Anticipating the wave train of AI
We’ve been poor about trying to predict the real, lasting impact of generative AI.
It’s not through lack of trying: some have talked about rethinking the way our economies run and how we think about our lives, to treating it as an existential risk, to treating AI as a foundational, or general purpose, technology that will change everything.
Oliver Noonan/Associated Press
I’m not above a bit of grand predicting, and I’ll make some here, but it took me a while to realise why all these efforts sat awkwardly for me. We’re used to tech players having a very clear and reasoned notion about the likely impact of their technologies, even if they’re wrong: we want information to be free; we want to unlock the value in owned property — cars, houses; we want to empower everyone on the planet by connecting them to the internet.
All fine, but AI, or in particular generative AI, seems to lack one of these clarifying bumper stickers. It’s vague, amorphous, grand but somehow silly, as if we’ve given the pub mic to the only person in the room who has never done stand-up.
It’s partly, I suppose, because of the nature of AI. We have found it useful for stuff: identifying cats, driving cars, taking better pictures, sorting search results. GAI, meanwhile, is a different beast. In a way we’re struggling for a use case that justifies the vast expense in building and running large language models.
But it’s something else. It’s because Silicon Valley has, for much of the past 20 years or so, been built on the idea of disruptive innovation — that technology will always find a way to do something differently, which will somehow find a way to dislodge an incumbent technology.
And not only have we really worked out what that incumbent technology is when it comes to GAI, we have been too uncritical in our belief in the concept of disruptive innovation. We need, I believe, to overhaul the concept, in order for it to be more useful at this point in our technological progress.
Let me show what I mean by talking about humble transistor radio.
This cute little fella, smaller than most modern smartphones, is often wheeled out as a great example of disruptive innovation, where a newcomer to the scene spots an opportunity to undercut expensive vacuum-tube radios with cheap and cheerful devices that were “good enough” for consumers. Here’s how the founder of disruptive innovation, Clayton Christensen, told the story in his “The Innovator’s Dilemma”:
In the early 1950s, Akio Morita, the chairman of Sony, took up residence in an inexpensive New York City hotel in order to negotiate a license to AT&T’s patented transistor technology, which its scientists had invented in 1947. Morita found AT&T to be a less-than-willing negotiator and had to visit the company repeatedly badgering AT&T to grant the license. Finally AT&T relented. After the meeting ended in which the licensing documents were signed, an AT&T official asked Morita what Sony planned to do with the license. “We will build small radios,” Morita replied. “Why would anyone care about smaller radios?” the official queried. “We’ll see,” was Morita’s answer.
This isn’t accurate, on several counts. Bell Labs, co-owned by AT&T and Western Electric, had been showing off its transistors’ radio capabilities as early as 1947, though use cases and bugs were still being fixed. It’s unthinkable that AT&T had not already thought of the transistor radio — their problem had been finding consumer manufacturers to partner with.
Christensen continues:
Several months later Sony introduced to the U.S. market the first portable transistor radio. According to the dominant metrics of radio performance in the mainstream market, these early transistor radios were really bad, offering far lower fidelity and much more static than the vacuum tube-based tabletop radios that were the dominant design of the time. But rather than work in his labs until his transistor radios were performance-competitive in the major market (which is what most of the leading electronics companies did with transistor technology), Morita instead found a market that valued the attributes of the technology as it existed at the time — the portable personal radio. Not surprisingly, none of the leading makers of tabletop radios became a leading producer of portable radios, and all were subsequently driven from the radio market.
The little gadget that could
In fact, Sony weren’t actually the first to introduce a small transistor radio, nor were they the ones who first recognised the transistor’s commercial potential. Already by 1950 the idea of the transistor radio was being described as the “find of the century” for commercial usage; three years later the first prototype transistor radios were being appearing in the wild, and one or two journalists were already writing about them. In January 1953 freelance science writer Leonard Engel wrote a piece headlined Curtain Lifts on Little Gadget Likely to Revolutionise Radio:
The delay wasn’t only due to indifference. Commercialisation was in part held back due to technical constraints — engineers had yet to master how to mould germanium, a key material, into the special forms required for transistors — but also in part due to the fact that most transistors were “going to the military for secret devices.” (Transistor technology originated in Allied research into radar during the war.)
And while Sony’s Morita may have been quick to recognise the opportunity, the first to market was Texas Instruments, with the launch of the Regency TR-1 in October 1954. They weren’t cheap: adverts at the time showed them retailing around $100 — roughly what an iPhone 15 Pro Max would cost today.
And while cheaper versions eventually brought the price down to $5 or less, Sony’s initial offerings such as the TR-63 cost more or less the same as the TR-1. (See The Transistor Radio in Nuts & Volts Magazine.)
This reality doesn’t fit the myth Christensen heard, and it doesn’t fit the disruptive dogma still powering Silicon Valley.
‘Destroying the social life of mankind’
But it’s not just the chronology of those first few years that is skewed. The focus on disrupting market incumbents misses out so much.
The introduction of the transistor radio, and its rapid spread across the globe, had a profound effect on mass communication, including the building of a mass infrastructure to support the radio’s reach and influence. In 1950 there were 46 countries without any radio broadcasting transmitter at all. By 1960 that number had fallen to 14. In 1950 more than half of the 159 countries surveyed had less than 10 radio receivers per 1,000 people. By 1960 that number had halved. The number of radio receivers in the world doubled between 1950 and 1960 — about half of them in the U.S. (Statistics on Radio and Television, 1950–1960, UNESCO, 1963) . The numbers rose even faster as the Beatles gained popularity: sales in the U.S. almost doubled, to 10 million transistor radios, between 1962 and 1963. (Source: Transistor radios: The technology that ignited Beatlemania — CBS News)
UNESCO had realised early on that radio could be a powerful tool for education — particularly given illiteracy, which was as high as 85% in some countries. (UNESCO Technical Needs Commission, Recommendations of the Sub-Commission on Radio, Aug 1948). War, too, had already convinced many countries of radio’s awesome political power: onetime BBC Governor Sir Ian Fraser had called it “an agency of the mind, which, potentially at least, can ennoble or utterly destroy the social life of mankind” (Broadcasting in the UK and US in the 1950s, edited by Jamie Medhurst, Siân Nicholas and Tom O’Malley, 2016). Indeed, both British and German governments had pushed for cheap radio sets to ensure their propaganda could spread as widely and quickly as possible during the war. ( Source: Television and radio in the Second World War | National Science and Media Museum).
Radio stations were a required target for any respecting coup plotter. Edwin Luttwak, in his Coup d’Etat: A Practical Handbook (1968), prescribed the seizing of the key radio station and establishing a monopoly on information by disabling any others, using “cooperative technicians” where possible. He ascribed part of the reason for the failure of Greek King Constantine II’s counter coup in late 1967, was the fact that the government radio station, Radio Larissa, reached only a fraction of the population because of its weak transmitter and unusual wavelength.
A Wall of Tinny Sound
Transistor radios, in other words, changed the way people got information, what they listened to and their habits. The TR-1 was released in October 1954 into a fast-fermenting musical world: Bill Haley released his Shake, Rattle and Roll the same month. By 1963, the now untethered average American teen listened to the radio for slightly more than three hours per day, according to Steve Greenberg’s piece on the Beatles. Bruce Springsteen believed the power and importance of his transistor radio could not be overstated: “I lived with it, tucked it in my schoolbag during the day, and tucked it beneath my pillow all hours of the night.” Radio became a “medium of mobile listening”, in the words of academics Tim Wall and Nick Webber (PDF), and music producers and song-writers adapted accordingly
Phil Spector, for example, built his “wall of sound”, filling the recording studio with mini-orchestras to create a big fat sound that was then fed through special “echo chambers” (See Lance LeSalle’s answer to the question What was Phil Spector’s ‘wall of sound’? on Quora). Spector’s innovations have fed pop music until today.
News spread faster, more immediate: initial reports of JFK’s assassination in November 1963 were largely heard on transistor radios, in fields, buses and on the street. James Sperber described how he heard from a friend who had smuggled a radio into their school in San Diego that JFK had been shot. Meanwhile the teachers, summoned one by one to learn the news themselves from the principal, decided not to burden their pupils with the news unaware many already knew.
The power of the transistor had been proven: from then on the race was to find more uses and scale the technology. Televisions, hearing aids and computers soon followed. Famously IBM president Thomas J. Watson bought 100 Regency radios and ordered that its computers immediately shift to using only transistors. “If that little outfit down in Texas can make these radios work for that kind of money, they can make transistors that will make our computers work, too,” he told executives. (Source: Crystal Fire: The Invention of the Transistor and the Birth of the Information Age)
The Soviet Union’s secret weapon
All this to say: The most disruption caused by transistor radios had very little to do with dislodging the old expensive valve radio manufacturers. It had to do with seismic shifts in behaviours, of consumption, of movement, of accessibility, of attention. It’s not hard, once you’ve absorbed the above tale, to draw a direct line between the transistor radio with the smartphone — via the Walkman, the CD Walkman, the iPod — and to see a certain inevitability. Miniaturisation and mobility became the bywords of consumption: so much so the Soviet Union earned much needed foreign currency by building the world’s smallest radio, the Micro, in the 1960s. It became the rage in Europe and the U.S., especially after Nikita Khrushchev gave one to Queen Elizabeth II (See: Soviet Postcards — “Micro” miniature radio, USSR, 1965–69 ).
We tend to think of disruption in terms of what it does to incumbents — companies, industries, workers etc. Really we should be looking beyond that: Much is missed if we assume the first wave of a tsunami is the last. The first is often not the most powerful: what follows is a ‘wave train’, the timings between them uneven, including edge waves, bores and seiches. We lack the imagination to predict and the tools to measure these serial disruptions that follow an innovation.
It is, I agree, not easy to think through the potential long-term impacts of a technology, especially one like AI. But it might help for us to at least be informing ourselves about the technology, its capabilities and its limits. I am an avid user of whatever tools I can lay my hands on, including both commercial services and any open source offerings. But we are too vulnerable to binaries: critical rejection or unquestioning embrace, which the transistor radio story shows us is both predictable and unhelpful to understanding its wave train.
Take, for example, hallucination. Loyal readers will know I’m not impressed by GAI’s ability to be honest. Indeed I would argue, 18 months in, that we’re no closer to solving the problem than we were before. I won’t bore you with the details but every tool I’ve used so far has failed basic robustness tests. If we choose to use these tools to create supposedly factual content, we will accelerate our decline into content mediocrity where every press release, every article, every email, every communication that might conceivably have been done with the help of AI will be dismissed as such.
Spikes and spin
Pretending this is not an issue is impossible, but the trick now seems to be to say that it’s just a blip. Despite its efforts to put a positive spin on the results, McKinsey’s recent survey on AI was forced to acknowledge that
Respondents to the latest survey are more likely than they were last year to say their organizations consider inaccuracy and IP infringement to be relevant to their use of gen AI, and about half continue to view cybersecurity as a risk.
That’s putting it mildly. The percentage of respondents who believed that inaccuracy was a risk rose from 56% to 63% between 2023 and 2024, making it their biggest concern out of the 10 or so on offer — the next one being IP and then cybersecurity, all three being deemed relevant by more than half of respondents. And sure, respondents are looking more closely at mitigating the problem but the number is around half of those who consider it a relevant problem (32% last year, 38% this.)
I’m sorry, but that to me essentially says a good chunk of those asked believe the technology is dangerous. And these people are apparently near the coalface. The 1,363 respondents are described as “representing the full range of regions, industries, company sizes, functional specialties, and tenures.” Nearly three quarters of them said their organizations “had adopted AI in at least one business function” and two thirds said their organizations “were regularly using gen AI in at least one function.” So for them to say in those numbers that GAI is making stuff up is the headline.
And yet the title of the paper is: The state of AI in early 2024: Gen AI adoption spikes and starts to generate value. The race to adopt AI, the race to sell AI-related services and consulting (which of course is what McKinsey is doing here) will unfortunately override the annoying voices in the boardroom saying, what are we doing here exactly?
Take the IP thing. There are some extraordinary, creative GAI tools out there — for making images and videos, music (try Udio), sound effects and text to voice (try out ElevenLabs) , but once again it’s inevitable that the companies behind these tools are either bought out by the bigger players or are bankrupted by intellectual property litigation. Those that survive will either leverage their own IP or retreat to a much more cautious approach, producing bland stuff that doesn’t fool anyone. (Just as it’s now relatively easy to spot visual content that’s been created by AI, so some people are already able to distinguish between AI-produced pop, such as Rick Beato’s son. )
So what we are left with? I’m told that LLM tools that use RAG, Retrieval-Augmented Generation, which confine the corpus an AI draws on, is proving useful in some verticals such as helping call centre support staff better help customers. That’s great but I’m not holding my breath. My experience with RAGs has not been a positive one. So I guess we’re left with writing faster, better emails, product descriptions and the like, none of which sound very exciting to me, nor particularly disruptive.
And, indeed, that is the problem. Companies hate to spend money on customer support, and so any money-throwing at GAI will be in the hope of saving more money later. Use cases like this will be the low-hanging fruit, because it’s relatively easy to sell, replacing one underpaid group of human drones with non-human ones.
Underwhelming. GAI may well have to rethink itself to avoid another AI winter or survive a race to the bottom, where every product and service has an AI feature that ends up being as little used as Siri or Clippy.
But if we’ve learned anything from the transistor radio, it’s that we have learned very little from transistor radios.
‘A revolutionary device’
Take for example what happened both before and after its launch. While Christensen’s founding myth of the cheap transistor radio was off, he was right that there was widespread scepticism about the transistor’s potential, with both academia and engineers only gradually being won over:
After indulging themselves for a year or two in the resentful skepticism with which academia generally greets revolutionary new concepts, physicists and electronics engineers gradually warmed to the expansive new possibilities offered by semiconductor electronics. By 1953, The Engineering Index, an annual compendium of scholarly monographs in technical fields, listed more than 500 papers on transistors and related semiconductor devices. A year later there were twice as many. (Source: The Chip: How Two Americans Invented the Microchip and Launched a Revolution — T.R. Reid)
But the transistor radio’s success quickly threw up another problem: scaling. Even as pocket radios were jumping off the shelves in the late 1950s, “the tyranny of numbers began to emerge. Enthusiasm gave way to frustration, even desperation,” according to Reid. The first transistor radios had 2–7 transistors: The problem was finding an economic and reliable way to cluster the thousands, even tens of thousands necessary for the devices which stood to benefit most from the transistor revolution. A Naval aircraft carrier had 350,000 electronic components; a single computer like the Control Data CD 1604 contained 25,000 transistors, according to Reid.
Solving that problem led to the integrated semiconductor circuits of our era, and the rest, as they say, is history. This story arc, however, is left out of the paeans to disruptive innovation, despite it being one we should be familiar with. A great idea comes along and takes a few visionaries (or good business minds) to see how it might be used. That in turn sparks a movement whose expansion is checked only by some limitations of the technology — usually related to scale. Solving those problems leads to even greater expansion, and we are into that wave train of change.
This describes, arguably, the evolution of AI, and even within AI for the evolution of instances of AI, such as neural networks, and, more recently generative AI. At any point someone predicting the long-term impact of the transistor radio might have looked foolish. Journalist Leonard Engel had been writing about the potential of transistors as far back as 1947, but he was a rare voice writing excitedly about the transistor’s potential. Even he was careful about when this revolution would start. In January 1953 he wrote:
A radio small enough to fit into the palm will be available in a few years. Your portable TV set will be the size of a small typewriter. And for the hard-of-hearing there will be a hearing aid the size of a matchbox which will run for a year on a single set of batteries. What’s even more important, air travel will be safer because better radio aids can be found. These are but a few of the revolutionary devices now possible thanks to a recent invention — a tiny device the size of a kernel of corn called the transistor.
He relegated to the bottom of his piece the real visionary:
Dr. E(lmer) W(illiam) Engstrom, head of the RCA laboratories, says that in less than five years transistors have come as far as vacuum tubes did in 20. He predicts that in another few years this amazing little gadget will be used not only in home radios and TV sets, but in all sorts of revolutionary devices that can scarcely be imagined today.
That last sentence stands the test of time, and that offers a helpful blueprint for thinking about AI. Not the self-serving (and misdirecting) noise about the existential threats, not the banal attempts to make customer service, legal research or content creation simpler and cheaper, but about what new things we might do, how society, health, war, creativity, might change and and how it may impact our lives in the same way that little transistor radio did.
