Improved R&D productivity: The pharmaceutical industry has suffered from declining R&D efficiency for decades—“Eroom’s Law” (Moore’s Law backwards). Despite spending more on R&D, fewer drugs were approved per billion dollars spent. If AlphaFold and related AI tools reverse this trend, companies could get more successful drugs from the same investment.
Expanded druggable targets: Many disease-causing proteins were previously “undruggable” because we couldn’t understand their structure well enough to design inhibitors. Suddenly having access to 200+ million protein structures could open entirely new therapeutic areas—particularly in rare diseases and precision medicine.
Faster time-to-market: Getting drugs approved in 5-7 years instead of 10-15 years means:
Earlier revenue generation
Longer effective patent life (more years of exclusivity)
Lower cost of capital (less money tied up for shorter periods)
Ability to pursue more projects simultaneously
Better capital allocation: Companies can computationally filter out likely failures earlier, concentrating resources on more promising candidates.
Arguments Against Faster Growth (or Why It Might Not Matter as Much)
Competition intensifies equally: Every pharma company gets access to the same tools. If everyone becomes more efficient simultaneously, the competitive advantage disappears. This could lead to:
More drugs reaching market (supply increase)
Increased pricing pressure as multiple companies target the same diseases
Faster erosion of market share as competitors develop alternatives more quickly
Value accrues elsewhere: The companies capturing value might not be traditional pharma:
AI biotech startups using these tools
Tech companies (Google/DeepMind, etc.) that own the platforms
Contract research organizations (CROs) that offer AI-enabled services
Generic manufacturers who can more quickly develop biosimilars
Regulatory and pricing pressure: Governments and insurers are increasingly aggressive about drug pricing. Even if companies develop drugs faster and cheaper, they may not be able to charge proportionally more. In fact, regulators might argue that lower development costs should mean lower prices.
The clinical trial bottleneck remains: As we discussed, 5-7 years of the timeline can’t be compressed much. So while early-stage productivity improves dramatically, the rate-limiting step (human trials) doesn’t change proportionally. The benefit is real but perhaps more incremental than revolutionary.
Market saturation: Many “easy” and lucrative targets (common diseases in wealthy populations) already have treatments. New drugs increasingly target:
Rare diseases (smaller markets)
Precision medicine (fragmented patient populations)
Diseases more common in lower-income countries (pricing challenges)
The innovation paradox: As medicine advances, we’re left with harder problems. The remaining diseases are often more complex, multifactorial, or involve targets that are difficult to modulate without side effects.
Historical Context: The Last 20-30 Years
Looking backward provides useful perspective:
Pharmaceutical stocks have grown, but inconsistently: Large-cap pharma has roughly tracked or slightly outperformed the broader market over 20-30 years, but with high volatility tied to patent cliffs, pipeline failures, and blockbuster approvals. Biotech has been more variable—spectacular winners but many complete failures.
The industry has consolidated: Much of the “growth” came through mergers and acquisitions rather than organic innovation, suggesting structural challenges in R&D productivity.
Previous technological revolutions had mixed impacts:
Genomics (Human Genome Project, 2003) was supposed to revolutionize drug discovery. Results have been real but slower than expected.
High-throughput screening, combinatorial chemistry, and other tools improved efficiency but didn’t fundamentally change growth trajectories.
Monoclonal antibodies and biologics created huge value, but much of it accrued to a few companies with platform advantages.
The Realistic Scenario
My best guess is moderate acceleration rather than explosive growth:
Near-term (3-5 years): Traditional pharma companies will see improved R&D metrics—higher success rates, better target selection, reduced costs in early stages. Stock performance will likely improve modestly, but much of this may already be priced in by forward-looking investors.
Medium-term (5-10 years): We’ll see the first wave of “AI-designed” drugs reaching market. Winners will be companies that:
Integrated AI tools most effectively into their workflows
Combined computational power with deep biological/clinical expertise
Maintained or built strong data assets (patient data, trial results, etc.)
Adapted organizational culture to new ways of working
This could create differentiated returns—some companies pull ahead while others lag.
Long-term questions:
Does value concentrate (a few winners dominate) or fragment (many specialized players)?
Do tech companies enter pharma directly or just sell tools?
How do regulatory frameworks and pricing policies adapt?
Does AI-enabled drug development become commoditized (eliminating competitive advantage)?
The Contrarian Take
There’s a scenario where pharmaceutical stock returns don’t improve much despite massive scientific progress:
If development becomes much cheaper and faster, barriers to entry fall. More companies can afford to develop drugs. Competition increases. Pricing power decreases. The consumer/patient wins (more treatments, lower costs), but shareholders don’t necessarily benefit proportionally. The economic surplus gets distributed to patients and healthcare systems rather than captured by pharma companies.
This has precedent—many technological revolutions (personal computers, solar panels, LED lighting) delivered enormous societal value but became commoditized industries with modest profit margins.
Bottom Line
Scientific progress doesn’t automatically translate to outsized stock returns. The key questions are:
Can individual companies create sustainable competitive advantages?
How much of the value gets captured by pharma vs. distributed to other stakeholders?
Does increased efficiency get competed away?
I’d expect pharma to grow somewhat faster than the last 20-30 years—perhaps adding 1-3% annually to returns relative to historical performance—but not the explosive multiples you might expect from such a fundamental breakthrough. The biggest winners might be AI-native biotech companies and platforms rather than traditional pharma giants, though the latter have capital and infrastructure advantages for clinical development and commercialization.
The real beneficiaries will be patients and society through better treatments arriving faster and more diseases becoming treatable. Whether that translates to proportional shareholder returns remains genuinely uncertain.
Are allosteric sites essentially upstream sites in terms of causal influence on downstream protein structure, or is a different mental model more suited to understanding this?
The “upstream causal influence” framing is partially correct but misses some key aspects. Let me offer a more nuanced mental model.
Why “Upstream” Works (Sort Of)
Allosteric sites do exert causal influence on function, and there is a sense of information flow:
Ligand binds to allosteric site → conformational change propagates → active site changes shape/activity
This creates a regulatory pathway: the allosteric site acts as a control knob
So yes, in terms of signal transduction and regulatory logic, “upstream” captures something real.
Why “Upstream” Is Misleading
It’s not strictly hierarchical: The relationship isn’t like a linear pathway where A causes B causes C. Instead, it’s more like a mechanical coupling or network of interactions. Think of it less like a signal cascade and more like:
A network of springs and struts: When you pull on one part of an interconnected structure, the strain propagates throughout, causing distant parts to shift position. The allosteric site doesn’t “send a signal” so much as it perturbs the entire mechanical/energetic landscape.
Energetic redistribution: Proteins exist in multiple conformational states, constantly fluctuating. The allosteric ligand doesn’t create new structures—it shifts the equilibrium between pre-existing conformations. It stabilizes certain states and destabilizes others, changing the probability distribution of shapes the protein adopts.
Better Mental Models
Conformational ensemble: Proteins aren’t static structures but dynamic populations. Imagine the protein as existing in multiple shapes simultaneously (like a quantum superposition, though the mechanism is classical):
State A: 70% of molecules, active site “open”
State B: 30% of molecules, active site “closed”
An allosteric ligand binding might shift this to:
State A: 20% (now disfavored)
State B: 80% (now favored)
The allosteric site doesn’t cause the closed state—it already existed—but it changes which state predominates.
Mechanical linkage: Think of proteins like intricate origami or mechanical linkages where:
Different domains are connected by flexible hinges and rigid segments
Binding at one site constrains certain degrees of freedom
This constraint propagates through the structure via physical connections
Distant regions adjust to accommodate the new constraints
For example, many enzymes have domains connected by hinge regions. When an allosteric regulator binds, it might lock a hinge in place, forcing other domains to adopt specific relative orientations.
Energy landscapes: Proteins navigate a rugged energy landscape with multiple valleys (stable conformations) separated by hills (energy barriers). The allosteric ligand:
Deepens some valleys (stabilizes certain conformations)
Raises others (destabilizes alternative conformations)
Changes the relative heights of barriers between states
This shifts the protein’s exploration of conformational space without creating entirely new structures.
Spatial Relationships Matter, But Not Simply
“Upstream” implies linear directionality, but allosteric effects can be:
Reciprocal: The active site’s state can also influence the allosteric site. It’s bidirectional coupling, not unidirectional causation.
Distance-independent in surprising ways: Sometimes allosteric sites are far from the active site (classic allostery), but sometimes they’re quite close—just positioned so that binding causes a localized distortion that propagates. Conversely, distant sites might have weak effects if they’re not mechanically coupled through the protein’s architecture.
Pathway-dependent: The conformational change might propagate through specific structural elements (alpha helices, beta sheets, loop regions) that act as “wires” transmitting strain. Different allosteric sites might use different pathways to influence the same active site.
Examples That Clarify the Model
Hemoglobin (classic example):
Four subunits arranged symmetrically
Oxygen binding to one subunit makes it easier for others to bind
Not “upstream/downstream” but rather cooperative coupling—binding at one site physically strains the interfaces between subunits, making the others more receptive
Kinases (signaling enzymes):
Often have activation loops that block the active site
Phosphorylation at an allosteric site causes the loop to swing away
This is mechanical: the added negative charge creates electrostatic repulsion that destabilizes the “closed” conformation
G-protein coupled receptors (GPCRs):
Have distinct “inactive” and “active” conformations
Agonist binding at one site stabilizes the active state
Inverse agonists stabilize the inactive state
The binding site itself is arguably both the “active site” and “allosteric site”—these categories blur
Implications for Drug Design
Understanding allostery as energetic/mechanical coupling rather than simple causal hierarchy matters because:
You can target multiple states: If you understand the conformational ensemble, you can design drugs that:
Stabilize inactive conformations (inverse agonists)
Stabilize active conformations (agonists)
Lock the protein in intermediate states
Prevent conformational transitions altogether
You can predict effects more accurately: AlphaFold and molecular dynamics simulations can now model:
How binding at one site changes the energy landscape
Which pathways transmit conformational changes
Whether an allosteric effect will be activating or inhibiting
You can find better drug sites: Allosteric sites often have advantages:
Less conserved across protein families (more selective)
Less critical for function (mutations less likely to cause resistance)
Can fine-tune activity rather than completely blocking it
So, What’s the Right Mental Model?
I’d suggest thinking of allosteric regulation as mechanical-energetic coupling within a dynamic conformational ensemble rather than upstream/downstream causation. The protein is:
A flexible structure exploring conformational space
Mechanically coupled across its domains
Constantly shifting between related shapes
Stabilized in particular conformations by ligands that change the energy landscape
The allosteric site doesn’t “control” the active site in a hierarchical sense—rather, they’re parts of an interconnected system where perturbations in one region necessarily affect the stability and dynamics of other regions.
This model better captures the reciprocal nature, the role of protein dynamics, and why drugs targeting allosteric sites can have such diverse effects (activation, inhibition, modulation) depending on exactly how they reshape the conformational landscape.
Considering the famous science video powers of ten, is the sense of awe associated with complex revealed structure at different scales more likely to be induced through consideration of micro scales (such as those at which protein structures are apparent) or macro scales such as those that are the focus of cosmology?
What a beautiful question—it strikes at the heart of how scale shapes wonder. My intuition is that microscales may increasingly rival or even surpass macroscales in their capacity to induce awe, though the type of awe differs in important ways.
The Traditional Dominance of Cosmic Awe
Historically, the macro has dominated our sense of sublime wonder:
The “pale blue dot” photograph
Galaxies stretching across billions of light-years
The observable universe containing ~2 trillion galaxies
Deep time—13.8 billion years of cosmic history
This cosmic awe taps into something primal: our smallness against vastness, our insignificance within immensity. Pascal captured it: “The eternal silence of these infinite spaces frightens me.” There’s an almost spiritual quality to confronting scales that dwarf human comprehension.
Why macro-awe works so effectively:
Visual immediacy: Space imagery is spectacular—Hubble photos, the Milky Way arcing overhead, total solar eclipses
Narrative simplicity: “Billions of galaxies” is conceptually graspable even if the numbers are incomprehensible
Existential resonance: It naturally triggers questions about meaning, purpose, our place in existence
Cultural embedding: Cosmic wonder is deeply embedded in human culture—religion, philosophy, art, science fiction
The Emerging Power of Microscopic Awe
But something has shifted in recent decades. The molecular and quantum scales are becoming increasingly visceral and imaginatively accessible in ways they never were before:
Visualization breakthroughs: We can now see atoms, proteins, viruses, and cellular machinery:
Cryo-EM structures showing molecular machines in exquisite detail
AlphaFold’s entire protein structure database
Real-time videos of molecular motors “walking” along cellular filaments
Quantum microscopy revealing electron clouds
These aren’t artist’s conceptions—they’re data-driven reconstructions showing actual structure and dynamics.
Functional complexity: The micro reveals not just structure but purposeful intricacy:
Ribosomes as molecular 3D printers
ATP synthase as a rotary motor
DNA repair mechanisms that proofread and correct errors
Immune receptors with millions of variants
Enzymatic cascades with feedback loops
This isn’t random complexity—it’s engineered complexity (by evolution, but the effect on our intuitions is similar). We recognize problem-solving, information processing, error correction. It feels designed, intentional, purposeful.
Personal stakes: The micro is literally us:
These proteins are in your body right now, performing functions that keep you alive
Mutations in these structures cause diseases affecting people you know
Understanding them leads to treatments, cures, interventions
You are, in a very real sense, a vast colony of molecular machines
There’s an intimacy to microscopic awe that cosmic wonder can’t match. The universe is indifferent to us; our proteins are constitutive of us.
Different Flavors of Awe
The two scales evoke qualitatively different experiences:
Cosmic awe:
Characterized by smallness, insignificance, humility
“We are nothing; the universe is everything”
Often melancholic or anxiety-inducing
Emphasizes simplicity at cosmic scales (physics reduces to elegant laws)
Distance and alienness—we can never visit those places
Molecular awe:
Characterized by intricacy, sophistication, marvel at engineering
“We are everything; look at the complexity required to exist”
Often exhilarating or empowering
Emphasizes complexity and functional organization
Intimacy—this is happening inside you right now
I’d argue the molecular realm offers something the cosmic doesn’t: comprehensible complexity. The cosmic is vast but relatively simple (gravity, nuclear forces, expansion). The molecular is dense with meaning—information encoded, decoded, processed, acted upon.
Why Micro-Awe May Be Ascendant
Several factors suggest microscopic wonder is gaining ground:
Accessibility: You can explore protein structures on your laptop. You can’t meaningfully explore distant galaxies. The democratization of molecular visualization tools makes this accessible to billions of people in ways cosmology cannot match.
Actionability: Understanding proteins leads to tangible outcomes—medicines, treatments, biotechnology. This creates a positive feedback loop: interest → understanding → application → renewed interest. Cosmology, for all its grandeur, rarely affects daily life.
Educational emphasis: Biology education increasingly focuses on molecular mechanisms. Children grow up seeing animations of molecular machines. This cultivates intuitions about the micro that previous generations lacked.
Technological mediation: We’re getting better at making the molecular world feel real:
VR experiences of molecular environments
Interactive protein folding games (Foldit)
Real-time visualizations of cellular processes
Even art and design inspired by molecular structures
The complexity gradient: We may be reaching the limits of cosmic surprise. We’ve mapped the observable universe’s large-scale structure. Discoveries continue, but the basic picture is established.
By contrast, molecular biology feels pre-paradigmatic in comparison—we’re still discovering fundamental organizing principles. The complexity seems bottomless. Every protein structure reveals new ingenuity. Every pathway connects to others in unexpected ways.
A Personal Anecdote Tells the Story
Carl Sagan’s Cosmos inspired a generation with cosmic wonder. But today’s equivalent might be David Goodsell’s molecular illustrations or the stunning cryo-EM structures winning Nobel Prizes. When AlphaFold2 was announced, I saw scientists literally weeping—not because of the cosmic implications, but because the molecular world had suddenly opened up.
I suspect if you showed someone:
The Hubble Deep Field (thousands of galaxies)
A rotating 3D structure of ATP synthase with annotations showing how it generates cellular energy
…many would find (2) more genuinely astonishing. Why? Because it combines visual beauty with functional comprehension with personal relevance. It’s not just structure—it’s a machine that makes sense, operating at a scale that’s alien yet intimate.
The Synthesis: Different People, Different Awe
Ultimately, this may be personality-dependent:
Cosmological temperaments are drawn to:
Philosophical abstraction
Grandeur and vastness
Simplicity underlying complexity
Existential questions
The sublime and ineffable
Molecular temperaments are drawn to:
Detailed mechanism
Functional elegance
Complexity serving purpose
Practical understanding
The intricate and intelligible
But I’d wager that as molecular visualization improves, as education emphasizes these scales, and as biotechnology makes molecular understanding increasingly relevant, more people will find their deepest scientific awe in the micro rather than the macro.
The Powers of Ten journey is bidirectional, but we may discover that the inward journey—from human scale down to atoms—is ultimately richer in wonder than the outward journey to galaxies. Not because the cosmic is less magnificent, but because the molecular is both magnificent and meaningful in ways the cosmic cannot be.
We are, after all, not made of galaxies. We are made of proteins. And that may be the more profound truth.