The Deep Feed

01 — Dwarkesh Podcast

The Compute Trap

Why the path to trillion-dollar AI revenues requires a massive spike in hardware costs

By Dwarkesh Patel · 8 min read

Editor's note: As AI labs chase astronomical revenue targets, they face a mathematical bottleneck: the cost of the silicon required to power them may soon skyrocket.

The current trajectory of AI labs looks less like a steady climb and more like a vertical ascent. Anthropic, for instance, is seeing revenues grow tenfold year over year. If this pace holds, we are looking at a company generating $1 trillion in revenue by the end of next year. Such numbers suggest a world where AI is not just a tool, but the fundamental substrate of the global economy. However, this growth creates a tension between revenue and the physical reality of compute. To hit these targets, labs cannot simply rely on more hardware; they must solve a complex equation involving margins, inference costs, and the sheer price of electricity and silicon.

The Inference Dilemma

There is a specific tension in how labs spend their money. Traditionally, the goal of generating revenue from inference—running the models for users—is to fund the next generation of training. Investors give money to build bigger, smarter models. But if a lab starts spending the majority of its compute on inference rather than training, it signals a stagnation. It suggests that the era of massive capability leaps is over and the company has transitioned into a mere cloud provider. To avoid this, labs must ensure that every dollar earned from a user today directly facilitates a smarter model tomorrow. This requires margins to expand at an almost impossible rate.

If you’re spending most of your compute on inference, you’re declaring that AI progress has stalled.

This brings us to the most likely driver of the next phase: the rising cost of compute. We are already seeing a divergence between 'spot prices'—the cheap, leftover capacity on the market—and the premium prices labs must pay for guaranteed, secure, and massive-scale hardware. Google is reportedly paying $900 million a month for a specific tranche of GPUs. This is double the market rate. As models become more capable, they become more valuable to use. If a model can perform the work of a high-level software engineer, the cost of the hardware running that model should logically reflect that value.

Drivers of Compute Inflation

Shift from spot instances to high-security, guaranteed capacity
Increased demand for specialized inference-heavy hardware
The massive value-add of human-level reasoning per unit of silicon

The consequence of this price surge is a widening moat. If the cost of compute rises by 15x because the models are so much more productive, new entrants will find it impossible to compete. A startup with no revenue cannot afford the entry fee to play in the league of the giants. We are moving toward a winner-take-all economy where the barrier to entry is not just code, but a massive, expensive, and increasingly scarce physical resource.

Key Takeaway

The economic success of AI depends on a paradox: the more useful the models become, the more expensive the hardware required to run them will likely be.

02 — Simon Willison

The Word Worm

How prompt injection turned Microsoft Word into a carrier for self-replicating malware

By Simon Willison · 5 min read

Editor's note: Security researchers have identified a new class of attack that uses LLM integration to spread malicious instructions through documents.

We have long understood the risks of prompt injection—the ability to trick an AI into ignoring its instructions and following new ones. Usually, this is a contained event: a user interacts with a malicious prompt and the AI misbehaves. However, a new variant discovered by Håkon Måløy changes the nature of the threat. By targeting Microsoft Word's Copilot integration, an attacker can create a document that acts as a biological worm, infecting every subsequent document it touches.

The Mechanism of Infection

The attack works by hiding instructions within a document—perhaps through white-on-white text or metadata that a human eye misses but an LLM processes. When a user asks Copilot to summarise or edit that document, the AI interprets the hidden text as a command. The command instructs the AI to copy those very same instructions into the new document being drafted. The result is a self-replicating loop. The original malicious file doesn't even need to be present for the infection to continue; the 'virus' is now part of the user's own work product.

This is the first time we have seen a prompt injection that deliberately copies instructions to self-replicate itself.

This represents a shift from 'exploiting a model' to 'exploiting a workflow.' In a modern enterprise, documents are constantly being passed between assistants, summarised for executives, and integrated into larger reports. If the AI is the engine of this movement, then the AI is also the vector for the spread. Once a document becomes a carrier, the infection moves at the speed of business communication.

Why this is different from traditional malware

It uses the legitimate reasoning capabilities of the AI to propagate
The payload is natural language, not binary code
It bypasses traditional antivirus by hiding in plain sight within text

Microsoft has been notified, but the difficulty in fixing this lies in the fact that it isn't a bug in the code, but a feature of how LLMs process context. To stop the worm, you have to stop the AI from following instructions found in its source material—a task that would significantly neuter the utility of the tool. We are entering an era where the very intelligence we rely on to process information can be used to corrupt it.

Key Takeaway

Prompt injection is no longer just a way to trick a chatbot; it is a method for creating self-propagating digital worms within professional workflows.

03 — Simon Willison

The Cryptographic Stress Test

Can Claude find the flaws in our digital foundations?

By Simon Willison · 6 min read

Editor's note: Anthropic's researchers are using LLMs to hunt for mathematical weaknesses in encryption, proving that AI can be a powerful tool for high-level research.

The field of cryptanalysis is traditionally the domain of the world's most brilliant mathematicians, working through dense proofs and complex algebraic structures. It is a slow, methodical process. However, Anthropic researchers have demonstrated that LLMs, when prompted correctly, can act as highly capable research assistants in this field. Using a system called Claude Mythos, they successfully identified mathematical flaws in both HAWK and certain versions of AES.

The Art of the Prompt

What is most striking about this work is not just the result, but the process. The researchers found that the models often default to a sense of 'impossibility'—they tend to believe a problem cannot be solved and therefore stop trying. Success required intense, human-led prompting to push the model past its own perceived limits. The researchers had to explicitly tell the model: 'find something that is worth publishing.' They had to encourage it to look for genuine research breakthroughs rather than low-hanging fruit.

The main human interventions were to encourage it not to give up and 'find something that is worth publishing'.

This experiment cost roughly $100,000 in API fees and 60 hours of compute time. It suggests that the value of an LLM in scientific research is not in its ability to work autonomously, but in its ability to expand the search space for a human expert. The model doesn't replace the mathematician; it provides a tireless, high-speed engine for testing hypotheses that a human might take months to verify.

Lessons from Claude Mythos

LLMs require 'persistence prompting' to overcome mathematical hesitation
The cost of high-level AI research is significant but scalable
AI is best used as a partner in 'hard' science rather than a solo agent

As we move toward post-quantum cryptography, the stakes could not be higher. We are currently redesigning the very algorithms that protect the world's data. Having a new, massive capability for cryptanalysis—even one driven by AI—is a double-edged sword. It allows us to stress-test our new standards before they are deployed, but it also provides a roadmap for those looking to break them.

Key Takeaway

AI is transitioning from a tool for generating text to a tool for discovering mathematical truth, provided humans know how to push it past its own limits.

04 — The Marginalian

The Ego State Map

Eric Berne and the psychology of our social games

By Maria Popova · 10 min read

Editor's note: A look back at the 1964 classic that decoded the repetitive, self-defeating patterns in human interaction.

Most human suffering is not the result of grand tragedies, but of a fundamental clumsiness in how we seek what we want. We often move through the world without a clear understanding of our own motivations, reacting to others with patterns that are both predictable and self-sabotaging. In 1964, psychiatrist Eric Berne gave these patterns a name: 'games.' His work, *Games People Play*, sought to map the hidden currents of social interaction through a framework called transactional analysis.

The Three Ego States

Berne proposed that every individual operates from one of three 'ego states.' The Child is the seat of our spontaneity, vulnerability, and creative vitality. The Parent is the part of us that mimics the attitudes and responses we observed in our own caregivers. The Adult is the rational, competent part of us capable of making objective decisions. While we possess all three, we often fail to coordinate them, leading to social friction and emotional confusion.

Communication proceeds smoothly as long as transactions are complementary; the wounding happens when the lines of communication cross.

A 'game' occurs when a social interaction is not a direct exchange between two people in stable ego states, but a series of 'ulterior transactions.' For example, one person may issue a stimulus from a 'Parent' state while concealing a 'Child' need. When the other person responds to the surface message, the first person reacts negatively to the response, frustrating both parties. This creates a predictable, repetitive loop that provides a sense of connection—however hollow—at the expense of true intimacy.

The Ego State Framework

Child: Spontaneous, emotional, and creative
Parent: Mimetic, judgmental, or nurturing
Adult: Rational, data-driven, and objective

Though Berne's language is dated, his insight remains sharp: we often choose the safety of a predictable game over the terrifying vulnerability of an honest relationship. By understanding these patterns, we gain the ability to stop playing them. The goal is not to eliminate the Child or the Parent, but to allow the Adult to navigate the interactions, turning repetitive games into genuine, transparent communication.

Key Takeaway

Social friction is often the result of 'games'—repetitive, unconscious patterns where we hide our true needs behind mismatched psychological roles.

05 — Simon Willison

The Post-Quantum Transition

Why the current era is the perfect storm for AI cryptanalysis

By Matthew Green · 3 min read

Editor's note: As the world moves toward new cryptographic standards, the arrival of AI-driven analysis creates a unique moment of both risk and opportunity.

We are currently in the middle of one of the most significant shifts in the history of digital security. For decades, our world has been protected by specific mathematical problems—primarily those based on Elliptic Curve cryptography and RSA. However, the looming threat of quantum computing is forcing a global migration toward 'post-quantum' algorithms. These are new standards, such as HAWK, designed to withstand the unique capabilities of quantum machines.

A Perfect Storm

This transition period is uniquely volatile. We are implementing new, unproven standards at scale, making this the ideal time for a massive new capability in cryptanalysis to emerge. If AI models become proficient at finding mathematical flaws, they will be testing these new standards exactly when they are most vulnerable. It is a high-stakes race between the architects of new security and the automated hunters of its weaknesses.

If there was ever a perfect time for a massive new public cryptanalysis capability to come on line, we’re in it.

The outcome of this collision will define the next century of privacy. In the best-case scenario, AI-driven cryptanalysis acts as a rigorous stress test. It allows researchers to find and fix flaws in post-quantum algorithms before they are widely deployed, making our digital foundations more robust than ever. In the worst case, AI could undermine the very 'hard problems' we rely on to keep the world's secrets.

The Stakes of the Transition

The obsolescence of RSA and EC-based cryptography
The deployment of unproven post-quantum standards
The emergence of AI as a primary tool for mathematical attack

Whether this leads to a more secure world or a more transparent one depends on whether the defenders can use AI more effectively than the attackers. We are no longer just defending against human ingenuity; we are defending against automated, high-speed mathematical discovery.

Key Takeaway

The global shift to post-quantum cryptography coincides with the rise of AI cryptanalysis, creating a critical window for either unprecedented security or total systemic failure.

06 — Simon Willison

The MCP Frontier

Bridging the gap between chat interfaces and custom data

By Simon Willison · 4 min read

Editor's note: The Model Context Protocol (MCP) is changing how we connect LLMs to our own tools, but the setup remains a hurdle for the average user.

The true power of a Large Language Model is not in its ability to write poetry, but in its ability to interact with the world. Currently, most users interact with AI through isolated chat windows. The Model Context Protocol (MCP) aims to change this by providing a standardized way for models to access local data, custom tools, and specific software environments. It is the bridge between a clever chatbot and a functional agent.

The Integration Gap

While the promise of MCP is significant, the practical application remains difficult. Both Claude and ChatGPT have begun to support MCP servers, allowing them to pull in context from custom sources. However, the process of setting these up is not yet intuitive. It requires a level of technical configuration that keeps the most powerful version of these tools out of reach for the majority of users. We are in a period where the plumbing is being laid, but the faucets are still hard to find.

Connecting a custom MCP server to Claude and ChatGPT's standard chat interfaces is possible, but can take quite a few steps.

The value of solving this friction cannot be overstated. If an LLM can seamlessly access your local files, your database, or your specific project management tools via a standardized protocol, it ceases to be a consultant and becomes an operator. The transition from 'chatting' to 'doing' depends entirely on the ease with which we can connect these models to our existing digital lives.

The Promise of MCP

Standardized access to local and remote data
Reduced friction between LLMs and specialized software
The foundation for true autonomous AI agents

As the protocol matures, we will likely see a surge in custom MCP servers—specialized 'knowledge modules' that users can plug into their preferred AI. This will turn the AI interface into a modular cockpit, where the user defines the scope of the model's expertise by the servers they choose to connect.

Key Takeaway

The Model Context Protocol is the essential infrastructure for turning LLMs from conversationalists into functional agents, though its current complexity limits its adoption.