The radiologist, the dynamo, and the dangerous allure of the “retrofit”
David Peterson

Two weeks ago I wrote about the bottleneck cascade and how every breakthrough shifts scarcity somewhere else. This piece zooms in on one of the most seductive dead ends that cascades produce: the retrofit. 

In 2016, AI pioneer Geoffrey Hinton famously declared that we should "stop training radiologists now," as it was obvious that deep learning would soon outperform them. The prediction made perfect sense. AI excels at pattern recognition, and a radiologist’s job, it seemed, was to spot patterns in images.

And yet today, not only are radiologists still employed, but the demand for them is growing.

Why was Hinton so wrong? For a complete telling of the story, read this excellent piece from Works in Progress. But the short version is that the attempt to replace radiologists was a classic retrofit: inserting a new technology into a single step of a complex workflow. And the effort quickly slammed into a cascade of deeper bottlenecks that technology alone couldn’t solve, from messy real-world data to the thorny issues of legal liability and the simple fact that a radiologist’s job is more about judgment and consultation than just spotting shadows.

This is the bottleneck cascade in action. The attempt to solve a single problem (that a technology is well-suited to solve) reveals deeper, systemic ones (that said technology may or may not be relevant for).

And this dynamic isn’t unique to healthcare. It is the central challenge of the AI transition for every industry. How many of today's celebrated startups raising gobs of capital at sky-high valuations are selling sophisticated retrofits? How many of these products accelerate one piece of an old process, without attempting to meaningfully change the process itself? 

As investors and entrepreneurs, this is a particularly important line of inquiry because the long-term prospects of the retrofit strategy are decidedly uncertain. Retrofits are, by their very nature, destined to be ripped out and replaced once businesses begin the far more difficult, and valuable, work of completely re-architecting their operations. Some retrofits may catalyze that shift (which is why FDEs are the startup go-to-market strategy du jour, and VC/PE roll-ups are so hot), but most will not. Understanding the difference between these two approaches is the single most important strategic challenge for founders and investors today. To navigate it, we need a map. Thankfully, history provides one.

The dynamo’s lesson

In 1990, economist Paul A. David wrote a landmark paper to address a puzzle that vexed his peers: with powerful computers appearing everywhere, why weren't they showing up in the productivity statistics? This "productivity paradox" felt like a deep contradiction.

To answer this question, David looked back in time to the diffusion of electricity and the rise of the electric dynamo (the precursor to the electric motor) at the turn of the 20th century. In 1900, factories were installing dynamos, cities were lit with electric lamps, and streetcars were powered by electricity, yet the massive economic boom everyone expected had not materialized. That boom, he pointed out, only arrived in the 1920s, a full forty years after the dynamo's introduction.

The delay, David explained, was caused by a critical lack of imagination about the unique capabilities of the dynamo. Factory owners initially treated the new technology as a substitute for the old one. They would replace a single, massive steam engine with a single, massive electric motor, but keep the old, inefficient factory design, which relied on power coming from a central source. 

This is what happens in the early days of any new revolutionary technology. It’s not that the factory owners weren’t trying to employ electricity to make themselves more productive. They just thought they were re-architecting, when they were actually retrofitting.

The revolution began when engineers realized the dynamo's true potential wasn't just generating power, but distributing it. This insight led to the “unit drive” (that is, a small motor dedicated to each machine). Freed from a single, central power source, factories could be completely re-architected. Production was now organized in logical assembly lines, making factories lighter, safer, and radically more efficient. And the payoff was enormous. Between 1915 and 1929, U.S. manufacturing productivity jumped from ~1.4% to ~5% annual growth.

This reveals an undeniable truth. Truly revolutionary technology doesn't just speed up the old way of doing things, it demands we invent an entirely new one. The productivity gains don't arrive until we stop retrofitting and start re-architecting.

What is AI’s “unit drive”?

We find ourselves in a similar moment right now. AI is the next great general-purpose technology. But to unlock the next wave of productivity, we must identify the "unit drive" of our time. What is the fundamental, granular building block that will allow us to fully re-architect our world?

I don’t know. I don’t think anybody knows. But here are three hypotheses. If intelligence is best understood as a unit of labor, then we’ll need autonomous agents that can reason, plan, and execute tasks. If it’s best understood as a standardized utility, like electricity itself, then we’ll need a shared intelligence layer that any AI “appliance” can plug into. And if it’s best understood as a trustworthy answer, then we’ll need verified components that can provide auditable and insurable responses ready to be built into critical systems.

The form this “unit drive” takes is still an open question. The real test will be whether it enables true re-architecture or whether it leaves us stuck in ever more sophisticated retrofits.

So what’s the opportunity for entrepreneurs and investors today? Retrofits will be ripped out. The models will eat most of the margin from the application layer (as Jerry Neumann persuasively argued in his fantastic essay “AI Will Not Make You Rich”). So the only lasting winners (other than the labs themselves) will be the re-architectors. The firms that reorganize themselves to capture the downstream productivity boom. That leaves a narrow but critical band of opportunity today: the tooling that makes re-architecture possible. In other words, the unit drive. Or at least the standardized wiring and plugs that made the unit drive possible. Unglamorous, yes, but necessary.

Neumann is right that most of today’s AI profits will be competed away at the model and application layer. But that doesn’t mean there isn’t real opportunity right now. When microprocessors first appeared, Intel imagined calculators. It took outsiders to realize the personal computer was the real prize. The same is true now. Somewhere in the crowded field of seeming retrofits lies AI’s “unit drive. For entrepreneurs and investors, the work today is not in betting on the shiniest retrofit, but in backing the toolmakers, experimenters, and tinkerers who are laying the wiring for what comes next.

The AI-Native Enterprise Playbook
Ten real-time observations on a rapidly evolving playing field

No More Painting by Numbers
It’s the end of the “SaaS playbook.

The Age of Artisanal Software May Finally be Over
Every wave of technological innovation has been catalyzed by the cost of something expensive trending to zero. Now that’s happening to software.

Founders as Experiment Designers
David on why founders should run everything as an experiment.

ENTERPRISE/TECH NEWS

New week new models. The AI race continues at full speed with new models launched this week from both Anthropic and Deepseek.

$100M in ARR just ain’t what it used to be. The Information dug into the new crop of AI companies that broke into the $100M ARR club and questions how durable those revenues are. One thing is for sure: founders are under incredible pressure to pump up their “ARR” numbers as much as possible - a pressure that only seems to increase with scale. “For some of the startups, their revenue is well above $100 million. Last week, Brendan Foody, CEO of AI recruiting firm Mercor, said on X that the company hit $500 million in annualized revenue, up from $1 million, in just 17 months. That’s an impressive feat although it’s much less impressive when you realize that the company pays a large portion of that revenue to the contractors that handle AI tasks for its customers. Its net revenue is about 30% of the ARR figure, according to a person with knowledge of its financials, meaning that its actual ARR is likely closer to $150 million these days.

That’s just one more reason why you can’t always take ARR figures at face value, especially for businesses that sell services rather than products, as we’ve detailed in past columns here, here and here. The most obvious flaw with the ARR metric is that a company might hit the $100 million ARR mark based on taking the last month’s revenue and multiplying it by 12, but that doesn’t mean it’ll stay at that level. One AI company we’ve chatted with, for instance, currently generates more than $100 million in ARR. But due to seasonal changes in its customers’ spending, that number can fluctuate below or above the $100 million mark depending on the month, according to its CEO.”

Workslop. According to a new report by Harvard Business Review, “employees are using AI tools to create low-effort, passable looking work that ends up creating more work for their coworkers. On social media, which is increasingly clogged with low-quality AI-generated posts, this content is often referred to as “AI slop.” In the context of work, we refer to this phenomenon as “workslop.” We define workslop as AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task.”

HOW TO STARTUP

Consumer to enterprise. The fantastic Kyle Poyar examined what he calls the “consumer to enterprise” playbook, using Fyxer as an example. This is motion is being attempted by a large number of AI-native companies, spurred by strong prosumer demand for initial products and a recognition that enterprise products are, ultimately, more valuable in most cases. “Fyxer is aggressive about inviting your team. Not only is the team invitation in the onboarding flow, Fyxer offers a $50 credit for each teammate you invite to Fyxer. While I rarely see such a bold team invitation push, I have confidence that this has been rigorously tested by the team. Fyxer has four Growth Engineers (10% of the company) and each Growth Engineer runs 1-2 experiments per day around improvements to onboarding, referral programs, and the dashboard itself.”

Designing NotebookLM. Jason Spielman, who led design for NotebookLM (probably the most useful AI tool I use in my daily work - GD) published a detailed breakdown of design decisions that led to the product users are experiencing today. “The mental model of NotebookLM was built around the creation journey: starting with inputs, moving through conversation, and ending with outputs. Users bring in their sources (documents, notes, references), then interact with them through chat by asking questions, clarifying, and synthesizing before transforming those insights into structured outputs like notes, study guides, and Audio Overviews. By grounding the design in this linear but flexible flow (Inputs → Chat → Outputs) we gave users a clear sense of place within the product while keeping the complexity of new AI interactions digestible and intuitive.”

HOW TO VENTURE

Don’t believe the hype. Eren Bali, CEO of Carbon Health, wrote an outstanding post on how “The zero to $100m in a year is the new norm” narrative is getting out of hand.” The key point is that the drivers of these outliers are usually temporarily specific and unlikely to be durable. “These unnatural early stage growth curves (will call it ultra high growth) happen for a few different reasons. Most of the time there’s a temporary supply demand imbalance for a good that becomes highly in demand over night. There were COVID testing labs that went from 0 to $1b+ revenue in a year when the pandemic broke. These days it’s the AI infra / compute companies. Initially both demand and prices skyrocket simultaneously which creates ultra high growth. Eventually competition comes in, prices normalize, margins shrink, growth slows down and valuations tank. If you sold your company before that, congrats.”

Maybe biggest is better? As it turns out, A16Z is doing pretty well. According to data leaked to Eric Newcomer, “Andreessen Horowitz’s $900 million third fund from 2012 appears to be its best performing fund at 9.4x net total value to paid-in capital as of March 31, 2025. The firm reported the total value of the fund as $13 billion to limited partners. The value of that fund peaked at 10.5x eight years into the fund, fell to 6.8x when tech stocks tumbled, and then has climbed back.”

Lunar announces it has joined the Google Cloud Partner Advantage to advance secure AI gateway governance.

Valohai announces partnership with Oracle.

FalkorDB
Core Developer (Tel Aviv, hybrid)

Groundcover
Backend Engineer (Tel Aviv)

Motorica
Platform Engineer (Stockholm)