The Deep Feed

01 — Not Boring

The Silicon Sovereignty: Why Westmag Matters

The strategic necessity of American motors in an electric age

By Packy McCormick · 12 min read

Editor's note: A look at why domestic manufacturing is a security requirement, not just an economic preference.

The era of cheap, globalised components is ending. For decades, the economic logic was simple: design high-value intellectual property in the West and outsource the heavy lifting—the motors, the actuators, the physical movement—to the massive, subsidised factories of China. This worked until it didn't. Today, the reliance on foreign-made hardware for autonomous systems is a strategic vulnerability. If every drone, robot, and electric vehicle in America relies on a black-box component manufactured by a geopolitical rival, then the software control we pride ourselves on is built on sand.

The Manufacturing-Design Loop

Westmag, a company emerging from the industrial fringes of South San Francisco, is betting on a different path. They aren't just making hats; they are building the electric stack. The core argument for domestic manufacturing isn't just about avoiding 'bugged' hardware or foreign interference. It is about the feedback loop between design and production. When you build the motor, you understand the limits of the physics. You learn how to iterate on the hardware to support better software. You cannot maintain design leadership if you have no grip on the physical reality of how your machines move.

In the Electric Era, maintaining design leadership without manufacturing leadership is not a coherent strategic position.

This is a direct challenge to the traditional comparative advantage model. Critics argue that competing with China's scale is a fool's errand. They are right, if you are trying to win on volume alone. But Westmag is not playing the volume game; they are playing the integration game. By building motors and actuators in the West, they allow American robotics companies to iterate faster and more securely. It is about creating a domestic ecosystem where the hardware and the intelligence are developed in tandem, rather than as two disconnected layers.

The Strategic Drivers for Domestic Hardware

Security: Reducing the risk of hardware-level vulnerabilities in autonomous systems.
Iteration Speed: Closing the gap between software requirements and physical implementation.
Resilience: Building supply chains that are not subject to geopolitical blackmail.

The transition to an electric, automated economy requires more than just better code. It requires a physical foundation that is as reliable as the logic it executes. As we move toward a world of drone-filled skies and domestic robots, the companies that control the magnets and the copper will hold as much power as those that control the algorithms.

Key Takeaway

Control over the physical stack is the only way to secure the digital future.

02 — Simon Willison

The Token Tax: Uber's AI Budget Crisis

Managing the explosive costs of agentic coding

By Simon Willison · 6 min read

Editor's note: A practical look at how massive enterprises are struggling to contain the costs of AI-driven productivity.

Uber recently discovered a hard truth about the AI revolution: it is incredibly expensive. The company reportedly blew through its entire 2026 AI budget in just four months. The culprit wasn't inefficient marketing or failed R&D; it was the sudden, massive adoption of agentic coding tools. When engineers move from simple autocomplete to autonomous agents like Claude Code or Cursor, the volume of tokens consumed doesn't just increase—it explodes. These tools don't just suggest a line of code; they read entire repositories, reason through bugs, and rewrite entire modules. This level of reasoning comes at a steep per-token price.

The $1,500 Limit

In response, Uber has implemented a monthly cap of $1,500 per employee for agentic coding software. This is a pragmatic move. It acknowledges that while these tools provide immense value, they cannot be allowed to run unchecked. Interestingly, the limit applies per tool, meaning an engineer using two different agents has a $3,000 monthly allowance. At current market rates, this represents roughly 11% of a median US software engineer's salary. It is a calculated trade-off: the company is willing to pay a significant premium for speed, provided that premium is capped.

The sudden popularity of token-burning coding agents has turned AI from a cost-saving tool into a significant budgetary line item.

This shift changes the way we think about developer productivity. Previously, software tools were largely one-time capital expenditures or low-cost subscriptions. AI agents are variable costs that scale directly with usage. If an engineer spends all day in a deep reasoning loop, the cost of their work increases. This introduces a new tension in management: how do you encourage the use of powerful tools without incentivising wasteful, 'token-maxxing' behaviour?

The Economic Reality of AI Agents

Variable Cost Scaling: Productivity is now tied to real-time compute expenditure.
The Agency Premium: Reasoning models cost significantly more than standard LLMs.
Management Tension: Balancing the speed of agents against the predictability of budgets.

For the agency owner, this is a warning. As you integrate these tools into your own workflows, the 'free' or 'cheap' era of AI is ending. You must account for the fact that the more intelligent the tool, the more it will cost to run. Efficiency in the AI era will not just be about how much work you get done, but how much compute you consume to get it done.

Key Takeaway

AI productivity is no longer a fixed cost; it is a variable expense that requires active management.

03 — Stratechery

The Ghost in the Machine: The Limits of the AI PC

Why Nvidia's new silicon might be a step backward

By Stratechery · 10 min read

Editor's note: An analysis of why local AI hardware may be failing to meet the needs of the agentic era.

Nvidia is attempting to do for the personal computer what it did for the data centre: make silicon the centre of the AI universe. The new RTX Spark chip, developed alongside Microsoft, is designed to bring high-performance AI directly to the edge. On paper, the specs are impressive—massive GPU cores and significant memory bandwidth. The goal is to allow autonomous agents to run locally on your laptop, providing unmetered intelligence without the latency or privacy concerns of the cloud. But there is a fundamental mismatch between this hardware and the direction AI is actually moving.

The Reasoning vs. Agentic Gap

The AI landscape has shifted rapidly. Three years ago, we were in the era of simple chatbots, where local inference made sense. Then came the reasoning era, which demanded massive memory to handle long context windows. Now, we have entered the agentic era. Agents are not just predicting the next word; they are performing complex, multi-step tasks that require heavy CPU performance to manage logic, tool-calling, and system interaction. The RTX Spark, however, prioritises GPU die space at the expense of the CPU. It is a chip built for a version of AI that is already becoming obsolete.

The ideal setup for a local agent is strong local CPU performance paired with cloud-based inference, not a GPU-heavy chip that lacks processing power.

Furthermore, there is the software problem. Windows on ARM remains a compromise. For an AI PC to be useful, the software stack must be as seamless as the hardware. If the local agent is bogged down by OS inefficiencies or limited by the inability to call out to the cloud effectively, the massive GPU becomes a wasted asset. We are seeing a collision between the desire for 'local intelligence' and the reality that the most powerful intelligence remains in the data centre.

Why the AI PC Faces Headwinds

Architectural Mismatch: Over-investment in GPU cores when agents need CPU logic.
Software Friction: The ongoing struggle of Windows on ARM ecosystems.
Cloud Dominance: The inherent advantage of massive, remote model parameters.

Nvidia and Microsoft are betting that users want a self-contained intelligence engine. But the reality of the agentic era suggests that the PC will increasingly act as a sophisticated interface—a terminal that manages local tasks while leaning on the cloud for the heavy cognitive lifting. The 'AI PC' might end up being less about the silicon inside the box and more about the connectivity of the system.

Key Takeaway

Hardware cannot solve the intelligence problem if it is built for the wrong era of software.

04 — Lenny's Newsletter

The Mirror of Gemini: Cloning the Self

The technical and emotional reality of AI avatars

By Claire Vo · 8 min read

Editor's note: A first-hand account of the speed and the strangeness of digital identity replication.

The ability to replicate one's likeness has moved from the realm of high-budget Hollywood VFX to a fifteen-minute task on a smartphone. Using Google's Gemini Omni and the Flow interface, it is now possible to scan a face, generate a storyboard, and produce a high-fidelity video of an AI clone. This isn't just a tool for content creators; it is a fundamental shift in how we perceive digital presence. When a video can be generated with zero production skills, the scarcity of 'presence' vanishes. The question is no longer whether we can clone ourselves, but what happens to the value of the original.

The Uncanny Valley of Emotion

The process is remarkably efficient. A user can move from a physical scan to a complete, edited video in minutes. However, the efficiency hides a persistent technical hurdle: the uncanny valley. While the visual likeness can be near-perfect, the emotional resonance often fails. The AI clone might nail the facial structure but miss the micro-expressions that signal genuine human intent. The physics of movement—the way a head tilts or a shoulder shifts during a laugh—can feel slightly off, creating a sense of unease rather than connection.

The uncanny-valley moments occur when an AI clone fails to nail the subtle physics of human emotion.

This creates a paradox of production. We are gaining the ability to produce massive amounts of high-quality video content, but we are losing the 'human' signal that makes video effective. If an agency owner can generate a hundred personalised video messages for clients in an afternoon, the clients will eventually realise they are talking to a mathematical model. The value of video will shift from the ability to produce it to the ability to prove it is real.

The Workflow of Digital Replication

Capture: Using mobile devices to create a high-resolution facial scan.
Storyboard: Using LLMs to act as creative producers and scriptwriters.
Generation: Converting text prompts into consistent video scenes.
Assembly: Stitching disparate AI clips into a coherent narrative.

As these tools become ubiquitous, the barrier to entry for video production will effectively drop to zero. This will lead to an explosion of content, but it will also lead to a crisis of trust. In a world where anyone can be anyone on screen, the most valuable asset will be the one thing AI cannot yet simulate: a verifiable, lived experience.

Key Takeaway

When presence becomes cheap, authenticity becomes the only premium.

05 — The Marginalian

The Moral Weight of Failure

Václav Havel and the cost of self-deception

By Maria Popova · 15 min read

Editor's note: A profound reflection on the difference between tactical success and moral integrity.

There is a specific kind of failure that does not offer the comfort of a learning curve. It is not the mistake that leads to growth, nor the setback that builds resilience. It is the failure of the self—the moment when one's actions betray one's values. Václav Havel, the Czech statesman and playwright, documented this in his prison writings. He describes a moment of 'honorable cleverness' that turned into a lifelong stigma. In a desperate attempt to be tactical and secure his release from detention, he wrote a request to the authorities that, while technically true, was designed to meet them halfway. He thought he was being smart; he was actually being complicit.

The Trap of Tactical Cleverness

Havel’s error was not a lack of intelligence, but a lack of moral guard. He treated a moment of profound political significance as a routine administrative task. He believed he could navigate the situation through linguistic precision, saying nothing he didn't believe while still providing the authorities with the weapon they needed to discredit him. This is the danger of the 'tactical' mindset: it often ignores the context of who is speaking, to whom, and why. In the end, the cleverness he felt in the moment became a source of intense inner suffering.

The haunting price of self-knowledge is that you always know exactly what your own moral failures would cost you.

For those in positions of leadership, this serves as a warning against the seductive nature of the 'middle ground.' In business and politics, there is a constant pressure to be pragmatic, to find the compromise that allows for progress. But there is a line where pragmatism becomes a surrender of principle. Once that line is crossed, the damage is often irreversible, not because the act itself was large, but because it signaled a fundamental shift in character.

Lessons from Havel's Reckoning

The Danger of Nuance: Using precision to mask a lack of conviction.
The Illusion of Control: Believing one can manipulate a system without being changed by it.
The Cost of Complicity: Understanding that minor concessions can have major consequences.

Accepting failure is the only way to reclaim integrity. Havel eventually returned to prison, attempting to pay the debt of his earlier mistake with years of hardship. He understood that true redemption is not about erasing the past, but about living in a way that acknowledges it. In a world obsessed with 'failing forward' and 'pivoting,' Havel reminds us that some failures are heavy, and they must be carried.

Key Takeaway

Integrity is not found in the absence of mistakes, but in the refusal to use cleverness to escape them.

06 — Simon Willison

The Microsoft Model: Data, Lies, and LLMs

Unpacking the reality of 'clean' training data

By Simon Willison · 7 min read

Editor's note: A critical look at the marketing claims surrounding the licensing of AI training data.

Microsoft recently announced its new MAI models, claiming a breakthrough in how they are trained. Specifically, they asserted that models like MAI-Thinking-1 and MAI-Code-1-Flash were built from the ground up using 'enterprise-grade, clean, and commercially licensed data,' without the problematic distillation from third-party models that plagues the industry. For a market increasingly wary of the legal and ethical implications of AI training, this was a significant claim. It promised a new standard: models that are not just powerful, but legally and ethically 'clean'.

The Reality of the Web Crawl

However, a closer look at the technical documentation reveals a different story. Despite the marketing emphasis on 'clean' data, the underlying reality is that these models are still trained on massive crawls of the public web. Microsoft's process involves filtering a trillion-page corpus down to a few hundred billion, removing adult content and AI-generated spam, but the foundation remains the same: a proprietary and Common Crawl of the internet. The 'cleanliness' refers to the removal of noise and low-quality content, not necessarily to the resolution of the massive copyright questions that define the current AI era.

The distinction between 'clean' data and 'licensed' data is where the industry's marketing meets its legal reality.

This discrepancy highlights a growing trend in AI development: the rebranding of standard practices. By using terms like 'enterprise-grade' and 'appropriately licensed,' companies can signal safety to corporate buyers while still relying on the same vast, unvetted datasets that everyone else uses. The technical paper confirms that the majority of the corpus is still derived from web HTML, which brings us back to the same fundamental question: does 'publicly available' mean 'legally usable' for training a commercial model?

The Disconnect in AI Marketing

Marketing vs. Documentation: The gap between press releases and technical papers.
The Definition of 'Clean': Shifting from 'legally sourced' to 'high-quality/low-noise'.
The Persistence of the Web: The continued reliance on massive, unvetted internet crawls.

For the business leader, this is a lesson in due diligence. When a vendor claims their AI is 'safe' or 'ethically sourced,' do not take the adjective at face value. Look for the technical specifications. The industry is currently in a phase of linguistic gymnastics, attempting to build a sense of security around technology that is still navigating a legal minefield. The models may be getting better, but the questions surrounding their origin remain unanswered.

Key Takeaway

In the AI industry, 'clean data' often describes the quality of the content, not the legality of its acquisition.