The Deep Feed

01 — Not Boring

The Ultrasonic Mirror: Midjourney’s Leap into the Biological

Why an AI image company is building a medical scanner

By Packy McCormick · 8 min read

Editor's note: A look at how generative AI companies are pivoting from pixels to physiology.

When Midjourney announced a hardware product, the assumption was predictable. We expected a digital canvas, perhaps a pair of glasses, or some new way to manipulate pixels through thought or gesture. We were thinking too small. Instead of refining the way we see art, Midjourney is attempting to refine the way we see ourselves. Their new venture, Midjourney Medical, is moving into the realm of the biological, aiming to turn the high-resolution complexity of an MRI into something as casual as a spa visit.

The Golden Pool

The proposed technology, which they call Ultrasonic CT, involves stepping into a shallow pool of golden light. As you descend, a ring of underwater sensors uses echolocation—much like a dolphin—to send ultrasonic waves through your body from every conceivable angle. This produces terabytes of data that computers then convert into a 3D map of your internal organs, accurate to the millimetre. The goal is a sixty-second procedure: you go in, you come out, and you have a complete picture of your health.

It is not an MRI, but the idea is to use compute to reach the same destination.

This is a massive departure from generative art. However, the logic is consistent. Midjourney is built on the ability to take noisy, incomplete data and use massive compute to infer a high-fidelity reality. In art, that means a beautiful image from a text prompt. In medicine, it means taking the 'noise' of ultrasonic waves and turning it into a clear diagnostic map. They are applying the same mathematical intuition to the human body that they applied to the human imagination.

The mechanics of Ultrasonic CT

Underwater ultrasonic sensor rings
Multi-angle wave propagation
Terabyte-scale data processing
Sixty-second scan duration

The partnership with Butterfly Network, which produces ultrasound-on-a-chip, suggests this isn't just a science fiction concept. It is a bet on the convergence of high-end computation and sensor hardware. If they succeed, the medical scan ceases to be a scheduled, intimidating event in a hospital and becomes a routine, almost invisible part of life. The tension lies in whether a company trained on aesthetics can master the absolute precision required by clinical medicine.

Key Takeaway

The next frontier of AI isn't just generating content; it is interpreting the physical reality of our bodies.

02 — Dwarkesh Podcast

The Data Black Hole

Why AI progress is a matter of scale, not just intelligence

By Dwarkesh Patel · 10 min read

Editor's note: An investigation into the massive data disparity between humans and machines.

We often talk about intelligence as a quality of reasoning, but in the world of machine learning, intelligence is increasingly a function of sample efficiency. This is the measure of how much data a system needs to see before it can perform a task competently. If you look at the current state of AI, we haven't actually made much progress in making models more efficient. Instead, we have simply built bigger and bigger pipes to feed them more data.

The Million-Fold Gap

The scale of this disparity is staggering. A human being, from birth to adulthood, will consume roughly 200 million tokens of information. A frontier AI model is trained on hundreds of trillions. This is a million-fold difference. We see these models as brilliant, but their brilliance is actually a brute-force result of consuming more information than a human could process in a thousand lifetimes. They aren't smarter; they are just better fed.

At the centre of the glittering galaxy of AI capabilities is an unimaginably massive black hole of data.

This explains why robotics and self-driving cars are lagging behind language models. A teenager can learn to drive a car after twenty hours of practice. A Tesla or a Waymo requires orders of magnitude more data to achieve the same reliability. The 'intelligence' of the model is trapped by its inefficiency. Until we solve the problem of sample efficiency—teaching machines to learn from a few examples rather than a trillion—robotics will remain a slow-moving industry.

Why sample efficiency matters

Reduces the cost of data acquisition
Enables learning in physical environments (robotics)
Allows for faster deployment in niche domains
Moves AI from brute force to true reasoning

The industry is currently trying to bypass this through 'expert data'—hiring thousands of lawyers, consultants, and specialists to write perfect examples for the models to mimic. We are essentially building a Frankenstein’s monster of human expertise, stitched together with compute. The real breakthrough won't come from more data, but from models that can learn the underlying rules of the world from the small amount of data they already have.

Key Takeaway

AI is currently a brute-force engine of data consumption, not a master of efficient learning.

03 — The Marginalian

The Insomniac's Creative Crucible

Kafka and the strange utility of sleeplessness

By Maria Popova · 7 min read

Editor's note: Exploring the link between the breakdown of sleep and the emergence of original thought.

Insomnia is usually described as a failure of the body, a breakdown of the biological rhythm that keeps us sane. It is a state of anxiety where the mind turns inward, ruminating on its own inability to rest. Yet, for Franz Kafka, sleeplessness was not merely a symptom of distress; it was a condition that facilitated his most intense creative work. In the liminal space between wakefulness and exhaustion, the rigid structures of the conscious mind begin to fail, allowing something else to emerge.

The Fraying of Order

When we are well-rested, our thoughts are organized by logic, social norms, and the practicalities of the day. We are governed by the 'conscious will'. But fatigue acts as a solvent. As the day's cognitive labours wear thin, the boundaries between the conscious and the unconscious blur. The 'unbidden thoughts' that emerge from the recesses of the mind begin to collide in ways that a rested brain would normally prevent. This is where originality often lives: in the chaos of the uncommanded mind.

Sleeplessness comes only because I write.

Kafka viewed his insomnia as a double-edged sword. On one hand, it brought him to the 'nearness of insanity', a state of exhaustion that made writing difficult. On the other, it made him sensitive to the 'imminent possibility of great moments'. He felt that his sleeplessness was a direct consequence of his creative sensitivity—a mind too ablaze with thought to be silenced by sleep. He was not just a victim of insomnia; he was a participant in it.

The tension of the sleepless mind

The struggle between conscious will and unconscious impulse
The anxiety of compromised daytime function
The emergence of unexpected cognitive combinations
The feeling of being 'torn open' by possibility

To embrace insomnia as a creative tool is a dangerous proposition. It requires accepting a certain level of instability. Kafka’s experience suggests that the creative impulse is a way of calibrating how much reality we can actually bear. When the night cuts us to pieces, it is because the protections of the day have been stripped away, leaving us exposed to the raw, unmediated stream of existence.

Key Takeaway

Creativity often thrives in the breakdown of the very structures that allow us to function normally.

04 — The Marginalian

The Search for Meaning in a Material World

Oliver Sacks and the spirit of nature

By Maria Popova · 6 min read

Editor's note: A reflection on how to find purpose when traditional structures fail.

Meaning is rarely something we find waiting for us in a book or a temple. More often, it is something we are forced to construct from the fragments of our own experiences, especially the painful ones. For the neurologist Oliver Sacks, meaning was not a steady state provided by religion or philosophy. Instead, it was a recurring achievement—something he had to re-attain, again and again, through moments of intense connection with the world around him.

The Cosmogenic Spirit

Sacks was a man of science, yet he rejected a purely materialistic view of the universe. He did not believe in a transcendental spirit separate from nature, but he felt a profound sense of awe at the 'deep ordering positivity' of the natural world. To him, the complexity of biology and the vastness of history were not just facts to be catalogued; they were sources of a kind of secular spirituality. He found meaning in the way order emerges from chaos.

There is no love of life without despair of life.

His perspective was shaped by his work with patients whose lives had been shattered by neurological disorders. In observing people who continued to affirm their existence despite devastating physical loss, Sacks found a powerful truth: there is an inextinguishable power of affirmation within the human spirit. Meaning, therefore, is not an abstract concept, but a practical act of love and engagement with the present moment.

Sacks' pillars of meaning

The emergence of order in nature
The resilience of the human spirit in illness
The capacity for love and connection
The continuous act of re-achieving purpose

Ultimately, Sacks suggests that we must formulate meaning for ourselves. It is a personal responsibility. We cannot borrow it from cultural institutions and expect it to hold when the storms of life arrive. We must find it in the 'higher domain' that art, music, and selfless acts reveal to us—a domain that is not separate from nature, but is the highest expression of it.

Key Takeaway

Meaning is not a destination to be reached, but a continuous act of creation and re-creation.

05 — Simon Willison

The Sandbox Revolution

How Datasette Apps are changing data interaction

By Simon Willison · 5 min read

Editor's note: A technical look at the power of sandboxed, interactive data applications.

The traditional way we interact with data is often a choice between two extremes: a static, read-only view or a full-scale, complex application. There is very little middle ground for lightweight, interactive tools that live directly alongside the data. Datasette Apps aims to fill this gap by allowing users to host custom HTML and JavaScript applications within a tightly constrained, sandboxed environment directly on top of their data.

Security through Isolation

The primary challenge in running user-provided code on a sensitive data platform is security. How do you allow a user to run JavaScript without giving them the ability to steal cookies, access local storage, or exfiltrate data to an external server? The solution is a combination of the HTML `<iframe>` sandbox attribute and a strict Content Security Policy (CSP). By locking down the environment at the browser level, the system can run untrusted code while ensuring it remains isolated from the parent application.

The magic combination is a sandboxed iframe plus an immutable CSP header.

Once the environment is secured, the next hurdle is utility. A perfectly secure sandbox is useless if it cannot interact with the data. Datasette Apps solves this by using a controlled communication protocol via `MessageChannel()`. This allows the sandboxed app to send requests to the parent application—such as 'run this specific SQL query'—which the parent can then verify and execute. It is a system of 'permissioned interaction' rather than open access.

Key features of Datasette Apps

Self-contained HTML/JS applications
Strictly sandboxed iframe execution
Immutable Content Security Policies
Secure SQL querying via MessageChannel

This architecture has implications far beyond simple data visualisation. It creates a pattern for how we might interact with LLM-generated 'artifacts'. Imagine an AI that doesn't just write code, but deploys a fully functional, sandboxed application that has direct, secure access to a persistent relational database. This moves us from chat-based interfaces to a world of interactive, data-driven tools that are both powerful and safe.

Key Takeaway

True interactivity in data requires a architecture of controlled, sandboxed communication.

06 — Stratechery

The Anthropic Paradox

Safety as a business strategy in the age of AI

By Stratechery · 4 min read

Editor's note: An analysis of how corporate values become strategic assets.

In the high-stakes race to develop artificial intelligence, 'safety' is often discussed as a technical hurdle or a moral obligation. But for companies like Anthropic, safety has become something more: a strategic superpower. By positioning themselves as the 'safety-first' alternative to other frontier labs, they have created a brand identity that allows them to navigate political and regulatory pressures in a way their competitors cannot.

The Alignment of Interests

The recent controversy surrounding the model 'Fable'—which faced export controls from the Trump administration—highlights this tension. When a company's core mission is safety, every business decision they make can be framed as a safety decision. This creates a powerful feedback loop. Even when their actions appear self-serving or commercially driven, they can be defended through the lens of their commitment to AI alignment. This makes them incredibly resilient to external criticism.

Anthropic's safety superpower is that every action it takes looks, from the outside, to be self-serving, even as the company becomes convinced its motivations are pure.

This is a classic example of how a company's internal culture can become its most potent external weapon. In a landscape where the public and the government are increasingly terrified of AI, being the 'responsible' player is not just a moral stance; it is a way to secure market share and regulatory favour. It allows them to challenge government mandates by claiming those mandates might actually compromise the safety of the models themselves.

Strategic advantages of the 'Safety' brand

Increased trust from enterprise clients
Political leverage in regulatory debates
A shield against criticism of business decisions
Differentiation in a crowded market

However, this strategy carries its own risks. If the gap between the 'safety' rhetoric and the actual technical reality ever widens too far, the backlash will be catastrophic. For now, Anthropic has successfully turned a technical problem into a competitive advantage, proving that in the era of AI, the most important thing you can build is not just a model, but a mythos.

Key Takeaway

In emerging industries, the perception of responsibility is as much a competitive tool as the technology itself.