The Deep Feed

01 — Not Boring

The Return of the Miracle

How we are rewriting the rules of energy, intelligence, and biology

By Packy McCormick · 12 min read

Editor's note: A look at the rare moments when scientific progress moves from theoretical possibility to physical reality.

For decades, the idea of domestic nuclear innovation felt like a relic of a previous century, a stalled engine of American industry. That changed this week. Antares announced that its Mark-0 low power reactor reached criticality at the Idaho National Lab. This is not merely a technical milestone; it is the first time a novel reactor design has undergone a fueled test in over half a century. The achievement marks a shift from theoretical physics to practical engineering. The team has moved past the stage of merely making neutrons; they are now positioned to start making electrons. This transition from laboratory success to electricity production represents a massive leap in the timeline of energy independence.

The Knowledge Bottleneck

While energy is being reinvented at the atomic level, the intelligence revolution is hitting a different kind of wall: the data wall. Many companies are currently throwing AI tools at their problems, only to find that the tools are only as good as the information they can access. If your internal data is scattered across different departments or sits in outdated spreadsheets, your AI will simply produce incorrect answers with greater speed. The problem is not the model; it is the context. Without a central, governed layer of truth, AI becomes a liability rather than an asset.

Fix the input once with Guru, and it improves every output, across every tool your team uses.

Companies like Spotify and Brex are addressing this by building a reliable knowledge layer. The goal is to ensure that when an AI agent queries a company's internal state, it is drawing from verified, maintained, and accurate sources. This is the difference between an AI that guesses and an AI that knows. In the race to implement automation, the winners will not be those with the most expensive models, but those with the cleanest, most accessible data architectures.

Rewriting the Biological Script

The most startling breakthrough, however, occurs at the cellular level. NewLimit, an epigenetic reprogramming company, is moving toward human trials with a therapy that targets the very mechanisms of ageing. Their work focuses on restoring the intrinsic ability of cells to withstand stress and regenerate. In animal trials, the results were binary: mice treated with their RNA-encoded therapy showed the resilience of young subjects, even when faced with extreme metabolic stress. This is not about metabolic speed; it is about cellular integrity.

The NewLimit Approach

Targeting epigenetic reprogramming rather than just symptom management
Using lipid nanoparticles to deliver transcription factors
Utilising the Ambrosia AI model to predict optimal gene combinations
Moving from mice to human trials in Australia by 2027

The engine behind this is Ambrosia, an AI model that ingests gene-function literature and protein sequences to propose specific cellular states. By running these models in reverse, researchers can specify a desired cell state and receive the exact combination of factors needed to achieve it. This represents a move away from trial-and-error biology toward a predictive, engineering-led discipline. We are entering an era where the limitations of the human body may become design problems rather than inevitable fates.

Key Takeaway

True progress in the next decade will be defined by our ability to control the fundamental inputs of our world: atoms, bits, and cells.

02 — Simon Willison

The Lethal Trifecta

Why OpenAI's Lockdown Mode is a necessary response to the reality of data theft

By Simon Willison · 6 min read

Editor's note: A technical breakdown of why preventing data exfiltration is the most effective way to secure LLM systems.

OpenAI has introduced 'Lockdown Mode', a feature designed to mitigate one of the most significant risks in the age of large language models: data exfiltration via prompt injection. For a long time, the industry has focused on preventing the injection itself—the moment a malicious prompt tricks an AI into behaving incorrectly. But preventing the injection is difficult because the prompt can be hidden anywhere, from a cached webpage to an uploaded PDF. Lockdown Mode takes a more pragmatic approach. Instead of trying to stop the trick, it stops the theft by limiting the outbound network requests an LLM can make.

Understanding the Trifecta

To understand why this matters, one must understand the 'Lethal Trifecta'. A truly dangerous AI system is one that possesses three specific capabilities: access to private data, exposure to untrusted content, and a method to transmit that data back to an attacker. If all three exist, the system is a security nightmare. An attacker can use untrusted content to trigger a prompt injection, which then uses the system's access to private data to steal information, and finally uses the system's network access to send that information to an external server.

The only way to solve the trifecta is to cut off one of the three legs.

Lockdown Mode attacks the third leg: the exfiltration vector. By restricting the ability of the model to make outbound calls, OpenAI is creating a deterministic barrier. This is a significant advantage because, unlike the AI's reasoning, which can be subverted by clever language, network permissions are hard rules. They are not evaluated by the AI; they are enforced by the architecture. This makes the defense much more robust against even the most devious attacks.

The Trade-off of Utility

Security is rarely free. Using Lockdown Mode comes with a cost to functionality. If your workflow relies on the AI being able to fetch real-time data from various web services or interact with external APIs, that utility will be diminished. This is why OpenAI's CISO has noted that the mode is not intended for everyone. It is a tool for those with an elevated risk profile—people handling sensitive corporate data or working in high-stakes environments where the cost of a leak far outweighs the convenience of an interconnected AI.

Who should use Lockdown Mode

Users handling proprietary corporate intelligence
Developers working with sensitive user datasets
Organizations in highly regulated sectors like finance or healthcare
Anyone whose primary risk is data exfiltration rather than model inaccuracy

The existence of this mode is a quiet admission that default AI settings are inherently vulnerable. As we integrate these models deeper into our business processes, the question is no longer whether we can trust the AI's logic, but whether we can trust the environment in which that logic operates. Security in the AI era will be defined by the strength of these digital cages.

Key Takeaway

In AI security, don't try to fix the unpredictable mind; fix the predictable pipes.

03 — Simon Willison

The Velocity Trap

Managing the collision between AI speed and engineering stability

By Simon Willison · 7 min read

Editor's note: An analysis of the internal organisational tension between rapid AI adoption and the preservation of technical reliability.

Modern engineering teams are currently split into two camps, often working side-by-side but seeing two different realities. On one side are the enthusiasts. They see the discontinuous leaps in capability that AI provides and understand that this is not a standard technology cycle. You cannot wait for the dust to settle when your competitors are using AI to outpace your entire development roadmap. For them, the threat is existential: move too slowly, and you become obsolete before you even understand the new tools.

The Entropy of Speed

On the other side are the skeptics. They are not anti-technology; they are pro-reliability. They see the danger in shipping code faster than human engineers can actually read or comprehend it. When you accelerate development to a point where no single person has full context of the system, you are making massive withdrawals from a trust account. You are trading long-term stability for short-term velocity. The result is often a system that is incomprehensible, a product that behaves inconsistently, and an engineering culture that burns out under the weight of unpredictable failures.

When you ship code faster than engineers can read it, you are making withdrawals from a trust account that took years to build.

Both groups are correct. The enthusiasts are right about the market risk, and the skeptics are right about the technical risk. This creates a profound organisational tension. If you lean too far into the enthusiasm, your systems become brittle and unmanageable. If you lean too far into the skepticism, you risk being outcompeted by more agile players. The challenge for leadership is not to choose a side, but to design a way for these two groups to communicate.

Bridging the Reality Gap

The core issue is the lack of a natural feedback loop. The enthusiasts see the output; the skeptics see the debt. To solve this, organisations need to build mechanisms that force these two perspectives to collide in a controlled way. This might mean implementing stricter observability requirements for AI-generated code, or creating 'stability budgets' that limit how much speed can be traded for technical debt in a given quarter.

Strategies for Alignment

Implement rigorous testing for AI-assisted code deployments
Create shared metrics that track both velocity and system reliability
Formalise the role of 'skeptic' in the deployment process
Establish clear thresholds for when technical debt must be repaid

The goal is to create a shared reality. An organisation that can harness the speed of AI without losing the ability to understand its own systems will have a massive advantage. The winner is not the fastest team, but the fastest team that can still explain why their system works.

Key Takeaway

Speed without comprehension is just a faster way to fail.

04 — Stratechery

The Gatekeeper's End

How power is shifting from legacy institutions to high-bar digital creators

By Stratechery · 15 min read

Editor's note: An examination of the shifting power dynamics in tech and entertainment, from Google's capital moves to the rise of YouTube directors.

The structures of power in the digital economy are shifting in ways that defy traditional industry logic. In the tech sector, we see Google making curious moves, such as issuing equity to Berkshire Hathaway. This suggests a future where capital itself becomes the ultimate commodity, and even the largest tech giants must navigate a world where liquidity and massive capital reserves are the primary drivers of dominance. Meanwhile, Microsoft's relationship with OpenAI continues to redefine the competitive landscape, moving the focus from simple software products to agentic platforms.

The Hollywood Disruption

Perhaps the most visible shift is occurring in entertainment. For decades, Hollywood has functioned as a collection of gatekeepers—studios, agents, and distributors who decided what reached the screen. That model is being dismantled by Gen Z YouTubers. These are not just 'content creators'; they are directors and producers who are now beating major studio productions at the box office. They are not just playing in a different sandbox; they are building a better one.

Succeeding on YouTube is a much higher bar than the gates that currently govern Hollywood.

The reason for this is simple: the audience feedback loop on YouTube is immediate and brutal. To succeed there, you cannot rely on the prestige of a studio or a massive marketing budget. You must capture attention through sheer quality and relevance. This has created a class of creators who are more attuned to audience psychology than any traditional studio executive. Hollywood is finding that its gatekeeping mechanisms are increasingly irrelevant in a world where the audience has already decided who the stars are.

The New Competitive Standard

This shift is not limited to entertainment. It is a broader trend of decentralisation where the 'bar' for success is being raised by the platforms themselves. Whether it is the ability to command massive capital or the ability to command massive attention, the old intermediaries are being bypassed. The power is moving to those who can operate at the intersection of high-scale technology and high-intensity creativity.

Key Drivers of the Power Shift

The decentralisation of distribution via platforms like YouTube
The shift from software products to agentic, capital-intensive platforms
The rising importance of capital as a commodity in the tech sector
The direct connection between creators and global audiences

For agency owners and business leaders, the lesson is clear: the gatekeepers are losing their grip. The new economy rewards those who can build direct, high-value relationships with their audience or their users, bypassing the traditional structures that once provided stability and protection.

Key Takeaway

When the bar for entry rises, the gatekeepers become obsolete.

05 — Simon Willison

The Architecture of Trust

Why running code safely requires more than just a simple wrapper

By Simon Willison · 9 min read

Editor's note: A technical deep-dive into the necessity of sandboxing in an era of untrusted, AI-generated code.

The plugin model is one of the most powerful ways to extend software, allowing applications to grow without constant core updates. However, it introduces a fundamental security flaw: plugins usually execute with the same privileges as the host application. A single malicious or poorly written plugin can read private files, connect to the network, or crash the entire system. As we move toward a world where code is increasingly generated by AI and executed on-the-fly, the need for a robust, isolated sandbox is no longer optional; it is a requirement.

The Sandbox Requirements

A true sandbox must do more than just isolate; it must provide control. It needs to manage memory and CPU limits to prevent a single process from hogging resources. It must strictly control filesystem access, allowing only specific, pre-approved files to be read or written. Most importantly, it must control network access. A sandboxed piece of code should never be able to communicate with the outside world unless it goes through a layer that the host application fully manages.

A sandbox isn't much use if I can't carefully expose selected platform features to the code that it's running.

The challenge has always been finding a solution that is both secure and easy to use. Many existing projects are either too heavy, like full V8 engines, or too difficult to embed in a Python environment. This is where WebAssembly (WASM) becomes a critical tool. WASM was designed from the ground up for isolation, providing a highly efficient and secure way to run untrusted code in a way that is already proven in the web browser.

MicroPython and the WASM Solution

The recent experimentation with MicroPython in WebAssembly offers a promising path forward. By compiling a lean Python interpreter to WASM, developers can run Python code in a highly constrained, secure environment. This allows for the flexibility of a dynamic language like Python while maintaining the strict security guarantees of the WASM runtime. It solves the 'plugin paradox': how to allow users to extend your software without giving them the keys to your house.

The Benefits of MicroPython + WASM

Strict memory and CPU limits
Controlled filesystem and network access
Easy integration with existing Python ecosystems
High performance with low overhead

As we build more complex, agentic systems that must execute arbitrary code to solve problems, the architecture of trust will become the most important part of the stack. We are moving from a world of 'trusting the developer' to a world of 'verifying the execution'.

Key Takeaway

In a world of untrusted code, isolation is the only true security.

06 — The Marginalian

The Lottery of Existence

Finding perspective in the sheer improbability of being human

By Maria Popova · 10 min read

Editor's note: A philosophical reflection on the mathematical impossibility of our own lives.

There is a persistent sense that we are living through uniquely difficult times, a feeling that the modern world is particularly heavy to bear. This is often an error of proximity. We live too close to our own immediate struggles to see the broader horizon of chance. When we pull back from the mundanity of our daily routines—the coffee, the deadlines, the endless to-do lists—we find something far more startling: the sheer, mathematical improbability of our own existence.

The Geometry of Chance

Physicist Alan Lightman explores this through the lens of probability. Every single human being is the result of a cosmic lottery with odds so astronomical they border on the impossible. Consider the arrangement of human DNA: there are more possible arrangements of our genetic code than there are atoms in the observable universe. Each of us represents one specific, winning combination out of a near-infinite field of possibilities. We are not just a biological occurrence; we are a triumph against staggering odds.

If you took a very long ruler that stretched from here to the planet Pluto, one inch of that distance would be you. The rest would be other possible human beings that could have been, but never were.

This visualization puts our uniqueness into perspective. The vast majority of 'possible' humans will never exist. We are the ones who won the lottery. To be conscious, aware, and capable of experiencing the universe is a gift of chance that we have no right to, yet we receive it every single day. This realization serves to strip away the absurdity of our small entitlements and replace them with a sense of cosmic overwhelm.

A Responsibility to Awe

If hope is the belief that the improbable is possible, then every living person is a living axiom of hope. We are the evidence that the wildest bets can actually win. This perspective does not make our problems disappear, but it changes our relationship to them. It moves us from a posture of entitlement to one of responsibility—a responsibility to acknowledge the miracle of our own presence.

Lessons from the Improbable

Distance yourself from immediate predicaments to see the larger scale
Recognize that existence itself is a statistical anomaly
Replace entitlement with a sense of awe
Understand that being alive is the most extraordinary luck you will ever experience

In the end, the only adequate response to the sheer improbability of life is a simple, quiet acknowledgement of the luck involved. When we stop taking our existence for granted, we find that the world is not just a place of struggle, but a place of constant, miraculous occurrence.

Key Takeaway

Existence is not a right, but a winning lottery ticket.