At its best, The Infinite Machine provides an intimate, immediate portrayal of AGI’s inventions. The book documents, in detail, just how contingent and, yes, improbable the development of modern AGI has become. Frontier invention and discovery require both obsession and compromise, resources and intuition, teamwork and visionary leadership.

Though it often gives voice to high-minded aspirations, the book also contains plenty of celebrity spotting for nerds. Hassabis rubs shoulders with scientists such as Geoffrey Hinton, Ilya Sutskever, and Jeff Dean, and commercial leaders, such as Sundar Pichai, Elon Musk, and Sam Altman, and they do things that sometimes border on the unbelievable. That will keep any reader engaged and elicit a range of reactions, from delight to disgust.

Constraint and Obsession

Hassabis’ early life follows a familiar narrative arc bordering on a trope. A child prodigy at chess from age six, he became one of Europe’s best under-13 players. He also experienced both exhilaration and burnout from a childhood diverted into such a high-pressure world. Predictably, he rebelled and developed a thirst for something else.

After finishing high school at 16, Hassabis joined a gaming company. A charismatic, manipulative founder led it and took Hassabis under his wing. Hassabis excelled at coding new game features. Though the experience contains shades of a modern Oliver Twist and Fagin, the episode does not leave a deep scar (to my surprise). Instead, it teaches Hassabis to enjoy the freedom and thrill of entrepreneurship.

After university, Hassabis founded a gaming venture called Elixir Studios. The effort encounters a gap familiar to any software entrepreneur: the one between the conceivable ambition and the computationally feasible. The ambitions brought Elixir into the spotlight, but the gap led to its downfall. After that experience, learning to manage that gap, whether software for gaming or science, becomes a recurring theme for Hassabis.

Hassabis returns to academia to pursue a PhD in neuroscience. That is not a misprint. The foundations of intelligence obsessed Hassabis. He became a lab rat, running experiments related to human memory and imagination. The experience becomes formative. Though he develops a taste for pursuing scientific goals, Hassabis learns that he does not have a taste for the (slow) rhythms of academic lab science. He also learns that he is drawn to modeling intelligence more than merely measuring human displays of it.

Boiled down to its essence, Hassabis reaches his mid-30s, wants to understand intelligence deeply, and feels at home in a gaming entrepreneurial setting. Those desires led to DeepMind’s founding, mission, and organizational form. That also explains why, at the time, there was nothing else quite like it.

Building DeepMind

The Infinite Machine frames the process of discovery not solely as a technical inquiry but as a deeply human endeavor. The underlying impulse remains essentially the same in many settings: the desire to push beyond known limits, to see what happens at the edge of possibility. That, more than any single achievement, is what this book captures best.

That topic takes the book into some highly technical material, and the author anticipates a non-expert reader, albeit an intelligent one. For example, it describes reinforcement learning and deep learning in depth without any math. It additionally provides a marvelous description of the invention of transformers and the variants that followed. All of this, frankly, requires an impressive delivery of intuitive verbiage.

The portraits of inventors in AGI are striking, too. When scientists confront problems with unknown solutions, they differ in how they approach them. Those reflect deep foundations and intuitions, as well as the imprint of faculty advisors, lessons from idiosyncratic histories, and rivalries.

When in doubt, inventors ask a friend for help. Lucky for Hassabis, he has loyal and exceptionally talented friends. The founding and growth of DeepMind is, in part, a story of collaboration among Hassabis, Shane Legg, John Jumper, and the itinerant David Silver. Their relationships and intellectual partnerships, which culminate in the development of AlphaGo and AlphaFold, are one of the other compelling threads in the book.

Speaking of partners, DeepMind’s early years put a spotlight on Mustafa Suleyman, who now leads a group at Microsoft and has penned his own book about the future. Every organization needs a right-hand man who is hungry for success and gets things done. With an unusual background and outlook, however, Suleyman both fits and defies the archetype. You have to read about his origins to believe it.

Hassabis grew up in the London area and wants to stay. The book gestures toward the region’s relative strength in technical talent and its comparative weakness in venture finance, with an emphasis on the latter. Aside from a few wealthy potential angel investors, Hassabis is out of luck when it comes to funding.

DeepMind confronts an inescapable dilemma. The pursuit of AGI requires the infrastructure and funding that only a handful of global firms and locations can provide. Hassabis eventually secured backing from Peter Thiel’s Founders Fund, and additional funding needs led Google to acquire the company in 2013. It is difficult to imagine an alternative path.

Dwell on that. London is not a financial or technology backwater. It makes anybody’s top cities for a world-class finance and technology hub. Yet, it seems to lack the necessary funders with sufficient capacity, patience, and risk tolerance for idiosyncratic, expensive entrepreneurialism. What a sobering observation. I was left wondering how many other UK-based ventures never materialized due to a lack of access to capital, appropriate mentors, and institutional support.

Themes

Though both Hassabis and Google gain from the combination, the book describes in much more detail how Hassabis gains and what he gave up. It also shows that better infrastructure is crucial, but Hassabis’s ambitions coexist uneasily with Google’s implementation of its mission. The book explores tensions over resource control and research discretion, as well as the difference between an organization with an engineering-driven and customer-focused mission and one with a scientific mission.

The book cannot help but compare Google and OpenAI, the latter of which was founded in late 2015. Both confront similar issues over public obligations and corporate missions. It is a massive oversimplification to say they manage them differently, and after all the complications settle, neither looks better than the other. The details defy a sweeping description other than “you cannot make this stuff up.” Again, it is one of the book’s compelling threads.

A question lurks behind all this: why pursue AGI at all? Every leader is fiercely competitive, and the desire to win for its own sake often takes over events. In his reflective moments, Hassabis also harbors a desire to understand the underlying structure of intelligence and, by extension, nature itself.

Among the other cast of characters, high-minded motives for undertaking a moonshot range from helping humanity, building a business, indulging in power, or acquiring gobs of money. On this foundation of influences from the citadel and street, distinct approaches to AGI emerge, both noble and profane.

The book never uses the phrase “the tech oligarchy,” but the shoe fits. In this portrait, Silicon Valley’s decision-makers come across as relentless and confident, moving with agency and entitlement, resistant to any limit on their discretion. The dynamics border on the astonishing and implausible: shifting alliances, governance crises, internal debates within firms that are themselves global institutions. Again, you can’t make this stuff up.

Not far from this topic are questions of safety, or, more precisely, the existential risks of AGI. The book accurately captures contemporary fears and explains why modern Cassandras see signs of things going awry. Relatedly, and once again, a frightening fact emerges: the preferences of only a small number of decision-makers shape safety-related decisions. Readers who take these issues seriously will react incredulously: The fate of humanity rests on the judgment of so few.

Where the Book Pulls Back

For all its expansiveness, The Infinite Machine occasionally drives right up to topics but does not completely cover them. It’s the reviewer’s job to notice. Here are two such topics that relate to core themes.

First, though invention receives attention, the book could have discussed deployment, integration, and adaptation. How does a discovery get transferred into Google services? The reader does not get a good sense of when discoveries shape Google’s services, nor why some work well in practical applications for widely delivered services.

A related question sits just below the surface: Why do some discoveries in this area of computer science fail to deliver commercial payoff? For example, in 2011, IBM’s Watson computer beat human champions at Jeopardy! Yet, during the 2010s, IBM eventually lost billions of dollars trying to commercialize Watson. What is the difference between AlphaGo and Watson? Was it a failure among IBM’s leadership, a flaw in the integration, or the market to which this technology was first applied?

Second, the book omits a crucial player: the National Science Foundation (NSF) in the US and its UK equivalent, UK Research and Innovation (UKRI). The NSF is a federal agency that explicitly aims to fund US science that private industry underfunds but could yield large gains. With just a few bullseye hits, NSF can make up for many duds.

Fei-Fei Li and ImageNet can illustrate what happens when the NSF gets it right. Li’s efforts accelerated breakthroughs in neural networks by Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky in 2012, and eventually led to Hinton receiving the Nobel Prize in 2024. The book accurately gives her credit for her initiative, foresight, and persistence.

What the book does not say: NSF gave Li half a million dollars in 2009, when no one else would. Society undeniably got an enormous return on this NSF investment. Relatedly, UKRI gave Hassabis nothing, and he had to search widely for funding. These could have been an important counterbalancing point to the numerous organizational tensions surrounding public obligations at Google and OpenAI.

A recommendation

Insider accounts with this much richness are rare. Mallaby has a storyteller’s sense of when to follow Hassabis closely and when to pull back to survey the broader landscape. As a result, the book is engaging and often astonishing.

Still, the book falls at the boundary between history and journalism, especially when covering current events in the 2020s. These accounts lack the benefit of hindsight. The Infinite Machine could not possibly have foreseen events such as the explosive growth of coding assistants or the bottleneck in advanced data centers. Read it soon before the latter parts of the book become dated.

Copyright held by IEEE Micro

Decades of experience inform this bit of wisdom. Roy Amara worked as a policy analyst at two think tanks, the Stanford Research Institute and the Institute for the Future. His law highlights a forecasting bias arising from two sources. Excitable humans lose sight of the mundane reasons why technologies have little immediate impact. Additionally, humans also overlook how consequences compound, producing dramatic results over long periods.

Amara made an art of systematic thinking about future possibilities. While a column cannot fully cover all his insights, there is enough room to focus on just the economics, which are rich.

The Short Run: Friction, Not Failure

Start with the short run. When a technology product designed by highly skilled engineers arrives in the world, the world looks nothing like the one in which it was conceived. Inside the lab or the startup, problems are defined, constraints are explicit, and incentives are aligned. Some engineers, especially inexperienced ones, also tend to have a narrow view of both the application of the product and the use.

Outside the lab or startup, a new supply meets the realities of demand. Users show up and, from a designer’s perspective, put assumptions about user behavior to the test.

The early overestimation of impact is not simply a result of enthusiasm. It is a systematic misunderstanding of demand, usually due to an underestimation of economic friction.

These frictions are not dramatic. They are mundane, repetitive, and stubborn. Most users work in organizations that are not optimized for any single mission. They are coalitions of departments, each with distinct metrics, budgets, and priorities, some explicit and some not. Technology that appears obviously superior in isolation must pass through procurement processes, compatibility checks, training constraints, and informal resistance from users with established routines.

Technical designers rarely see this obstacle course initially. The first round of introduction fails for reasons that become obvious in retrospect. If the design team wants to win, the second and third rounds of revision yield to reality

An excellent example comes from the deployment of image recognition to dentists. Dentists don’t mind getting help from algorithms identifying cavities in X-rays, but that is not what sold dentists on the services of one provider in this space, VideaHealth. They succeeded by leaving the lab and meeting demand in all its messy glory.

VideaHealth changed its design after learning that dentists were enthusiastic about colored overlays on X-rays. Why? Because these could be shown to patients. That helped persuade reluctant patients to fill the cavity, addressing one of the biggest problems dentists face every day: patients leave with a diagnosis but without treatment.

It turns out there are rarely any substitutes for trial and error with users. For example, this also shaped generative algorithms and accounts for some of their initial success. OpenAI’s models did not work right away in 2019 or 2020, but successive iterations and deep conversations with beta users led to the inclusion of novel features. By November of 2022, they had refined it enough, and we know what happened then.

A similar process describes the earliest evolution of Coding assistants, which started as a project engineered by GitHub management. Successive improvements have made their results more accessible to less technically adept users. The additional twist is that dozens of firms have entered this market and imitate one another. As soon as one firm learns about a useful feature from its users, many others develop their own versions.

This is not new.

It has always been so. The designers of Encarta, the first (and only) commercially successful encyclopedia on DVD, questioned the assumptions behind previous efforts.

After studying software buyers in retail outlets, they concluded that a potential customer, usually a parent, would spend no more than 3 minutes with the product before deciding whether to purchase it for their children.

The designers did not lament that buyers would not perform a thorough examination. They accepted it as a constraint, designing the pictures, sounds, and videos to grab attention in the store. In other words, they identified the friction and responded to it.

As another example, consider digital music in its early phase. MP3 compression was, by any technical standard, a technically marvelous invention. It reduced file sizes dramatically while preserving acceptable audio quality. From a design perspective, the problem was solved. Music could move across networks.

And yet, in the short run, the commercial system did not reorganize smoothly around that capability. Napster demonstrated demand, but it also revealed the absence of legal frameworks, licensing agreements, and viable payment models.

The design worked, but the market lacked a sustainable path forward. The frictions were everywhere: in copyright law, in record label incentives, in consumer payment habits, and in the absence of integrated distribution platforms. Those are economic constraints.

The development of Advanced Audio Coding (AAC) led to improvements for firms dealing with audio format for the music industry. Standards organizations defined the technical parameters. Alongside them, a different set of actors took on the task of defining how the standard would be monetized, reducing the licensing complexity for standards by assembling patent pools, defining standardized terms, collecting royalties, and distributing them among contributors.

AAC’s technical advantages, combined with a more predictable licensing environment, made it a suitable foundation for Apple’s iPod and iTunes. That did not eliminate piracy, but it did foster legal transactions, which many potential buyers wanted.

Don’t get me wrong. Sometimes these lessons are forgotten. Video codecs such as H.264 and later HEVC delivered remarkable gains in compression efficiency. But the algorithm was only the beginning. The system had to be negotiated. Any engineer involved in this area knows about the frictions, the grind of moving forward with multiple actors pulling in different directions. No participant calls the present solutions friction-free. Frictions have been the central problem in that area for many years.

More to the point, early forecasts are usually not naïve about technology, but they often fail to completely think through the economic frictions. The designers could not fully anticipate the organizational realities of the users they were trying to reach, or the extent to which divergent firm interests would distort outcomes. Firms “get things done” not because they are optimized, but because they are adaptable to these uneven frictions. A new technology must insert itself into that unevenness.

As the first half of Amara’s law predicts, the result is not failure, just a delay due to friction. Who is at fault for the errors about how long it would take? Blame the inevitable but overlooked friction, or those who had incomplete expectations?

The Long Run: Accumulation, Not Prediction

The long-run bias arises from different origins. Ironically, it arises because frictions do eventually get overcome, generating unforeseen consequences. It arises from the inherent challenges of foreseeing the combination of several accumulated changes.

For one, the dispersion of decision-making distorts economic outcomes. No single firm, designer, or planner controls the trajectory of the future. Understanding that dispersion requires a breadth that few possess.

Moreover, dispersion of experience creates traps related to the fallacy of composition, i.e., knowing something about one individual does not provide a representative example for the entire economy. The system is more than the sum of its parts.

Here is an example where we have had enough time to watch it develop in the long run. In 1995, the world was abuzz with excitement over the Internet and World Wide Web. What happened next? The actions of hundreds of thousands of firms, entrepreneurs, and users experimented in parallel.

What resulted? Firms learned from failure, imitated success, and gradually built complementary capabilities. Exaggerated claims were tested against user behavior, and flawed assumptions were discarded. Several generations of startups pushed new products, while some established firms retained their positions. That yielded something no one in 1995 could have possibly designed from scratch, namely, the commercial internet of 2025.

Why was that unforeseeable? Not because of a lack of intelligence, but because the relevant knowledge did not exist in any one place and probably could not. Over time, distributed experiments (and distributed human feedback) generated second-order effects. New business models appeared to take advantage of new infrastructure, such as 4G and broadband. User expectations shifted, and so did legal frameworks.

The Internet began as a narrow technical capability, but over time, it became part of the infrastructure, influencing a broader economic system. This process was slow, uneven, and chock-full of feedback loops. That yielded an economic system with great outcomes, but those were difficult to predict. Anybody who says they predicted this is selling something.

Return to digital music. After the turbulence of the Napster era, it would have been reasonable to conclude that the technology’s impact had been exaggerated. The industry was in conflict. Revenues were declining. Legal battles dominated headlines.

And yet, over the next two decades, a different system emerged. Streaming platforms aligned licensing, distribution, and user experience into a coherent model. The iPod, complemented by iTunes, turned out to be a bridge to a streaming future dominated by Apple Music on an iPhone and Spotify almost everywhere. Payment systems normalized subscriptions. Social features integrated discovery on social media with listening.

The long-run effect was not simply the digitization of music. It was the reorganization of how music is produced, distributed, discovered, and monetized. That outcome was not designed at the outset. It was assembled, piece by piece, through countless decisions made by a variety of participants responding to local conditions.

Now, think about Large Language Models in light of the experience of music. Early expectations are focused on first-order effects: faster computation, better comprehension, and more accurate models. All of that is fine, but the long-run impact comes from how these capabilities change as firms reorganize around these tools, how complementary systems develop, and how new forms of work emerge to support new services. Those cannot be forecast.

In other words, the economics behind Amara’s law teaches us to be humble about the long run. It may well be great, but such a characterization omits many relevant details. Those can only be discovered as the system unfolds.

A Shared Source of Error

It is tempting to treat the two halves of Amara’s Law as separate phenomena. Yet, both errors shared roots. Both stem from a focus on technology itself, rather than on the economic system into which it is introduced.

In both cases, the bias arises from the same community of technologists operating with the same mental models at an early moment. In the short run, they underestimate the economic frictions that slow adoption. In the long run, they underestimate the economic processes that amplify impact.

At the prototype stage, this focus is natural and necessary. Designers must abstract away from messy realities to make progress. Alas, economic life is both a messy swamp and a teeming system full of glorious variety.

Copyright held by IEEE Micro

Pick-and-shovel markets have been around for centuries. Has anything changed much in recent years with firms such as Intel, Qualcomm, NVIDIA, and Amazon? To answer, let’s start with Samuel Brennan, the first West Coast entrepreneur to (literally) exploit the sale of a widely used pick or shovel. Like many West Coast billionaires today, Brennan was, shall we say, prominent but not popular.

Brennan owned two general stores, one in San Francisco and another in Sacramento. The latter gave him access to rumors of a gold discovery on the South Fork of the American River, which he confirmed for himself in the spring of 1848. In early May, Brennan walked through the streets of San Francisco with a vial of gold in his hand, shouting, “Gold! Gold! Gold from the American River!”

Brennan did not spread this news with charitable intent. He had spent the previous month buying up every available piece of mining equipment and related supplies. He resold them at exorbitant markups during the rush he created, reinvested the profits in real estate, and became the first West Coast millionaire.

Aside from legal restraints on unethical monopolization, the modern version of pick-and-shovel economics involves far less gold and much more silicon. Is there any other crucial difference?

Centralized

After being selected as the microprocessor supplier for the IBM personal computer (PC), Intel became the dominant producer of Central Processing Units (CPUs) for personal computers. Based on that experience, one might think that pick-and-shovel businesses are most effective at a large scale, supporting widely deployed products with many applications. After all, large-scale enables an organization to amortize fixed and sunk costs over a larger revenue base.

There is a grain of truth to that observation. Two other leading firms in integrated circuit design, Qualcomm and NVIDIA, also support widely deployed products with many applications. Qualcomm dominates chipsets for digital communications equipment. NVIDIA dominates the market for Graphics Processing Units (GPUs) in hyperscale data centers. They get similar economic benefits from scale, despite different end users and products.

Still, the scale came about through a combination of foresight, skill, and luck. Intel invested in frontier manufacturing and design. Yet, they ultimately needed IBM’s blessing. Qualcomm’s designs led to the invention of Code Division Multiple Access (CDMA). They ultimately required the cooperation of US carriers. NVIDIA survived years in the volatile gaming market. It ultimately needed growth in scientific computing (more on this below).

Once they got large, management used scale to maintain the lead. For example, after it began to lead the PC market, Intel was given the option to purchase frontier photolithography equipment earlier. NVIDIA gets a similar advantage through its partnership with Taiwan Semiconductor Manufacturing Company (TSMC).

Intel, Qualcomm, and NVIDIA all sell to new frontier markets with backward-compatible designs to guide buyers’ upgrade paths. If that is done well and on a large scale, then it prevents any potential rival from gaining the same advantages. Relatedly, scale can arise from selling a standard design that others build around, supporting a broad array of applications. All three of these firms take advantage of that, too.

Don’t get me wrong. There was nothing easy about building such partnerships and upgrading them without losing customers to rivals. Whole courses are devoted to studying such business strategies.

Transitions become particularly challenging with significant shifts in demand. For example, Intel shifted from desktops to laptops, and NVIDIA shifted from gaming to scientific computing. Still, it is easy for a leading firm to get focused on guiding users in its core business and subsequently lose sight of future growth opportunities. There are not enough managerial platitudes in my dictionary (nor space in this column) to describe, for example, the challenges Intel has faced in smartphone markets and developing chips with parallel processing.

Distributed

A different set of lessons about frontier picks and shovels comes from tracing the journey of a new service developed by Amazon. In this case, let’s focus on Mechanical Turk.

The story begins after the turn of the millennium, when Amazon faced an unavoidable headache. As it expanded the scope of its website to include a wide range of products, it faced a vexing product labeling problem. Amazon sought an automated method to resolve simple questions, such as whether “Sony 42-inch LCD TV” and “Sony Bravia KLV-42 HDTV” were the same product.

All such issues were resolved manually. An employee inspected, evaluated, and resolved the open question. Why manually? Because certain simple decisions—like matching, labeling, or verifying information—were trivial for humans but intractable for computers at scale.

Rather than build a massive data-entry workforce, Jeff Bezos and his team sought something he would later call “artificial artificial intelligence,” a system in which humans could perform small digital tasks that machines could not yet perform. This also required new technologies, including APIs and scalable interfaces, to handle large volumes of small-scale human microtasks, such as image object recognition, text transcription, and data verification.

Long story short, the internal service acquired a set of externally facing features. It emerged online in 2005. Stressing the magic of service, Amazon called it Mechanical Turk, and users referred to it as MTurk. MTurk matched users from anywhere on the Internet who needed microtasks requiring human labor with those who could provide them. It was anticipated that thousands of other organizations—researchers, startups, and developers—faced similar gaps between automation and human judgment.

From the buyer’s perspective, MTurk was a new kind of pick-and-shovel. However, the early rollout revealed numerous short-term, startup problems. Worker registration was cumbersome, and payment mechanisms were limited, initially requiring U.S. bank accounts and manual approval. APIs were under-documented, and developers struggled to design effective interfaces. The system also needed better mechanisms for quality assurance and reliability. Another issue was fraud control. Due to these issues, MTurk saw little uptake.

Either out of remarkable patience or bureaucratic indifference, Amazon did not close the service in its first years. MTurk remained a niche tool and far from a commercial breakthrough. Still, it gradually improved.

It took a young professor, Fei-Fei Li, to demonstrate that MTurk could scale to support massive operations. In 2007, Li needed a method for labeling millions of images for what would become ImageNet, a database of images for developing algorithms for image recognition. MTurk unlocked the bottleneck Li had been facing for years. For a time, her project became MTurk’s largest customer.

Most of you know how this turned out. Li and her team labeled millions of images and started a contest in 2010. In 2012, AlexNet placed first in the third ImageNet competition. Its Convolutional Neural Network outperformed the competition by a long shot, and the field of algorithms changed dramatically. This technical breakthrough transformed neural networks into the preferred frontier pick-and-shovel for algorithm development.

Trace the line of inventions: Before 2012, image recognition had limited uses, such as handwriting recognition. Among the earliest applications after 2012, neural networks proved valuable for automating camera focus in the iPhone. All the phone cameras use this approach today. In other words, one new pick-and-shovel (in microtasks) helped create another (in algorithm estimation), which improved a widely used product (the smartphone camera), which ultimately helped smartphone camera users.

Note a crucial difference with the experience in integrated circuits. While MTurk continues to thrive today, it cannot stop others from labelling. Data labeling has become its own distinct industry, with many firms specializing in supplying it. MTurk does not dominate labeling the way Qualcomm and Intel did in their respective markets for decades.

More to the point, sectors based on advancing neural networks today are not compensating Amazon for addressing Fei-Fei Li’s bottleneck at a crucial moment. If you make a living estimating algorithms today, remember to thank Amazon.

Conclusion

What do all these pick and shovel suppliers have in common? Believe it or not, they all possess similar incentives to innovate.

Here is what I mean. It took luck and foresight to get to a prominent place. In each case, the firms could control that investment in innovation, but not the luck. For example, Intel invested heavily in R&D to support its exceptional manufacturing record of frontier designs. Qualcomm invested heavily to figure out how to adapt the logic of package switching to wireless communications and offered an integrated chipset to support it. NVIDIA invested heavily in the Compute Unified Device Architecture (CUDA) to support scientific computing, well before it paid off handsomely with the growth of neural networks. Amazon took risks to develop Mechanical Turk. Only in this example, Amazon did not get a handsome payoff.

More generally, the economics of innovation are the same, regardless of distinct managerial issues. When they received payoffs, stockholders benefited from the inventions because the frontier firm made more money from the core invention and subsequent upgrades. Users benefited, as businesses often enhanced their products and sold them to their customers.

Here is the rub. While these inventive investments are unambiguously good for society, the same incentives extend to any action that helps maintain or exploit the monopoly that arises from such inventiveness, implemented at scale. The same incentives also motivate significant investments in the legal defenses and attempts to corner the market for a crucial input, just as Brennan did in his day.

In other words, viewed with a broad perspective, frontier pick-and-shovel markets will necessarily be dramatically inventive, crucial to the economy (eventually), and occasionally exploitative of their users. That also means these firms will be both prominent and, on occasion, necessarily unpopular.

Copyright held by IEEE Micro

Reiterating what should be obvious, the awards are worth nothing. Their only value is fifteen seconds of fleeting fame on this blog post.

Are you excited? As a preview, this year’s awards include numerous annual awards, such as Best Supporting Actor, Best Fiction, the Dr. Doolittle Award, the Best Meme Award, and the James Bond Award. There are several new awards this year as well. Taylor Swift will win something, and so will OpenAI. Lepidopterists will also be honored, and there will be a moment of remembrance for Skype. Also, please be assured that no prize will be awarded to Elon Musk, as he is disqualified for taking himself too seriously.

With that, let’s start.

As with the Oscars, we begin with the Best Supporting Actor and Actress. This year, the best supporting actor goes to songwriter Aubierre Rivaldo Taylor. He is the creator of the AI singer Breaking Rust, who has a No. 1 country single, “Walk the Walk.” Music has come a long way from the days of Milli Vanilli lip-syncing others’ vocals. Now, everyone plainly admits that there is a false front. There is even an AI-generated image of the artist, who vaguely resembles the Marlboro Man without a cigarette, showing just how much progress society has made. I look forward to the live tour.

For the female category, the award must go to Eline Van der Velden, Dutch comedian and owner of the production company Particle6, which operates its own AI talent studio (yes, a talent studio just for AI), again, underscoring how far we have come in such a short time. Eline created Tilly Norwood, the most polarizing new AI actress introduced this year. Why do so many actors feel threatened by Tilly? For one, Eline gives Tilly an ageless preternatural look that falls somewhere between Wonder Woman (Gal Gadot) and Miss Teen USA. In addition, Tilly can have a British accent or a soft, sweet US-neutral accent. The accent. Yeah, that’s it. Tilly is getting all that attention because of her voice.

The next award is for Best Movie. This category is conventional, but the winner is anything but. The award goes to Javier. Javier is one strange dude. If you missed his films, take a moment on TikTok to watch their weirdness. Javier’s movies have a simple premise: He is from the near future (i.e., 2027), and he wakes up to find everybody gone. Yes, you did not misread. Somehow, he is coming back to our time and showing us what the world will look like after we all pass away from some sort of apocalypse. They are strangely alluring films, even if the details behind the science of time-travel are a tad sketchy. Just suspend disbelief and go with it. His movies primarily feature empty streets in major cities with heavy dust and blowing debris. He also has eight million followers. Yes, eight million is plenty for any cousin of Cassandra, so he deserves an award just for getting an audience. Also, the film editing is strong, which predicts Best Picture. Still, I cannot help but ask about one minor detail: In the future apocalypse, where does Javier get electricity to power the video cameras?

In the age of fake news, there are many nominees for Best Fiction. The winner is a Christmas market at Buckingham Palace. Wait, what? Is this an annual event? Um, no. This never took place. It is all made up. Watch the video. It is so convincing that it lured many tourists to visit the gates to see what was up. As it turns out, nothing was up, except for the changing of the guard and a small pop-up shop in the Royal Mews. That’s it? A fake video to encourage people to go to a makeshift gift shop? *sigh*

Honorable mentions in this category deserve some attention. This goes to the person who maintains a database of AI-generated hallucinations in legal filings. And a special commendation goes to the writer at the Chicago Sun-Times who wrote a feature on the best summer books; many were AI-generated fiction. Hey, all you AI startups out there, are you looking for a business idea? How about a tool that checks to see whether a lawyer cites an actual or fictitious case, or a writer cites an actual or fictitious book?

The Dr. Doolittle Award. Every year, an award goes to the most interesting use of digital technologies in the service of some bit of zoology, which, as a reminder, also includes entomology. In the unlikely event that you need an additional vocabulary word, the winners are a team of lepidopterists from several universities and non-profit organizations. The lepidopterists developed a small antenna and attached it to a monarch butterfly. It is powered by a small solar cell and affixed to the thorax with eyelash adhesive. Read that slowly. This seems utterly impossible. Amazingly, it enables the scientists to map the routes the Monarchs take as they seek warmer pastures in the winter.

Anyway, this got me wondering. What do these lepidopterists do for a second act? I have a suggestion. This team has an opportunity to test one of the oldest scientific principles and puzzles. They can observe a butterfly flap its wings, right? How about this? They should send a team abroad, record a butterfly flapping its wings in China, and then determine whether it causes hurricanes in the Caribbean. That would be a grand experiment, don’t you think?

Every year, an award goes out for the Best Internet Meme. Why? Short answer: Because the internet needs to stay young at heart. Success needs to be celebrated. Long answer: Every year, someone manages to do something that inspires the online collective brain to invent irreverent, creative memes and bring a smile to the world. The world needs smiles, so it is worth our time.

Alright, let’s get to it. This year, the award goes to Astronomer’s former CEO and HR head, Andy Byron, and Kristin Cabot, who inspired others with their KissCam antics at a Coldplay concert. Later details surfaced that Kristin had already separated from her spouse, so everything fell on Andy, who, um, lost his job. That inspired many low-brow but creative memes. Here is a sampling:

An honorable mention goes to the firm Astronomer, which ran an ad with Gwyneth Paltrow, the ex-wife of Coldplay’s Chris Martin. With tongue firmly in cheek, Paltrow became a “spokesperson on a very temporary basis” for Astronomer and made a short, irreverent commercial. She thanks everyone for the surge in interest in Astronomer, briefly highlights a few of the questions they have been receiving, such as “How is your social media team holding up?” and assures them, “Yes, Astronomer is the best place to run Apache Airflow.” It is brilliant.  It is worth a minute of your time.

The Bear Necessity Award. While we are on the topic of preserving the irreverent spirit of the internet, we need to give an award to Mike Fitz, former park ranger at Katmai National Park, for initiating Fat Bear Week. Why? Because it is overdue. Now in its eleventh year. Fat Bear Week is a pointless, silly contest. Park rangers maintain cameras at Katmai, a national park in Alaska, so the entire world can waste time watching bears feeding on salmon. To be sure, it is a bit gory, gruesome, and grizzly to watch a salmon in its last moments being torn apart by a large wild bear. Still, setting that aside, the whole point is to watch these animals safely from a distance and revel in nature. Yep, let’s watch their instinctual impulse to fill up before hibernating, much like a wayward uncle at Thanksgiving. It is hard to look away. This year, the winner is a bear named “Chunk”. He weighs 1200 pounds. (This is an estimate. Nobody put him on a scale.) He has a broken jaw, and no one is quite sure how he got it. He deserves to win for playing through the injury. Here is a picture of him. Whoa. That is not a dude you want to meet on the front line of Chicago’s football team.

Young Frankenstein award. It’s alive! It’s alive! Is this a comedy, a tragedy, or maybe a slow-motion train wreck? What genre is it? A hero action movie, a drama, a horror flick, or perhaps a parody of a dramatic horror flick? And to announce the award, here comes Frau Blucher (neeeeeeeighhhhhhh).

This year, the award goes to OpenAI. Do you really need an explanation? Alright, let us count the ways. First, OpenAI finally completed its restructuring of its complex hybrid governance model. It is a monster, and so legally and managerially convoluted that only one aspect matters: Sam Altman retains complete discretion. Second, OpenAI has convinced many investors to pour hundreds of billions into building massive data centers, which has generated a widespread fear of missing out among the financial elite, and may cause a monster financial bubble. (Look, yours truly, ever the techno-optimist, believes the future of frontier AI holds incredible promise. However, that is not a license to pursue monstrously bad economics.) Finally, we have to mention the monster one-billion-dollar deal with Disney. OpenAI will pay Disney $1 billion, allowing Sora users to create 30-second videos featuring Disney characters without violating copyright. This deal includes several restrictions to ensure the resulting content is tasteful. Why does anybody think this deal will just generate anything other than a bunch of tasteful, legal, mediocre slop? It all seems like a wish upon a star.

The Mamma Mia Award. Taylor and Travis are getting married! Woohoo! The country’s most eligible football player and most eligible mega-popstar are engaged! Taylor and Travis received this award for their Instagram post announcing their engagement, which garnered more than 31 million likes. Good for them for finding love in this crazy world, and for allowing 31 million to share in it. What a fairy tale, and just perfect for a podcast as well as People Magazine. Here is the key point: there are many songs to choose from for their walk down the aisle and first dance. Because Taylor has completed her music buybacks this year, she can now play both the old and new versions of a song! Taylor definitely should play her latest hit, “The Fate of Ophelia.” It is not very subtle, as Taylor appears to be thanking Travis for helping her avoid a nunnery. Still, it is a catchy tune. But I digress.

Back to the key point: There has not been a wedding this culturally focal since Queen Victoria’s daughter, Princess Royal (yes, that was her name). That wedding introduced the Western world to “Here Comes the Bride” by Wagner, and after that, it became a staple of weddings for the next century. By that benchmark, the globe is about to experience one of the most significant cultural touchpoints: Taylor Swift will choose songs for her own wedding. It will be added to a Spotify playlist and become the new standard for brides worldwide.

Think of the improvements this will bring to so many weddings. After that silly movie, Mamma Mia!, everybody got stuck on ABBA songs at weddings. Look, it is ok to play “Dancing Queen” during the reception, but please, anything else is too much. Who thought it was a good idea to play “I’ve been waiting for you” and “I do. I do. I do. I do?” Please. Anything else. I am so looking forward to learning Taylor’s choices.

The James Bond Award is an annual award, but this year it could have been renamed for Casino Royale, after the 1967 misfit of a parody of a James Bond movie. The award goes to The Atlantic editor Jeffrey Goldberg, who published a story about being accidentally included in the Signal chat as US government officials discussed sensitive details of the bombing of Houthi positions before the bombings took place. What was more foolish – to have the conversation in front of the editor of The Atlantic, or to have this conversation about top-secret matters on Signal at all, even though it is not secure? There is an old saying. When a clown moves into a palace, he doesn’t become a king. The palace becomes a circus.

Alright, it is time for what seems like an annual award for making sassy fun of something in the US government. This year, it is called The Ministry has Fallen Award, named for a scene from Harry Potter. Explanation: Long ago, cryptocurrency had fallen far short of the ideals envisioned by libertarian techno-idealists. Credible reporters who do their homework have shown that a big corner has become a currency exchange cesspool, becoming the primary tool for money laundering by those wishing to escape scrutiny, particularly the North Korean government, Asian-based drug traffickers, and Russian Oil ships evading international sanctions. Any sane government would be finding ways to discourage such activity. But, of course, the sane ship sailed. The US government has completely taken its hands off of crypto. Why? Because meme coins. This would be comic if it were not so tragic.

In so many words, it is now possible to buy some Trump-meme coins. This is a new business for the US royal family, created after the election, and owned entirely by them. How does a meme coin work? It actually doesn’t. It is not possible to purchase anything with them except a politician. And it’s a rigged business: any transaction with the coins generates fees that go directly into the owner’s pocket. Here is the best part: none of the purchases can be tracked or documented. That makes meme coins ideal for untraceable favors, donations, and gifts, particularly from lobbyists representing foreign governments. This has been widely reported, and the best estimate so far is that the family has earned over $700 million in fees and holds several billion in virtual assets. Where is the humor in this? There is an old sarcastic joke: If you think this is an honest business, I have a bridge to sell you.

The funnier question is, “Who put that kind of money into these meme coins?” No one knows for sure, but nobody is guessing that the MAGA faithful put up hundreds of millions of dollars. Can we say, “Favors for sale?” Hmmm, let’s see. Trump has made many suspicious pardons, including Binance’s founder, Changpeng “CZ” Zhao, who was convicted of money laundering in his crypto exchange. He now helps the family with their meme coin business. What a coincidence.

As the saying goes, there is a fine line between farce and fecklessness.

The previous award was all sass and sarcasm, but a bit of a downer. We need an award with a pick-me-up, especially some absurd comedy. This next award is called the Life Imitates Art Award. In this case, the art is Andor. Season 2, which came out this spring on Disney+. Season 1 is sort of slow, but Season 2 is the best writing ever in the Star Wars Universe. The series focuses on the small acts of human resistance against advancing, bureaucratically imposed, militarily enforced, suffocating fascism. The writers downplay the Star Wars mythology and emphasize the daily human struggles of people who gradually lose their freedom to a heavy-handed government. The series makes no attempt to relate it to contemporary events. (I swear. Really. Nada.) But life has a way of imitating art, which is why the newly named award goes to the individual in the Subway Sandwich viral video.

Some detail. His name is Sean Charles Dunn. He is a former Department of Justice paralegal. Frustrated by all the new law enforcement in DC, he threw a Subway Sandwich at a Federal officer in August, just after they invaded, er, showed up to, um, assist with local policing. All of this would have disappeared into the abyss of trivial events, except that it was captured on video and went viral. Nobody with good sense could calm things down, and prosecutors felt they had to bring charges. From the outset, the prosecution looked like overkill, and it generated lots of online memes about how any grand jury would indict a ham sandwich. Except that the grand jury did not. The prosecutors could not convince a grand jury to issue a federal indictment. Why? Because, um, ridiculous. Really. Who is threatened by a flying sandwich? The crust was not even toasted, and the officer is wearing body armor. Egged on by the ravenous few and not content to stop there, the prosecutors reduced the charge to a misdemeanor assault, and Dunn stood trial in November. You might have thought this was a straightforward case, since the prosecutors have a video of the alleged assault. Well, no. The jury acquitted Dunn. Why? It is worthwhile to repeat this again. Because, um, ridiculous.

The lesson? Gosh, it was a good idea to enshrine in the Constitution the right of a suspect to a jury of peers. Smart dudes, those founders. Lin-Manuel Miranda ought to make a musical about them.

Are you still reading? Thanks for such endurance! Time for the last award, which recalls those who were once with us. We honor them with the Sunset Blvd Award. This year, the award goes to Skype, which was officially retired in May, more than twenty-two years after its release. The younger generation has no idea what this is. Almost everyone of a certain age, however, remembers their first Skype video call with some degree of sentimentality. It was magical. Not endlessly magical, but more like that first boyfriend or girlfriend who, in retrospect, did not quite work out. To be sure, there were moments of joy and fleeting excitement, especially at the beginning, but then communication started to drag. Too much trouble warming up the conversation, static responses at random moments interfered with getting a rhythm, and then, when you most depended on them, it was silence and unresponsiveness. To quote a famous line, “To fall in love with you, that was the idiotic thing.” It is behind us now. We have teamed up to meet the moment and zoomed in on better communication. After all, good communication is the key to a satisfying relationship.

This is the end, my friend. You have completed this year’s review. Thanks for reading until now. Have a wonderful year, remember to retain your sense of human decency, and please do not forget to occasionally stop, smile, and laugh at all the silly things people do with digital technology.

Before we fully embrace this new era, let me suggest that we review lessons from the most recent era of technology adoption. In particular, let’s focus on consumer computer technologies (CCTs) – the mix of the mobile ecosystem and the widely used internet, enhanced by Web 2.0.

Some research partners and I recently analyzed returns from co-invention in CCTs. By co-invention, we mean the invention of new applications by firms that utilize CCTs in their business. That investment typically involves developing business processes and practices to complement the adoption of CCTs and support the introduction of new services and business models.

This column provides a brief overview of the study’s key points. Some of the implications arising from these points might seem obvious, but it surprised us that all the consequences would stem from the same framework. For more information, please see the reference at the end of the column.

Framework

Consider a straightforward model of the two types of projects undertaken by firms, where one utilizes the new technology as an intermediate input and involves incremental co-invention. The other also uses the technology as an intermediate input, but requires something more ambitious and innovative for the firm. Referred to by many names in everyday speech, for brevity, we will label it as novel co-invention.

Incremental co-invention in CCTs was exceedingly common. That is because, when faced with a low-cost expansion of existing services, most firms choose to invest in it. For example, should a firm build an app? This was a comparatively straightforward economic decision, especially when the benefits were directly measurable in terms of web traffic that enabled a marketing funnel into sales leads or ad revenue. Accordingly, most firms did build apps.

Novel co-invention differed. Success was difficult, costly, and rare, but you knew a successful novel co-invention effort when you saw it. Many users found the service compelling. It established a new category of use or altered leadership in an existing end market. Think of Netflix in its early years, or the elaborate efforts required by your pharmacy to send you text reminders to refill a prescription.

For a firm involved in novel co-invention, adopting CCTs was usually less straightforward, often because the payoffs were several years out from the initiation of investment. For example, when Netflix first began transitioning from mail-order DVD rentals to streaming, it faced considerable uncertainty about how to do so at scale. It took them more than seven years to find a predictable and reliable process.

In an entrepreneurial setting, the economic constraints bind in different ways. For example, should an entrepreneur build an app for a use case that no other firm covers? Consider Snap, which introduced ephemeral messaging with Snapchat. This novel social experience drove rapid adoption among young users, creating network effects that fueled explosive early growth before competitors introduced similar features. Indeed, Snap reached a place that few entrepreneurs ever attain. Alas for them, their time at the top was brief.

Here is my point: novel co-invention was risky in the CCT era. Private returns were uncertain. Yet, as we all know, many firms did build these. Some succeeded, and most failed, sometimes spectacularly.

Implications

What does all that imply? For one, comparisons between investors in incremental co-invention should reveal a specific pattern of returns, one proportional to the pre-existing customer revenue.

Here is what that means. If one observes a single industry as it evolves into an online market over time, and if every firm made incremental changes, then everybody would have received a 2%, 3%, or 4% return, and nobody should miraculously gain a 50% return on their website.

We tested that prediction. We rolled up our sleeves and collected information from the Internet Archive about the online experiences of terrestrial radio and newspapers, in each case in 2013. We compared that with their listenership in 1993, twenty years earlier. In radio, we were able to track the experiences of over 2000 stations. In newspapers, we examined approximately 100.

The framework predicts that radio stations with more listeners in 1993 would have a larger online audience in 2013 than those that started with fewer listeners in 1993. That is what we see. In other words, radio appears to be an industry without any novel successful co-invention as of 2013. We make a similar prediction for the audience size for newspapers, but see a slightly different outcome. Newspapers with larger circulations gained a more than proportionate online readership compared to those with smaller circulations.

Compare incremental and novel.

Here is another implication: Incremental co-invention should be as widespread as the businesses that use it. In contrast, the returns on novel co-invention should be skewed with only a few big winners and many losers. Novel co-inventions should also be tied to the labor market for technical talent or industry specialties.

Testing this contrast is no small task. It requires an examination of all industries. We rolled up our sleeves again and found a way to collect data on private returns for 2400 leading online properties in 2013, roughly split between smartphone apps and websites. Through extensive (and sometimes tedious) work, we classified all of these into incremental and novel co-invention efforts that led to their production. Some of these industries are so new that we also had to build a new product classification to capture the main categories in the data.

Figures A and B illustrate the distribution of value creation across firms. These should differ between incremental and novel innovations. Indeed, we find that it does. Incremental co-invention grows from firm assets that support existing business, while novel co-invention grows from an entirely different origin. Local labor markets with appropriate technical and commercial domain knowledge supported investment in novel uses of CCTs.

We then examined a related, open question. There is no theoretical reason why the total value created by novel co-inventors should be bigger or smaller than that made by incremental co-inventors. A small number of big hits could exceed the total of a large volume of small contributions, or vice versa. Only data can settle the question.

As it turned out, the aggregate private value created by incremental co-invention is smaller than that made by novel co-invention, and by a considerable amount. Empirically, we estimate that incremental co-invention accounts for approximately 6% of the total observed value among the top 2,400 firms. If every bit of value we do not observe in the long tail of small firms is attributed to some incremental investment, it still accounts for less than 18% of the total value.

In other words, novel co-invention drove the vast majority of new private value creation in CCTs, and in a small set of industries where novel co-invention thrived; like it or not, Google, Facebook, Netflix, Amazon, and other top hundred CCT-based companies owned a significant portion of the profits from adapting CCTs.

Figure A: Incremental Co-Invention Distribution Across Industries

Figure B: Novel Co-Invention Distribution Across Industries

To be clear, sometimes these companies came up with the novel co-inventions themselves (e.g., the Google Search Engine and Ad network), and sometimes they bought parts of it (e.g., Google bought YouTube after they tried and failed to develop something similar), absorbing it in the big company. Sometimes these received enormous investments, and sometimes these purchases went nowhere.

That result raises questions about why novel co-inventions sometimes originate in entrepreneurial firms and sometimes in leading firms. That is a bigger topic for another day. Today’s column focuses on the lessons from characterizing these fact patterns.

Concentration

There is one more subtle prediction. If all co-inventions were incremental, there would be little impact on competition and market structure. Similarly, the industrial and geographical distribution of economic activity would change little, as they are based on the existing choices of firms and their existing assets. However, novel co-inventions should not be widespread. It should concentrate on a few industries and in a few places.

As expected, four related media and entertainment industries, as well as retail trade, collectively account for 97% of the value from novel co-invention, with targeted advertising being the most significant. Most of the manufacturing and several additional significant sectors, such as healthcare, are absent.

To be sure, incremental co-invention is more widespread industrially. However, it is also less impactful in creating value.

Novel co-invention should also concentrate geographically. To check this, we conducted another forensic investigation into the location of each product’s co-inventing firm’s headquarters.

Again, as expected, we find that the location of incremental and novel co-invention differs.

Because incremental co-invention builds on existing business, it is geographically distributed widely, following the existing location patterns of the co-inventing firms. In contrast, the geographic outcome of novel co-invention differs sharply, with four prominent regions emerging for novel innovation: the greater San Francisco region, New York, Seattle, and Los Angeles.

San Francisco and Seattle are no surprise, as they have long been centers for technical talent, and many firms there zig-zagged their way into media and entertainment markets. New York and Los Angeles firms had the opportunity to join them during this technology wave because their cities possessed the entertainment industry’s human capital, and many firms there had to zig-zag their way into technology.

Conclusion

We are not the first commentators to notice the bifurcation of results in the commercial internet. However, most commentators emphasize the winner-take-all competitive dynamics that result in a small number of massive providers.

We took another approach. You might call it reductionist for eschewing elaborate explanations about why things happened, but simple has advantages too. It boils results down to either/or and incremental/novel, making it easy to observe. If that trend continues into the new AI era, expect to be able to track project outcomes this way.

Reference
[1] Tim Bresnahan, Shane Greenstein, and Pai-Ling Yin. 2025. “New Economic Forces Behind the Value Distribution of Innovation.” National Bureau of Economic Research, Working Paper 34090. https://www.nber.org/papers/w34090

Copyright held by IEEE Micro

October/November 2025

His work populates the earlier chapters of textbooks on the history of economic thought, where he is credited with being a pioneer in starting the Neoclassical School of Economics.

Yet, on January 27, Nadella, posted a statement on social media: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

Why is Microsoft’s CEO posting this statement, and why at that moment? While he has warmed the hearts of whoever taught him economics as a student, that surely could not have been his primary purpose.

Indeed, why? This column provides an answer. Understood in context, Nadella is making a profound economic remark about the Chinese firm DeepSeek and the future of Artificial Intelligence (AI). To be sure, the profound parts are not obvious. This will take some explaining.

Running out of Coal

There were many parallels between Jevons’s era and today’s. Today, there is a public debate about the lack of supply of Graphics Processing Units (GPUs), the high-performance chips that are the preferred platform for executing frontier algorithms used for estimation. In Jevons’ lifetime, public debate focused on the potential shortage of coal, the fuel for steam engines. Just like today, there also were technological optimists and economists, and, furthermore, they came to different conclusions despite starting from the same facts.

More precisely, Britain had a lot of steam engines and abundant coal, but by the latter half of the 19th century, many observers worried that Britain could quickly run out of enough coal to fuel the fires of all its steam engines. Ironically, in retrospect, the technological optimists did not forecast how technological improvements would change the situation, namely, that large-scale oil refining and the invention of internal combustion engines would alter society’s energy usage. Instead, technological optimists argued that more efficient steam engines would save on coal use.

Restated in modern terms, the rate of technological improvement in the previous half-century had been miraculous, displaying improvement to steam engines that look equivalent to today’s Moore’s Law. These optimists projected that steam engine technology would continue to improve at the same rate. Like analysts today who forecast by drawing lines on a chart into future decades, the optimists forecasted another half-century of rapid gains in efficiency, and argued this would alleviate any coal shortage.

While the optimistic argument sounds plausible, Jevons pointed out the flaw in the economic reasoning. A more efficient use of coal would lower the cost of producing a wheel turn from an engine. As the cost per wheel turn declined, industrial firms would find more uses for a turning wheel, limited only by their imagination for finding additional applications. However, more applications would lead to more demand for engines in turn.

In other words, under any scenario, increasing the use of engines would work against the efficiency gains, producing not nearly as significant a decline in coal demand as the optimists forecast. Jevons went further and made a second point using the same economic reasoning. Namely, it was possible that the demand for more steam engines could grow so much that it would result in more coal use. In other words, more efficiency would not save any coal.

Let’s rephrase Jevon’s insight to heighten its generality. Jevons argued that creating more efficient input technologies reduced the per-unit output cost. That would translate into lower prices, generating more demand for final output. There is no reason why the decline in input demand from more efficiency is larger than the increase in input demand from more output demand. Net input usage could plausibly go up or down.

As an aside, please forgive a clarifying but pedantic semantic observation. Though “paradox” is a memorable label for Jevons’ insight, it is a misleading. A paradox is defined as a seemingly self-contradictory statement. Jevons did not reason his way to a paradoxical insight. It is more accurately labeled a counterintuitive conclusion.

AI and Coal

Now for the metaphorical payoff. What do coal usage and efficient steam engines have to do with AI?

Nadella wants his readers to use the same economic reasoning but replace the nouns. Replace “coal” with “GPUs.” Replace “execute more wheel turns” with “execute more algorithms.” Replace “more efficient use of coal in steam engines” with “more efficient use of GPUs in data centers.”

In other words, more efficient use of GPUs will lead to more AI algorithms being executed, but that does not necessarily mean it will lead to less demand for GPUs. It could lead to more.

That implies numerous generalities. If a clever firm such as DeepSeek invented more efficient AI software, would it lead to less use of GPUs? What if a dozen firms show up next year with imitations of DeepSeek’s clever approaches? Would that lead to less use of GPUs? Think a bit before answering.

Taking a lesson from Jevons, do not presume demand for GPUs will decline with advances in efficiency. More generally, do not be too quick to conclude that creating a more efficient execution of estimation algorithms on GPUs will lead to less GPU demand in the future.

Seen in context, Nadella was also arguing something subtle. By alluding to Jevons’ paradox, he wants his audience to stop predicting with certainty that GPU usage will decline significantly due to innovations by firms such as DeepSeek and all the imitators that will soon follow. Even if GPU usage declines, it plausibly should fall only a little, and perhaps not at all. More to the point, the value of GPUs might still be high.

What is going on?

Jevons’ paradox tells us not to overreact to efficiency gains. That is what motivated Nadella’s post. Why did he post it on January 27?

On January 20, DeepSeek revealed its second Large Language Model (LLM), R1, after revealing another called V3 in the prior month. V3 and R1 both met some audacious benchmarks. Moreover, DeepSeek claimed to have made both for cheap. By the time of Nadella’s post, DeepSeek’s achievement had begun morphing into a Rashomon example in which every observer could view another angle and comment on another aspect of the AI market. Panic had set in financial markets.

Against that cacophony of comments, Microsoft’s CEO wanted events interpreted in favor of his firm. Microsoft has invested tens of billions of dollars into frontier infrastructure to support LLMs, both for itself and as part of a partnership with OpenAI. Yet, DeepSeek’s achievement led many to panic that these investments no longer had as much value.

There were at least two related but separate panics. The first panic concerned the merits of the announcements from Microsoft, Google, Meta, Apple, Amazon, Oracle, and other large firms, not to mention the Venture Capital community. All have spent vast sums on data center infrastructure, and all plan to spend more. Should the assessment of the value of these plans be revised downward?

The second panic concerns the financial valuations of the firms that supply frontier AI infrastructure for data centers, principally for the silicon stack. Should the forecasts for sales of Nvidia, TSMC, ARM, and ASML be revised downward?

DeepSeek’s achievements upended many market analysts’ assumptions, showing that more efficient use of hardware was possible when estimating an algorithm that imitates a frontier algorithm. Nadella wanted them to rethink their logic. More efficiency does not necessarily lead to less GPU sales.

Conclusion

I am always surprised by panic among experienced technology market analysts. They have lived through massive changes in the last half-century, brought about by the revolutions linked to the deployment of the PC, Local Area Networks, Internet, Web, and smartphones.

By now, experienced analysts should know better than to panic, no matter the news. They should keep those forecast wheels turning, going full steam ahead.

As it happened, since Nadella’s post, a couple additional factors have come into focus. These addenda further reduced the panic. Let’s enumerate three.

First, GPUs are helpful for executing algorithms for estimation but also essential for inference. If execution becomes more efficient, then more uses will be found for the algorithms, and more GPUs will be deployed for inference. If anything, that leads to increasing demand for GPUs for inference, just not for estimation.

Second, DeepSeek allowed users to see the algorithm’s reasoning, an appealing, innovative product design feature. That feature is easy to imitate and should generate many variants in new features among the established players. New features should raise the appeal of algorithms in the future.

Third, DeepSeek delivers censored results, reflecting the Chinese government’s preferences for an information product. Many users do not care, but some have tried DeepSeek, experienced the results, and declared they are uninterested in using such a censored product. That will reduce demand outside of China, reducing some of the anticipated competition with the established players.

More to the point, concerns about innovation and economics are at least a century and a half old. Costs always come down eventually, albeit unpredictably. Sometimes, innovation arrives slowly, and sometimes in bunches. Occasionally, somebody comes along and makes massive leaps in achievements. The timing cannot be gamed.

A voice from the past could teach the present to look beyond most short-run events and recognize additional long-run effects linked to economic factors. That is the essential lesson of the Jevons paradox.

Copyright held by IEEE Micro

April 2025

This model is misleading in numerous ways that would take a book to enumerate. Today’s column focuses on step two, aiming to one piece. How and why do businesses embed the invention in products and services while the invention spreads?

Cataloging the factors that help businesses embed inventions can yield valuable lessons. Today, we focus on knowledge spillovers, eliminating bottlenecks, and inventing ways to invent.

Spillovers of Knowledge

A spillover is a lesson learned for free or little cost, typically by intelligently watching and imitating somebody else. Importantly, no formal contracting or intellectual property governs how knowledge transfers from one individual to another.

Consider the invention of viral marketing, a low-cost approach to acquiring customers at the outset of the commercial web, as an illustration of spillover’s role. Beginning in mid-1995, several entrepreneurs built a browser-based email system. The idea for viral marketing emerged during a conversation between the founders of Hotmail, an early browser-based email provider, and their VCs, Tim Draper and Steve Jurvetson. By embedding an HTML link in an email, suggested by the VCs, the users performed the marketing on behalf of the email provider every time a user sent an email.

Using that technique, Hotmail acquired millions of users in a short time. More to the point, viral marketing made acquiring customers unexpectedly and incredibly inexpensive. Others watched, recognized why it worked, imitated it, and implemented it in new settings. Due to these spillovers from one experience to another, customer acquisition costs came down for many new commercial services on the World Wide Web.

Here is the point: The spillover made the World Wide Web more valuable and helped the technology diffuse.

Sometimes, knowledge spillovers involve the transfer of broad lessons, not precise details, which, once again, are learned through intelligent observation. The creation of Artist Shops illustrates this. Artist Shops is a service in which the graphic artist authorizes another firm—Threadless in this case—to produce items with prints of artists’ designs—shirts, socks, cups, and whatever.

Artist Shops emerged after Threadless CEO Jake Nickel began experimenting with digital printing in 2012 to bring the new technology to his crowd-sourcing site, which made prints on shirts using screen printing. Though digital printing was of low quality, he (correctly) anticipated that digital printing would soon acquire sufficient quality to substitute for screen printing. His firm mastered digital printing in 2015.

Artist Shops emerged after Nickell realized that his firm could further help the graphic artists with whom it regularly worked in its primary business. They could arrange to fulfill all the products that utilized graphic artists’ designs. This realization was inspired by watching what Shopify, Etsy, and other platforms did and noticing that none of them had tailored their services for graphic artists.

Threadless created early versions of Artists Shops within a year, and graphic artists reacted positively. Today, Artist Shops is one of the most popular services for graphic artists who sell online. Here is the point: As with any knowledge spillover, learning from the prior demonstrations cost Threadless little, and the new inventions of Artist Shops helped digital printing diffuse.

Spillovers of knowledge also happen in research when scientists learn from one another. Scientific inventions build on those spillovers. One well-known example changed artificial intelligence after the millennium. At that point, researchers were experimenting with tagging images but had yet to figure out how to scale the collections beyond a few thousand images.

A young Assistant Professor, Fei-Fei Li, learned from many other efforts and aspired to tag an order of magnitude more. One day, a graduate student told her about a new service called Amazon Mechanical Turk. How about hiring “Turkers” to tag images?

It took a few years, but Turk enabled a viable approach for creating a supply chain to tag millions of images. By 2009, there was enough to start holding a contest, ImageNet. As is well known, the 2012 winner, AlexNet, changed Artificial Intelligence forever.

In all these cases, spillovers alone do not make for advances. At the same time, spillovers help many entrepreneurs with the most challenging step: figuring out how to use a new invention.

Eliminating bottlenecks

Successfully embedding an invention typically involves eliminating a crucial bottleneck. Hugging Face illustrates such an experience.

After years of trying to develop a market for its chatbots, in 2019, an employee created a library of code that implemented Google’s 2017 idea about Transformers. However, in this case, the employee did it for his own model.

As it did with all its efforts, Hugging Face made the library available as open source for anyone to use. Management observed the reaction of other coders, who started building models with the library. Implementing Transformers had been a crucial bottleneck for others. Realizing that others valued the implementation by a Hugging Face employee, management pivoted to building a business around supporting all the activities unlocked by eliminating the bottleneck around implementing transformers.

Today, Hugging Face hosts the models created by users of their library. Virtually every new open-source artificial intelligence (AI) demonstration today calls functions in its library. In other words, their invention of a library of functions and code helped spread transformer technology.

The invention of GitHub CoPilot in 2021 also eliminated a bottleneck. GitHub Copilot is an AI-powered code assistant that raises productivity by helping developers write code by providing real-time suggestions. It leverages machine learning models trained on vast amounts of code.

To appreciate that invention, recall that Microsoft had acquired GitHub a few years earlier and appointed Nat Friedman as CEO and Thomas Dohmke as Chief Product Officer in 2018. Both had landed at Microsoft as acquisition hires in prior years, and they and their team aimed initially to develop the assistant within GitHub’s Integrated Development Environment (IDE).

Due to Microsoft’s earlier one-billion-dollar investment, Friedman, Dohmke, and their team had early access to OpenAI’s models, beginning in the spring of 2021. That motivated some trial and error with different applications. Frustratingly, they had little but their failures to show for it for a few months.

They learned that a crucial bottleneck was in the interface with users. They unlocked the crucial features after they tried the basic outline of architecture familiar to users today – short or long suggestions for code that the user could accept or reject. Quickly, they saw the success among their beta testers and began to release the software more widely, incorporating more tweaks and suggestions by observing more uses.

Today, we call this architecture CoPilot, recognizing its application outside of the GitHub IDE that first spawned it, trained on more than just code. Ever since the software tool has been adding features and improving. Knowledge spillovers have also been significant, as evidenced by many other entrepreneurs who have imitated the basic architecture.

Inventing a Way to Invent

An initial invention can become embedded in a product because a way to perform faster experimentation and iteration has been invented. That matters because many new products undergo iterative changes before launch.

That observation should be evident to anybody who has estimated AI algorithms. Fast iteration is an essential aspect of the new generation of AI algorithms. Hence, the experience at Hugging Face could also illustrate the importance of lowering the costs to iterate, which is what their hub does. Aside from the library of codes, the website further added an area to stage demonstrations of the results. Successful demos attract attention and feedback, which creates experiences that inspire further iteration.  

In other words, fast iteration helped AI applications grow and spread. The inventions at Hugging Face helped AI creators invent faster.

The largest online travel agency, Booking.com, provides another illustration of how to think about inventions that facilitate iterative learning. Management invested heavily in an internal infrastructure to conduct low-cost A/B tests on an enormous scale across its platform. Employees are trained to use it and have the discretion to experiment within the infrastructure as much as possible.

Only a minority of experiments succeed. Yet, despite the failure rate, the vast volume of experiments leads to many incremental improvements, which accumulate over time. The firm has gradually and steadily improved over time. Today, it is the largest online travel platform.  

In other words, Booking.com’s infrastructure helped its employees experiment faster. The invention of an infrastructure to support iteration enabled another technology to improve a comparatively mature invention: electronic commerce for travel.

Sometimes, the faster learning process takes on a systemic form. A few years ago, I interviewed the team that developed SageMaker for AWS. At the time, the suite of AI applications was aimed at users who did not have the time to learn technical coding languages but wanted something practical.

There was little distance between the users and the SageMaker team, who interviewed the users about their needs and current uses of the AI applications. The suggestions were then implemented, inspiring further application inventions. It is not an exaggeration to say Sage Maker adopted the new technology and set up a system to learn from experience quickly.

In all the above examples, providers anticipated that businesses would invent during the spread of the initial invention. Among other actions, suppliers converse with their users and learn by imitating the inventions of others or providing tools to make inventions easier for their users. No matter what form it takes, all of it helps the technology diffuse.

Conclusion

Unpriced inputs are tempting for innovative purposes because they are inexpensive and focal. Yet, none of this activity—knowledge spillovers, eliminating bottlenecks, or inventing new ways to invent—fits easily into traditional formal R&D processes, nor is it easily measured. The operational efforts may not necessarily appear as a separate line item in an income statement or balance sheet.

That makes it incredibly challenging to manage. How does a firm know how much to invest when there is no ready benchmark for monetary costs and gains? The lack of systematic measurement also reflects the ad hoc nature of many efforts. There are benefits but also dangers. A firm could put itself out of business with its own ingenuity.

Recognizing this pattern, this topic deserves more attention as a driver of productivity growth and economic change.

Copyright held by IEEE

February, 2025

The advent of digitization has significantly revolutionized the prototyping process over several decades in both commercial and scientific spheres. This aspect of the digital transformation lowered the costs of finding errors, sped up the iteration and addition of new features into prototypes, improved accuracy and precision, and enhanced the integration of prototyping with other activities, such as manufacturing.

Despite the progress, confusion persists about the characteristics distinguishing commercial prototypes from their scientific counterparts. That confusion grows when a firm resides in a setting where both types of prototyping occur, such as generative AI or disease detection from medical images. Let’s dig deeper to understand why changing economic conditions blurred some lines.  

Commercial

Commercial prototypes play a pivotal role in the innovation process, facilitating the evaluation of a product’s design, functionality, and usability. They also act as a focal point for expanding the application of cutting-edge technologies in new directions, identifying and addressing issues that need resolution before product launch. 

In the later stages of product development, especially in modern software, prototypes aim to understand user experience and usability. Designers often make prototypes with beta testers in mind. They want to learn if they correctly anticipated how users will react to new features. If issues arise, the prototype can be adjusted and refined before a design is frozen in a broader release. 

Here is an example. One of my former students is involved in a commercial venture to develop and diffuse algorithms that help dentists detect disease in routine X-rays – from cavities to cancer. In addition to everything else they do, the firm makes algorithms for different diseases, trying them in prototypes first. Why? So they can anticipate issues before a dentist ever touches their product.

Commercial prototypes incorporating existing product improvements can serve an additional purpose in some situations, what managers call maintaining a ‘flywheel.’ This ‘flywheel’ is set in motion when a firm develops software that attracts users and learns from users about desirable new features. In an ideal scenario, the prototype tests these new features before a new version’s release. If successful, these features should satisfy existing users and attract new ones, thus generating additional revenue and perpetuating the cycle.

A crucial determinant of the flywheel’s speed is the cost of generating and testing prototypes of new features. Firms would create a gazillion new features as often as possible if every prototype cost nothing. Of course, designers, coders, and managers take time and effort, even with the most agile methods, and the costs add up. So, managers consider options before prioritizing certain features in each prototype.

Maintaining a ‘flywheel’ with the aid of prototypes can involve high-stakes decisions. Prototyping costs and timing are balanced against the reputational costs from error if the prototypes do not deliver on promises and the strategic urgency of releasing a new version. Some mistakes can sink a launch and must be corrected in advance, but so can delays.

In my experience as an observer, assessing the benefits of features is often more complex than forecasting the costs. Firms differ in their soliciting user advice and determining which options to implement. Moreover, environments differ and matter. The specifics of products, the firm’s mission, the competitive setting, and other factors all play a role in determining priorities for prototyping.

Science

As with commercial prototypes, scientific ones represent a functional product or service, and may not embed all the relevant features, nor account for all the interactions among those features. Yet, different concerns shape scientific prototypes. They may confirm or refute a theory or enhance an idea. They may inspire other scientists to replicate the experiment or apply a similar process in a new context, especially if they demonstrate a novel method for testing an idea or concept. 

To illustrate, let’s contrast a setting with high and low prototyping costs. Consider prototyping in quantum computing, where the costs are high. Cryptography and possibly other high-precision measurements are so valuable to the US military that, despite large-scale functional and operational products being far off in the future, the US military regards the potential value from experiments as high. Most of those experiments build prototypes, not fully functional products.

Despite the high cost, the military willingly funds research to build prototypes. These test scientific ideas in that area. Relatedly, most of the work occurs on simulations in university and government labs before anything is ever built. Most interestingly, a few specific applications are now within reach. In these areas, some commercial firms are also beginning to prototype.

Contrast that with a setting infused with lower prototyping costs, namely, algorithms for detecting cancer in medical images. Like algorithms for dentistry, the main costs arise from obtaining and cleaning the appropriate images and then tagging them. That takes time. A person (or team) must also train the algorithm.

The prototype costs for disease detection are non-zero but, on average, much lower than in quantum computing. Unsurprisingly, there has been a massive increase in the creation of prototypes for scientific literature.

Like their commercial counterparts, scientific prototypes embed a contrast between the corporal and cerebral. The costs are immediate, while the gains are intellectual. Like the other type, valuing the benefits can be difficult before building a functional product that can operate at large scale, and the representation, while necessarily imperfect, must be designed to shed light on the goals. However, the goals and audience differ, and that makes a world of difference.

Scientific prototypes typically aim to communicate with other scientists with transparent documentation and meticulous detail. This level of detail raises the costs of scientific prototypes, but it is essential. It reflects the scientific community’s commitment to clarity and reliability, which are crucial for reproducibility and scientific knowledge accumulation.

Mixing prototype purposes

The most confusing situations are settings where both scientific and commercial prototypes exist. This happens in commercial areas where science is crucial to advancing products.  Designers of one type of prototype may be keenly aware of the other and influence each other. The same team in an organization may even do a single prototype with two purposes: commercial and scientific.

The interplay is most interesting to watch. At their most basic, new scientific prototypes can encourage new commercial ones, and vice versa.  The impact of ImageNet and ChatGPT3.5 illustrates how that can happen.

ImageNet is an example where a scientific prototype led to new commercial ones. After the turn of the millennium, Fei-Fei Li hypothesized that the range and size of the training data sets had held back scientific progress in image recognition algorithms. By 2009, she and her team had figured out how to compile tagged training data that achieved a scale of digital images far beyond what anybody had considered. She released ImageNet in 2010.

In 2012, ImageNet enabled convolutional neural networks to demonstrate their superiority convincingly. That had enormous consequences for commercial firms and entrepreneurs, who gained a new understanding of what was feasible and possible.

Here is what made that effort and result so remarkable. No researcher at a firm had thought of doing this, not even researchers in commercial firms with well-funded R&D labs and major search businesses that would directly benefit, such as Google or Microsoft. Yet, though they did not pay a cent for it, those firms benefited from the new understanding. Their commercial prototypes and products changed irreversibly.  

Compare that with ChatGPT3.5, introduced in November 2022. In this situation, commercial prototypes have influenced scientific direction.

Before that introduction, OpenAI and other organizations had been making prototypes of large language models for years. In addition to more tokens and a more extensive training set, the interface for 3.5 was new and tailored for use by a non-expert, showcasing improved conversational skills. OpenAI also implemented numerous guardrails for what computer scientists call safety and compliance issues, such as biased and offensive language.

Moreover, when it moved from prototype to product, OpenAI chose what problems to leave unresolved. For example, ChatGPT does not have a low-cost way to measure the incremental contribution of a specific set of training data. That would have been useful because it would help licensing negotiations and contribute to a standardized benchmark that could help anticipate (and perhaps settle) legal copyright claims.

In 2022, the computer science community had not built a prototype with a solution, and neither had OpenAI. After ChatGPT’s success in gaining users, everyone understood its importance. Commercial success has magnified the area’s importance. Scientific prototyping has responded with an explosion of new research on unresolved problems.

In other words, ImageNet’s example illustrates how success in scientific prototyping enabled a new direction for commercial prototyping. OpenAI’s example demonstrates that success in commercial prototyping can focus interest in scientific prototyping on unresolved problems.

These examples also illustrate how important communication is between scientific and commercial prototypes. Yet, there is an essential tension in communications. Scientific activities require transparency, while commercial activities usually remain proprietary and secret, especially if rivals are both prototyping, as is occurring in generative AI. This is why many observers forecast that recent commercial success may interfere with scientific progress in the near term. 

Conclusion

Science and commercial prototypes are not assessed using the same criteria and do not try to achieve the same goals. Each needs feedback from different communities, further reinforcing the differences.

When they overlap, commercial interests largely determine the visible direction of prototyping at firms with commercial stakes in a flywheel. However, even though they differ, intelligent commercial firms also engage in scientific prototyping to understand what is feasible and how the underlying technology likely will evolve.

Copyright held by IEEE

October 2024

This year, we have had many outstanding awards for Least Supporting Actor, Best Drama without Histrionics, and Best Short. We also have our annual Best Meme and Best Use of Digital Technology in Science awards and the James Bond and Rosebud awards.

Be assured that Elon Musk will win something. How could his antics not deserve at least some ridicule? Google and OpenAI will also receive awards since they have often been in the news this year. But newsworthy is not enough. No prize will go to Taylor Swift, who has contributed so much to the television ratings for the Kansas City Chiefs, which employs her boyfriend, Travis Kelce. Did you know that last year’s Superbowl, which the Chiefs won, was the largest streamed event ever, tying up almost a third of the internet’s traffic on that day? But I digress. Let’s start the fun.

The first award is for Best Original Play. Revenge is a dish best served cold, and it is incredibly wonderful if satire can be added to the mix. This award goes to The Onion, who had the brilliant idea to bid for the website InfoWars, from which Alex Jones once preached hateful conspiracy theories. Jones lost his suit with the Sandy Hook families, and nobody with any sense of human decency shed a tear for him. He was forced into bankruptcy and forced to put up the site for sale. As of this writing, a judge has held up the transfer, so this may or may not happen. Even so, the headline was worth it, as it brightened up the morning it first appeared. Bless you, Onion. Every slice of material brings tears to my eyes.

Let’s get the awards to OpenAI out of the way. They are a double winner this year. The next award is for the Best Short Film. The award goes to OpenAI for the demo for Sora. Have you seen this? It is possible to make a film with just prompts. Look at those puppies. Look at the wooly mammoth. Look at the ocean. If you have not looked at this video, you should. Wow. Imagine the things one could fake with such a fictional film creation. Whoa. I expect them to return next year with a contender for best original screenplay.

Best use of Digital Technology in Science. Everybody knows that ChatGPT 4.o has a tough time doing complex math problems, but nobody expected such a leap forward quite so fast. OpenAI’s model o.1 solved 83.3% of math problems in the Olympiad by combining the LLM with another geometry-solving algorithm. For that accomplishment, they get the award. That is a massive improvement over ChatGPT4.o, which got 13% correct. Rumor has it that the o1 model takes vast amounts of computing to operate, and OpenAI will charge thousands of dollars for each prompt. That is an expensive toy, especially in light of the leaked details behind their $6.5B funding round — over $3B last year in revenue and over $11B expected next year, yet they are losing billions each year. While this award was a no-brainer, it might better fit science fiction instead of science. After all, OpenAI could also get an award for the best modern imitation of Frankenstein’s monster. Isn’t this one more step towards AGI? Oh, but I digress.

The next award is for Least Supporting Actor. This occasional award goes to the infrastructure firm that failed most at its basic job. There are multiple winners. All starred in horror flicks. The first winner is the Roku Horror Picture Show. The plot? Roku has a data breach, and everyone’s private data gets released to the dark web. The scariest part? Roku takes no responsibility for anything, and nobody can do anything about it. What a time warp! A second winner was The Crowd Strikes Back. In this intriguing corporate thriller, an unknown upstream provider of one component of the system, CrowdStrike, pushed a bug into widespread use. More than 3,400 plane flights into and within the US were canceled in a day. The movie ended as many tech whizzes spent their day fixing systems with meticulous repair procedures. The last winner stars AT&T Wireless in the production titled Burning Down the House. A technician loaded a bug into the network, blocking 92 million calls and 25 thousand attempts to call 911. All three movies contain the same dystopian ending: As the sun sets, the camera pans over grade school water fountains drying up, unlit traffic lights swaying in the wind, and automated doors in an office building refusing to open. The camera continues and enters the building, goes down a staircase and into a hallway, and settles in a room in the basement where a hapless technician frantically searches for the password to restart the electrical grid. He stops, mumbling to himself, and the audience learns that he kept the only copy of the password on a file on the cloud. As the camera fades, he realizes he cannot access the cloud because, after all, the electrical grid is down. Ouch.

Best Drama without histrionics. Despite plodding along for the last few years as the first major antitrust case in over two decades, the Federal Antitrust case against Google is the winner of this year’s award for drama. To be sure, nobody noticed the trial. Google’s lawyers kept testimony out of the public eye under the guise that the information was competitively sensitive. So, there were few headlines until Judge Mehta returned with a guilty verdict and provided a two-hundred-page explanation for what the fuss was about. The two most damning pieces of evidence? First, Google manipulated the ad auction to raise prices, and then employees covered up their tracks by altering the releases of information about the auction, making it less transparent. Second, Google paid — and continues to pay — Apple $18 Billion a year to be its exclusive search engine, which also turns out to be enough to disincentivize Apple from developing its own. (For $18B, I could find excuses to slow-walk the development of a search engine. Wouldn’t you?) Here is the thing about this document: About a quarter of it analyzes why Google has market power in some markets and not others. Which motivates the observation: It is all good to eliminate rash decision-making from antitrust reasoning. However, did society need more than four years of legal back-and-forth to determine something as seemingly obvious as “Google has market power”? Four years? *sigh* While this movie had little box office appeal, there is good reason to think a sequel is in the works. Its tentative title is “Throw the Book at Them.”

Now it is time for the James Bond Gadget Award, and this year, it involves both information technology and blowing up stuff. The winner is the Israeli secret service’s placement of bombs inside pagers and Walkie-Talkies, which were sold to Hezbollah. They exploded on September 17. If you missed the backstory, the Secret Service imitated a Hungary assembly operation for pagers and black market provider of Walkie-Talkies. You might reasonably wonder why Hezbollah was buying such primitive communications equipment at all. That is because the Israeli spy agency – like every other competent spy agency – had learned how to tap cell phone networks successfully. Hezbollah believed older communications technologies – i.e., pagers and walkie-talkies – would be more challenging to tap. Here is the thing about the bombs in these pagers: unlike a counter-missile strike, these exploding devices targeted soldiers of Hezbollah (and, from news reports, their handlers in the Iranian Embassy too). What was the most ironic part of the whole episode? Hezbollah’s indignation at being bested, as their spokesman declared the explosive pagers were an “act of war.” To which observers reacted with, “Huh?” Since October 8, 2023, for eleven straight months, Hezbollah had been tossing scores of Iranian-supplied rockets over the border each day, displacing almost one percent of the Israeli population from their homes. Weren’t they already at war? Needless to say, these little explosions were the beginning of an Israeli counter-offensive, and matters have not gone too well for Hezbollah since then.

Speaking of international intrigue, the next award is The Honorary Robert Ludlum Biggest Mystery Award. This award goes to the events that most resemble a movie made from a Robert Ludlum book. The events must contain international intrigue and unanswered questions. The award goes to Pavel Durov, who owns and manages Telegram. In case you missed the news, he recently was arrested and detained in France for failing requests to answer information from the French government about human trafficking on Telegram. A Russian by birth but a French and UAE citizen by choice, his service, Telegram, has grown to (reportedly) over 900m users worldwide. It is used by many who seek to evade government surveillance, which means it is the communication service of many well-meaning citizens of authoritarian governments, and also many creepy people from the dark web –drug runners, porn addicts, human traffickers, and nationalist Russian bloggers. As such, the service has become a flash point for some strange bedfellows defending free speech. Durov is a quirky dude, obsessed with his physique and sarcastic to a fault (in a way that only someone who has defied the Russian government seems capable of). Alas, his libertarian inclinations lead him to neglect any content moderation, even in the presence of violations of basic human decency. One mystery is why he refused to answer calls to stop human trafficking, which Telegram facilitates, and he could quickly stop it if he wanted to. This is part of what the French government wants him to do. An even bigger mystery is how somebody so smart could be so stupid and take a plane to one of the few countries where a government could legally serve him an arrest warrant.

Oh, we need to lighten up the mood. Let’s give the award for the Best meme. This year, the award goes to the memes about Princess Kate, who released a doctored family photo after announcing she has successfully gone through cancer treatment. Kate is just about the best thing the British royal family has going for it, and she has been dealing with chemo and deserves the world’s love. She got that and more once the doctoring was discovered. Kate said, “Like many amateur photographers, I occasionally experiment with editing.” That just asked for an affectionate response from the internet, and a little affection never hurt anyone. Welcome back to public life, Kate. We are glad to see you smile, and here’s hoping a few more doctored photos make you smile further.

Mr Smith goes to Washington award. This award goes to an amateur politician who sincerely tries to do some good for the world. This year’s winner is Vic Miller of Cheyenne, Wyoming, who ran for mayor on the premise that ChatGPT could run the city. No, seriously. He did. Before you dismiss this out of hand, if you read his platform, it is not entirely crazy. His basic idea is pretty plausible if you know anything about modern chatbots. In so many words, he said that most city government employees answer citizens’ requests by looking up local ordinances, filing documents, and keeping track of small matters like permits to build a parking lot. He proposed letting a chatbot operated by ChatGPT do as much of that as possible so the mayor could do the ceremonial stuff, such as ribbon cutting and attending memorial services. Before we go down a rabbit hole discussing how to train a chatbot on city documents, let’s first find out how the election turned out for Vic. The citizens of Cheyenne voted for candidates with more conventional visions; Vic got 3% of the vote. Welp. Perhaps his vision for local government is ahead of his time. Still, what do you want to bet he inspired someone? There are some entrepreneurs somewhere who have read about Vic’s idea, and I bet next year we will hear about a startup selling such a service to city governments.

Speaking of politics, it is time to give the annual Rose Bud Award. Named for the sled featured in the movie Citizen Kane, this award goes to the billionaire that society would most like to put on a sled and send down a snowy hill, preferably to stay there. Musk and Zuckerberg have been trading this award in the last few years, and Elon Musk earned it this year with three actions. First, Elon got in a spat with the Brazilian government over enabling misinformation on Twitter/X. The government threatened to arrest his Brazilian managers, so he moved them out of the country. Then, the government threatened to shut down X, and eventually, it did. After three weeks of no service, he backed off and removed the offending material. Second, he donated $250 million to the Trump campaign, which made him its biggest bankroller. Look, I am not criticizing the business decision. With an ethically challenged transactionally-oriented prospective president, Musk has bought his way to regulatory relief from investigations by the Security and Exchange Commission and the Department of Transportation. Most investors also assume Musk has also purchased additional contracts to send up rockets for NASA. All in all, his stock value has increased by far more than $250m. On the other hand, it is a flagrant violation of conflict-of-interest rules that every mortal human lives by (and a further sign that we have entered the late stages of capitalism). More to the point, it will be fodder for Congressional investigations after the house flips in the midterms. Third, he has joined with Vivek Ramaswamy, another smarmy rich guy, in an unofficial performative group called DOGE, which ostensibly stands for Department of Government Efficiency but is a cringy shoutout to DogeCoin, a digital currency that Musk likes. Here is the rub: Nobody expects DOGE to accomplish much except make headlines and keep Musk from managing Tesla, Twitter, and SpaceX. However, the competent day-to-day managers at those firms are thrilled to see Elon distracted because when he shows up, Musk has no idea what is happening, and he just makes mistakes anyway. DOGE keeps him away. This raises the question of when we are all better off — when Musk mismanages the federal government or his own firms? Oh, but I digress again.

This next award takes no inspiration from movies but brings attention to a significant crisis. It is a new category of award called the Burning of the Modern Library of Alexandria. This refers to the denial-of-service attacks on the Internet Archive. The Archive operates the Way Back Machine, one of the Internet’s most valuable public services. Why, why, why? What evil effort motivates this attack? This is distinct from another problem the Archive has faced this year: selfish overuse. For example, on May 28, the Archive went down, providing the following details. “Apparently, an AI company is harvesting Internet Archive texts at an extreme rate.” Editorial to whoever this is: please refrain from illustrating theories about the overuse of the commons. Let the site alone so the staff can concentrate on fighting the truly evil folks out there. Everyone thanks you.

Let’s end this award ceremony in a lighter tone. This last award is always for some light but head shaking event that defies explanation. So it is called You Can’t Make This Stuff Up Award. The winner is Braintrust Air, who put out a job ad trying to hire someone to rewrite their AI-generated web pages so that they don’t sound like crappy AI. The punch line? Braintrust Air is an AI company. See the screenshots below from LinkedIn. Think about it.

************************************************************************

Those are the awards for this year! Bless you all for reading this far. And try not to forget the overarching lesson of these awards. Just because we live in an era of magnificent digital achievement, that is no excuse for neglecting a bit of human decency in your daily life.

Have a wonderful new year!

Fei-Fei Li is known for leading the development of ImageNet, which helped catalyze machine-learning approaches to vision recognition, and for being an essential voice shaping the science behind Artificial Intelligence (AI) today. She offers an extended, thoughtful, and heartfelt memoir in this book.  Beautifully written and grounded in many rich, thought-provoking observations, the book describes her journey from being a child immigrant from China to her present position as a Stanford Professor.  

Li’s book draws on interconnected vignettes. These vignettes, particularly the deeply personal ones, shed light on family crises and career choices, pointedly when neither could wait. They also offer profound insights into AI science’s antecedents and progress, providing a unique perspective on recent history.

Crucially, she does not spoon-feed the reader. The book anticipates an educated reader who understands the broad context. For example, Li never provides a layman’s overview or timeline to summarize the broad history of AI – its origins, fads, and winters. That should work well for most technically savvy readers.   

While the book is easy to devour, I found it challenging to digest. That is a compliment. Li’s point of view is clear, consistent, and thoughtful. Li offers many insightful, broad observations, but not all help the reader interpret every vignette. An engaged reader will think about those observations after putting the book down. 

Unlikely Origins

After the introduction, the book follows a chronological narrative. These early chapters contain little about AI, describing instead the Li family’s journey from China to the US and the massive adjustments required. The family settled in New Jersey, living an economically precarious life.

Li possessed an inherent interest in science from a young age, which her parents encouraged. While she had encountered resistance to girls pursuing physics in China, she experienced the social roughness of being an outsider in a US public high school.

Such a crucible at a young age would require much mettle from anyone. Amazingly, Li escapes the situation by following her science north star, which is inspiring. Alternatively, give the first half of this book to a whiny student. Then say, “If you think you had it hard, read this.”

Though Li’s high school was not a likely place to spawn a future academic star, her high school math teacher was the difference-maker. He exemplifies how a thoughtful and empathetic individual can change a student’s life. Though he starts as a science mentor, he eventually becomes more like a benevolent uncle.

Li’s parents’ financial situation overshadowed her ability to pursue her dreams. Her high school teacher gives Li her first financial break; he lends the family money to buy a cleaning business. Then came more successes: a full scholarship to Princeton, funding to be a summer research assistant in an experiment digitizing vision in Berkeley, and a scholarship for a Ph.D. at Cal Tech.

The funding mattered. Li recalls how close she came to taking the money of a standard American job in consulting or finance. She admits that she would not have stayed on the scientific path without her mother’s pointed questions and resolute willingness to sacrifice comfort for her daughter’s dreams. This funding was just enough to keep her on this path.

Funding also opens opportunities to the best of the American academy, starting with Li’s experience as an undergraduate. In graduate school, two professors mentored her on the science of sight, computational linguistics, and an eclectic mix of academic areas. In other words, many saw something in her, made a bet on her future, and then she produced path-breaking research that justified all that investment.

What is the lesson? Computer scientists do not come to breakthroughs quickly. Like other scientists, computer scientists follow a fragile path to the frontier. To be sure, the individual must be extraordinarily intelligent and resourceful. But science also requires hard work, perseverance, and curiosity married to an openness to insight from unexpected sources. A pro-social outlook from many mentors at many stages helps immensely. Li’s account raises the question of how many future Einsteins never made it down the path due to a lack of funds or support.

The atmosphere of science

It is often said that nobody should go into academic research if they crave immediate gratification. Li’s journey shows the extreme side of that platitude. The audacious ideas for ImageNet drew on multiple influences from her training as a graduate student. Yet, nothing paid off until years into her career as a professor.

In graduate school, she began to hypothesize the need to assemble and tag many pictures of various objects. Then, over time, she began to appreciate the order of magnitude required. Yet, the costs initially overwhelmed the aspirations.

Many vignettes illustrate what is often unheralded today: creating low-cost supply chains for tagging carefully curated digital photos was an enabler for next-generation visual AI. Moreover, it did not arise overnight. In Li’s case, after years of stops and starts, a creative and hard-working graduate advisee suggested using Amazon’s Mechanical Turk (MT) for tagging. MT got ImageNet to the next level.

When a convolutional neural network won the 2012 ImageNet contest, blowing away the next-best entrant by ten percentage points, Li saw it (correctly) as vindication of all her efforts. Most interesting, however, was that she was as surprised as anyone at the solution. In describing why, she studiously avoids retrospective bias. The account explains why this breakthrough was obvious with some hindsight but not until it happened. It is an uncommonly beautiful piece of science writing.

Li offers many vignettes in this part of the book in which she walks a highwire between personal and professional crises. She has some of the most significant moments of her career while helping her parents with the cleaning business or tending to her mother’s chronic near-death illnesses. Eventually, she and her academic spouse managed a “commute marriage” between California and Michigan for many years while pursuing independent career goals, ending up in the same location just before their first child arrived. 

She must have unbounded energy and an uncommon ability to juggle different pressures. She was also fortunate to have nurturing parents. In addition, she found a talented, patient spouse who would delight any set of parents. (He is Italian, has a Ph.D., and cooks!) Every reader will root for her.

The future of science

The last parts of the book touch on contemporary themes. Li covers a provocative list of concerns. For example, she worries about what private firms in AI neglect, such as ethical considerations. She identifies why complex neural networks make transparency about ethical issues so challenging to address. She also gives a window into the next frontier, ambient intelligence. She shows how it could benefit healthcare operations but violate privacy in the wrong hands. It is a credit to her thoughtfulness that she does not offer bromides and avoids the self-important posturing and simplistic declarations common in social media.

Still, no book can do everything, and I would also have liked Li to be more self-aware about a few incomplete observations. The reviewer’s job is to notice blind spots, so let me list a couple.

After an extended sabbatical at Google, Li concludes that today’s frontier AI models rely on more resources and talent than most university labs can afford. She explains why in general terms. A few more specifics would have helped me understand where universities specialize, where firms should focus, and where their interests overlap. Li must think about those questions daily, so hearing her views would have been enlightening.  

The tensions in the university/industry relationship also needed more scrutiny, especially how it shapes AI today. Li does talk a bit about the administrative challenges of conducting science at universities, but not much about the many dimensions of short-termism at private firms, especially when researchers in either setting pick projects or follow fads.

She also alludes to the symbiotic relationship between private and university interests. Still, she does not analyze how private interests can keep technical secrets or hollow out university personnel. How can technical training occur in universities if private firms’ investment shapes the technical frontier?   

Finally, Li talks about her status as an Asian woman in a male-dominated field. She is certainly justified in feeling lonely as a woman in computer science. While women have slowly worked their way into many academic fields in the last few decades, computer science has stubbornly stayed at a low percentage.

In response, she cofounded AI4ALL, a society to help women get into computer science and find mentors. That is a fascinating initiative for a challenging situation. I would have liked to know more. What has this organization done during its history?

I also wanted a half-empty-half-full perspective. The number and percentage of non-white faculty – i.e., particularly South and East Asians – has grown in the last few decades in computer science and related fields. Her perspective on these trends would be interesting to hear, but she says almost nothing.

Conclusion

The book ends with Li returning from her sabbatical at Google to teach a 600-student class on neural nets. It is a fitting end. Teaching energizes Li, and she hopes to see her younger and more curious self among her students.

I bet most undergraduates would be inspired to take a class from her. More pointedly, the book implies that no Ph.D. student could survive her supervision without high aspirations. Having someone like that as an anchor for the AI scientific community is fantastic.

Copyright held by IEEE Micro

First published V44, 3, May/June 2024.