We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. — Roy Amara
When we look at the giddy year since ChatGPT, are we looking at a short-term effect, it is just a year after all? Or are we looking at a pattern of events on the other side of the inflection point of a long-run trend?
Or, are we experiencing a compressed time where a decade of change happens in a year, a year in a month? And we’re still seeing, in other words, short-term effects play out?
We’re around ten years into the exponential transition.
It was, roughly, a decade ago when the giants of the oil age were surpassed by silicon stars, Apple and the like, as the largest firms in the world. So too a decade-or-so since deep-learning and data-meets-compute started to become useful. More than a decade since the smartphone became the ubiquitous computing device giving most of us an internet-connected super computer in our pocket. Beyond that, it’s 6-7 years since solar-generated electricity started to become consistently cheaper than fossil-generated electricity. And, let’s not forget, the same time since the cost of sequencing a human genome fell below $1000. Just shy of a decade from the first one-off successful trial of gene therapies. There are now roughly a dozen approved therapies based on CRISPR or CAR-T, and the drumbeat is speeding up.
Each one of those technologies has made remarkable progress since that decade-or-so-ago waypoint. Readers of this wondermissive don’t need to be reminded of what has happened with compute capability and AI performance.
How short is the long run?
So we need to return to the fundamental question: are we seeing effects of the long-run excitement or short-term exuberance-cum-disappointment? Or some combination of both.
Consider the argument that lots of data and compute, rather than hand-crafted algorithmic structures, symbols and architectures, were the key to unlocking the next paradigm of sophisticated machine learning systems. The Peter Norvig paper, The Unreasonable Effectiveness of Data, is fourteen years old. Deep learning’s moment began shortly after. Rich Sutton’s “bitter lesson”, arguing that computation will trump heuristics, is four years old. The Chinchilla Scaling Law which put a theoretical-emprical basis for scaling neural networks is merely a toddler, two years old this spring.
So with Amara’s lens, do we consider this — the data and compute strategy — a fourteen-year, ten-year, four-year or two-year trend?
Let’s go back to the data. Over that ten-ish year period, sequencing the genome has become 10x cheaper. Solar-generated electricity has fallen in price by a factor of four. The cost of lithium-ion batteries has dropped by about 80%. And all with attendant impacts on not just the tech industry, but the energy, transport and healthcare sectors, too.
So when we take stock at the beginning of 2024, it might not be unreasonable to assertion that we’re into the second half of Amara’s picture, the other side of the inflection.
And if we are, we shouldn’t be surprised to see rapid deployments of technologies into our economies. In many cases, we’ve passed both the science and engineering hurdles, to reach the market and commercialisation hurdles.
Transition time
Why do things suddenly go quickly? Why can this acceleration take place?
It’s about feedback loops. In the case of the new technologies, positive feedback loops spin up, and with them rapid improvement in price or performance.
For example, within AI, the systems-wide feedback loop is the improvement in performance from technical breakthroughs increases the utility of the technology triggering a cascade in demand. This drives investment further down the stack - the tooling and, most importantly, the semiconductors. As that market grows, it attracts competition, increases innovation, and further drives down prices. Wider availability of capable technology further stimulates end-user demand, and… stampede.
The same effect occurs with the technologies being replaced. Their advantages are path-dependent, based on their history. Scale is entangled in business processes and technical architectures, and is well understood. But that scale sits on a highly optimised industrial-economic stack, vulnerable to small changes in market share. As soon as part of the market shifts to the newer technology, assets start to lay fallow on the margins, talent flows, new entrants challenge incumbents who in turn struggle with the innovator’s dilemma.
Consider the electric vehicle transition: as consumers move to EVs, more firms have brought out EVs, increasing consumer choice. This puts pressure on the oil and gas business — where I live our closest petrol station became a block of flats a couple of years ago. Automakers put their best teams—and marketing—on the future. At some point, consumers making a buying decision will face the choice of looking forward or at the rear-view mirror to make their decision. A stampede may follow.
Technology transitions take a few years, fewer than you think. Where markets are in the long-run part of Amara’s assertion, the other side of the inflection, we could well be in the steepest part of the adoption curve.
Know nothings
But in truth what do we really know? What can we really forecast at times like this?
I like
’s phrasing: “we are in Knightian uncertainty rather than Bayesian risk.” Knightian uncertainty deals with the unknown and unknowable, where probabilities can’t be meaningfully assigned. Bayesian risk relies on knowable and known probabilities which can be updated with new information.A world of unknowability requires different strategies, from scenario thinking, portfolio approaches, flexibility and adaptability, resilience, judgement and preparedness. These approaches help deal with uncertainty and those situations where it is too expensive or inaccurate to constantly reappraise where we stand.
But this is indeed where we stand. Perhaps President Kennedy said it best when he addressed a crowd at Rice University some 61 years ago:
The greater our knowledge increases, the greater our ignorance unfolds.
This year’s ingredients
I’ve adopted a thematic approach to this year’s horizon scan. In a sea of uncertainty—that sense of the unknowable—I believe a horizon scan will help you make better decisions than a rear-view mirror or any point predictions.
I’ve built on my nearly three decades in industry, the last nine of which build the thesis behind Exponential View and investing in the team’s research capability.
I also listen to a lot of people. And, many of the conversations I had this year have helped inform this year’s analysis. In 2023, I had hundreds of conversations with a wide variety of people, including:
Cabinet-level ministers (including Heads of State/Government) in around a dozen countries in Europe, Asia and Africa.
Leaders at every one of the large AI labs, as well as a couple of dozen AI startups.
Public market investors in Europe, Asia and the US, across every asset class.
Private market investors in Europe and the US, including those involved in credit, infrastructure finance and private equity.
Venture capitalists in Silicon Valley, New York, across the US, as well as in the UK and Asia.
Policy-makers responsible for AI, net zero and energy transition policies in Europe and the UK.
Senior execs (up to the CEO level) of large public firms including big tech, automotive, financial services, media, insurance, luxury, biotech, and industrials.
Leading scientists in AI research, the energy transition, biotech as well as geopolitics and political science at universities like Stanford, MIT, Oxford, Cambridge, Harvard, UCL, NYU, Duke and Yale as well as in industrial research labs.
All of this has been percolated through the small black box neural net that sits in a protective cage above my neck. I can’t explain how it works, but I think the tokens it produces will give help you think and act more clearly over the coming years.
What I left out
We must acknowledge three major trends which I don’t address directly. However, they are important elements in this complex process of systems change.
The first is the growing evidence that climate change, through hotter temperatures, extreme weather and signs of hitting tipping points, is well underway. This will have substantial impacts on the risks people and nations face and how they respond.
Secondly, global demographics are getting complicated. Many nations are aging and shrinking, and others, particularly in South Asia and Africa, are getting younger and larger. This speaks to political and economic challenges. And different trade-offs from exponential technologies, especially AI.
Third, geoeconomics has shifted. Uncertainty is back, baby. Money is (and likely to continue to be) more expensive than it was for the past decade. Geopolitics is breeding a new geoeconomics. The exponential transition is creating not just new opportunities but vast swathes of stranded assets (both natural ones, oil in the ground that will soon serve no economic purpose, and human-created ones like firms, business models and infrastructure that will be leapfrogged or disrupted by the next).
The chart from Blackrock below summarises this new age of uncertainty. Equity analysts are less in concordance than they have been for 25 years. It is all a bit WTF (or as JFK might have said “unfolding ignorance”).
I don’t say much about how time-space compression impact this challenge of unknowability in this year’s horizon scan, although I think this framing is key for thinking through investment decisions. I won’t reprise those arguments here, although some of those ideas are available in 2023’s essays in Exponential View and their fundamentals underpin my 12 themes for 2024.
I’ve also tried to avoid point forecasts, like the launch date of the Apple Vision Pro or whether X will have a change of leadership.
And finally the themes
For this year, 2024, I’ve picked out 12 themes to think about.
Electrifying everything: Continuing progress in the energy transition
The deepening computing fabric: On the selling of shovels
The corporate AI agenda: How quickly will firms roll out AI?
The business model of AI
Compressing time with scientific AI
Smol is beautiful: The allure of smaller technologies
The society of AI: How we will move away from singular models
Helpful humanoids: What to look out for in humanoid robots
The state of insecurity: Geopolitics meets technology vulnerabilities
Democracy’s plebiscite and the epistemic quake
Decentralisation gathers momentum, and it isn’t just about blockchain
Moving at human speed
Subscribers to Exponential View, will receive this analysis in full in their mailboxes. Part one arrives today; part two will be available on Wednesday.
Happy New Year!
Azeem
P.S. As a bonus, I put together a 3-hour dynamic set, which blends nostalgic picks with contemporary electronica, spanning house, techno, trance, electro, progressive house and classical. You could use Deeper Rarebit to see the New Year in.
