The Deep Feed

01 — Lenny's Newsletter

The Silicon Wall: Why AI's Next Frontier is Physical

Caitlin Kalinowski on the transition from software models to the hardware that runs them.

By Lenny Rachitsky · 12 min read

Editor's note: As the software hype cools, the real battleground is shifting to the machines themselves.

For years, the conversation around artificial intelligence has lived almost exclusively in the realm of weights, biases, and transformer architectures. We have treated intelligence as a mathematical abstraction, something that exists in the ether of cloud computing. But we are hitting a wall. The next phase of this revolution will not be won by better code alone, but by the physical objects that execute that code. Caitlin Kalinowski, a veteran of Apple, Meta, and OpenAI, argues that we are only at the start of a massive hardware boom. The transition from pure software to embodied intelligence—robotics, specialized chips, and advanced sensory hardware—is the inevitable consequence of scaling models that can now reason about the physical world.

The Ghost in the Machine

The difficulty of building hardware is fundamentally different from building software. In software, a bug is a line of code you can patch in minutes. In hardware, a bug is a multi-million pound mistake that requires a complete redesign of a physical component. Kalinowski's experience at Apple and Meta provides a sobering perspective on this. When you build something like the MacBook or the Quest, you are fighting against the laws of thermodynamics, the limits of material science, and the brutal realities of global supply chains. The current AI boom has largely bypassed these constraints by running on general-purpose GPUs, but as we move toward robotics and edge computing, those constraints will become the primary bottleneck.

The next era of intelligence is not about better algorithms; it is about the machines that can actually touch, move, and sense the world.

We are seeing a shift in how capital is being deployed. It is no longer enough to fund a lab with a thousand researchers; you now need to fund the massive, energy-hungry data centres and the specialized silicon required to make those researchers' models useful. This creates a high barrier to entry. The companies that will dominate the next decade are those that can bridge the gap between the digital logic of a large language model and the messy, unpredictable physics of a humanoid robot or a pair of smart glasses.

The Hardware Bottlenecks

Memory price shocks and supply constraints
The thermal limits of high-density compute
The difficulty of training models for physical interaction (world models)
The supply chain for specialized robotics components

The convergence of AI and hardware will likely follow the path of the smartphone. Initially, mobile computing was a crippled, niche version of desktop computing. But as hardware caught up to the software vision, it became the primary way we interact with the digital world. We are approaching a similar inflection point with AI. The software is ready to be everywhere, but it is currently trapped in data centres. The hardware boom is the process of liberating that intelligence from the server rack and putting it into the world.

This transition will be messy. It will involve massive failures in manufacturing, intense competition for rare earth minerals, and a desperate scramble for energy. But for those who can navigate the physical constraints, the rewards are total. The companies that build the 'body' for the AI 'mind' will own the infrastructure of the 21st century.

Key Takeaway

Intelligence is moving from the cloud to the physical world, and the winners will be those who master the hardware.

02 — Not Boring

Cowboy Space: The Audacity of Orbital Compute

Why turning Robinhood money into space-based data centres is the ultimate billionaire move.

By Packy McCormick · 8 min read

Editor's note: A look at how branding and extreme ambition are colliding in the new space race.

There is a specific type of madness required to launch a rocket. It is a pursuit that defies standard economic logic, where the risks are astronomical and the timelines are measured in decades. But Cowboy Space Corporation, a rebranded evolution of Baiju Bhatt’s Aetherflux, is taking this madness to a new level. They aren't just trying to reach orbit; they are trying to turn the upper stages of rockets into foldable, solar-powered data centres. It is a concept that sounds like science fiction, yet it is backed by hundreds of millions of dollars in venture capital. It is the intersection of the 'compute is king' era and the 'space is the final frontier' era.

The Branding of the High Frontier

What makes Cowboy Space fascinating isn't just the engineering; it is the theatre. Packy McCormick notes the sheer weirdness of the company's presentation: a founder slapping cowboy hats on his team while discussing orbital energy. This isn't just a quirk; it is a deliberate choice to lean into a narrative. In an era where tech companies often struggle to differentiate themselves, Cowboy Space has embraced a persona. They are calling space 'The High Frontier' and adopting a rugged, frontier-style aesthetic. It is a way of telling a story that makes the impossible feel like a quest.

Turning Robinhood money into sci-fi energy and compute moonshots is exactly how you should billionaire.

The logic, however, is not entirely insane. As AI models grow, the demand for compute is becoming a planetary-scale problem. Data centres are hitting the limits of terrestrial power grids and cooling capacities. Moving the compute to space—where solar energy is constant and the heat sink of the vacuum is readily available—is a radical solution to a very real problem. If you can build a data centre that unfurls like a solar sail in orbit, you have solved the energy bottleneck for the AI age.

The Cowboy Space Thesis

Space offers unlimited solar energy for compute
Orbital environments provide unique cooling opportunities
The 'High Frontier' narrative attracts top-tier talent and capital
Decoupling compute from terrestrial power grids

Of course, the skepticism is warranted. Competing with SpaceX in launch is a fool's errand. But Cowboy Space isn't trying to be a launch company; they are trying to be a compute company that happens to use rockets. They are betting that the value of the data processed in orbit will far exceed the cost of getting it there. It is a high-stakes gamble on the idea that the future of intelligence is not just digital, but orbital.

Whether this results in a functional constellation of orbiting servers or a very expensive collection of cowboy hats remains to be seen. But in the current venture landscape, audacity is a currency all its own.

Key Takeaway

The next great compute bottleneck may be solved not on Earth, but in orbit.

03 — Dwarkesh Podcast

The Fragility of Scale: Why AI Training Fails

A deep dive into the numerical and causal errors that break frontier models.

By Dwarkesh Patel · 15 min read

Editor's note: Training a frontier model is less like writing software and more like conducting a high-wire act.

We often talk about AI progress as a linear progression of scale: more data, more compute, more parameters. This creates an illusion of stability. In reality, training a frontier model is a precarious operation where a single numerical error can ruin months of work and millions of dollars. Dwarkesh Patel's recent investigations into pretraining failures reveal a world of 'broken causality' and 'compounding bias'—technical ghosts that haunt the largest training runs. When you are operating at the scale of thousands of GPUs, the standard rules of computer science begin to warp.

The Causality Trap

One of the most insidious issues is the breakdown of causality during 'expert routing' in Mixture-of-Experts (MoE) models. To make training efficient, engineers use routers to decide which 'expert' should handle which token. To ensure every expert gets an equal workload, they sometimes use 'expert choice' routing. This sounds efficient, but it creates a fundamental flaw: the decision of which expert handles token *n* can depend on information from token *n+k*. This breaks the causal chain. The model learns patterns during training that it simply cannot see during real-world inference. This discrepancy—training on information that won't exist in deployment—can lead to models that are fundamentally broken, regardless of how much data you feed them.

Bias is much worse than variance. Variance can average out, but bias compounds.

Then there is the problem of numerical precision. In the rush to speed up training, engineers often use lower-precision formats like FP16. While this saves memory and time, it introduces rounding errors. In a massive collective operation where you are summing millions of small gradients, these tiny errors don't just vanish; they accumulate. If your precision is off, you aren't just getting a slightly noisy result; you are getting a result that is mathematically disconnected from reality. This is not a minor bug; it is a systemic failure that can cause a model to diverge entirely.

Primary Failure Modes

Causality breaking via expert choice routing
Token dropping leading to information gaps
Numerical drift caused by low-precision arithmetic (FP16)
Compounding bias in gradient accumulation

This reality challenges the idea that we can simply 'solve' AI training by throwing more hardware at it. As models scale, new and more bespoke ways to fail emerge. It is a game of whack-a-mole played at the limits of physics and mathematics. The ability to manage these subtle, non-obvious errors is becoming the primary differentiator between the labs that produce frontier models and those that merely produce expensive, broken experiments.

The implication is clear: the future of AI development belongs to the engineers who can master the discipline of numerics and the rigour of causal consistency. It is a move away from the 'move fast and break things' ethos toward a much more traditional, high-precision engineering discipline.

Key Takeaway

Scaling AI is not just a matter of more compute; it is a battle against the mathematical errors that emerge at scale.

04 — Dwarkesh Podcast

The Power Delusion

Why we mistake strategic competence for raw intelligence.

By Dwarkesh Patel · 10 min read

Editor's note: A necessary correction to our intuitions about how intelligence and power actually interact.

There is a persistent error in how we conceptualise the rise of Artificial Superintelligence (ASI). We tend to imagine a single, god-like mind that outmanoeuvres the world through sheer strategic brilliance—a digital Machiavelli. This intuition is driven by our fascination with high-stakes games like Go or Diplomacy, where intelligence is measured by the ability to predict and manipulate an opponent. But this conflates two very different things: intelligence and power. Intelligence is the ability to solve complex problems across domains; power is the ability to command resources and people. They are not the same, and they are not always correlated.

The Authority Gap

Consider the most powerful people in human history. Their power rarely stemmed from having the highest IQ or the most efficient optimization engine in their heads. Instead, it came from their ability to command authority and build trust. A dictator is powerful because a massive bureaucracy and a population believe in their legitimacy, not because they can solve differential equations faster than a physicist. When we talk about AI, we often assume that a 'smarter' AI will automatically become a 'more powerful' AI. This ignores the social and institutional structures that actually distribute power in the real world.

Power is the product of having the authority and trust to get people to collaborate, not a galaxy-brain scheming capability.

The AI we are building today is being trained for economically valuable tasks: coding, translation, scientific reasoning. This is pure intelligence. There is no inherent reason why a model that can write perfect Python code would also have the drive or the social capability to seize control of a nation-state. The 'power-seeking' AI trope assumes that strategic manipulation is a natural byproduct of intelligence, but history suggests that many of the most brilliant minds have been entirely powerless, while many of the most powerful have been mediocre thinkers.

Intelligence vs. Power

Intelligence: Problem-solving, abstract reasoning, domain mastery
Power: Authority, trust, resource control, social coordination
The Error: Assuming strategic manipulation is a universal byproduct of intelligence
The Reality: Power is often institutional and social, not just cognitive

The more likely scenario for the AI age is not the rise of a single, omnipotent digital tyrant, but the rise of automated firms. We will see companies that use AI to outcompete everyone else in the market through sheer efficiency and coordination. This is a capitalist transformation, not a coup d'état. The threat is not a single mind outthinking us, but a thousand intelligent agents out-executing us.

By separating these two concepts, we can better prepare for the actual changes coming. We should worry less about the 'Skynet' scenario of a rogue superintelligence and more about the economic and social displacement caused by the widespread deployment of highly intelligent, non-human agents.

Key Takeaway

Intelligence is a cognitive tool; power is a social reality. Do not mistake one for the other.

05 — Dwarkesh Podcast

The Science Problem: Why RL Might Fail the Laboratory

The difficulty of automating discovery when the truth is elusive.

By Dwarkesh Patel · 14 min read

Editor's note: Reinforcement Learning (RL) works for coding and math, but science is a much messier beast.

The prevailing optimism regarding AI in science rests on a single assumption: that scientific discovery can be turned into a reinforcement learning (RL) loop. In coding or mathematics, the 'truth' is easily verifiable. A piece of code either runs or it doesn't; a mathematical proof is either correct or it is not. This creates a tight feedback loop that allows an AI to learn through trial and error. But science does not work this way. In the real world, the verification of a new theory can take decades, and even then, the 'correct' theory often looks like a mess of errors before it is unified into a coherent whole.

The Epicycle Trap

History is littered with 'progressive' theories that initially looked worse than the models they replaced. When Copernicus introduced heliocentrism, his model was actually less accurate than the existing Ptolemaic geocentric model. Ptolemy had spent centuries adding 'epicycles'—complex mathematical corrections—to make his model fit observations. Copernicus discarded these in favour of a simpler, more elegant idea, but because he didn't yet have the concept of elliptical orbits, his model was actually a step backward in terms of predictive precision. If you were an astronomer in 1543, you would have had no empirical reason to prefer Copernicus over Ptolemy.

Science is not just about finding what works; it is about finding the framework that explains why things work.

This creates a massive problem for RL. An AI agent trained to maximise 'accuracy' or 'verifiability' will naturally gravitate towards the 'Ptolemaic' approach. It will learn to add more and more complex, ad-hoc corrections to existing models to fit new data, rather than proposing a radical, paradigm-shifting new framework. It will become a master of 'epicycles'—the very thing that prevents true scientific progress. The AI will be incredibly good at fitting the data, but it will be terrible at discovering the underlying laws.

Why Science is Hard for AI

Verification loops are too slow (decades vs milliseconds)
The 'correct' theory can initially be less accurate than the old one
Risk of 'regressive' research: adding complexity instead of insight
The role of human intuition and heuristic judgment is difficult to codify

To truly automate science, we need more than just an RL loop. We need a way to reward 'progressiveness'—the ability of a theory to make unanticipated, correct predictions. We need to teach AI not just to fit the data we have, but to prepare for the data we don't yet have. This requires a level of conceptual reasoning that goes far beyond the pattern matching seen in current large language models.

Until we can bridge this gap, AI will remain a powerful tool for scientists—a way to process data and test hypotheses—but it will not be the scientist. The leap from 'verifiable computation' to 'scientific discovery' is perhaps the hardest leap in all of artificial intelligence.

Key Takeaway

Science is not a game of pattern matching; it is a game of theoretical progress, which is notoriously difficult to automate.

06 — Simon Willison

The Open Source Retreat

The NHS decision and the growing tension between security and transparency.

By Simon Willison · 5 min read

Editor's note: A warning shot from the UK's digital services regarding the move toward closed-door security.

In the digital age, security is often used as a justification for secrecy. The recent decision by the NHS to restrict access to its open-source repositories is a prime example of this tension. Triggered by vulnerabilities identified during Project Glasswing, the move was intended to protect sensitive systems. However, it has sparked a fierce debate within the UK's digital infrastructure community. The Government Digital Service (GDS) has stepped in to offer a counter-perspective: that making everything private actually increases risk by reducing the ability for external scrutiny and reuse.

The Cost of Secrecy

The argument for 'open by default' is not merely ideological; it is practical. When code is open, it can be audited by more eyes, finding vulnerabilities faster than any internal team could. When code is closed, it becomes a black box, prone to the same hidden flaws that plague proprietary software. Furthermore, closing off repositories creates massive inefficiencies. Instead of building upon existing, proven tools, different departments end up reinventing the wheel, leading to wasted taxpayer money and a fragmented digital landscape.

Openness should remain the default posture, with closure used sparingly and deliberately.

The friction between the NHS and the GDS represents a broader struggle within the civil service. It is a clash between a risk-averse culture that seeks to mitigate threats through isolation, and a modern digital culture that seeks to mitigate threats through transparency and collaboration. The GDS's intervention is a significant signal that the move toward closed-source in the public sector is being viewed by many as a strategic error.

Risks of Closing Open Source

Reduced scrutiny and slower vulnerability detection
Increased delivery and policy costs
Loss of code reuse across government departments
Creation of siloed, unmaintainable systems

As we move into an era where AI and automated systems manage more of our public services, the debate over open source becomes even more critical. The code that runs our hospitals, our transport, and our social safety nets should not be a collection of secrets. It should be a shared, transparent foundation that can be continuously improved by the entire community.

The NHS decision may have been intended as a shield, but if the GDS is right, it may end up being a weight that slows down the very digital transformation the healthcare system so desperately needs.

Key Takeaway

Secrecy is not a substitute for security; in the digital world, transparency is often the better defence.