The Deep Feed

01 — Not Boring

The Cowboy Space Corporation: Branding the High Frontier

How Robinhood money is turning orbital mechanics into a sci-fi marketing exercise

By Packy McCormick · 8 min read

Editor's note: A study in how modern billionaires use narrative to sell hardware that sounds like science fiction.

When a company announces it has raised $200 million to build rockets whose upper stages are essentially foldable data centres, the rational response is skepticism. The space launch market is already a battlefield dominated by established players and well-funded newcomers. Yet, Cowboy Space Corporation—a rebrand of Baiju Bhatt’s Aetherflux—is not playing by the standard rules of aerospace competition. Instead of selling technical specs or launch cadence, they are selling a mythos. It is the kind of move that turns venture capital into a tool for world-building rather than just product development.

The Aesthetic of the Moonshot

The marketing for Cowboy Space Corporation is intentionally jarring. Their videos feature tumbleweeds, cowboy-adjacent soundtracks, and a founder slapping hats on his team. This isn't a mistake; it is a deliberate choice to distance the company from the sterile, engineering-heavy tone of traditional aerospace. By calling space 'The High Frontier,' they tap into a specific brand of American frontierism that suggests the next great era of wealth and energy won't be found in a lab, but in the wild, unmapped reaches of orbit. It is a way of making the impossible feel like an inevitable adventure.

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

The actual product is as strange as the branding. The concept involves using the upper stage of a rocket as a deployable unit that unfurls into a solar-powered data centre once in orbit. This solves a fundamental problem of space-based computing: the massive energy requirements. By moving the compute to where the sun is constant and the cooling is free, they are attempting to bypass the terrestrial constraints that limit AI scaling. It is a high-stakes bet that the future of intelligence is orbital.

Why the brand works

It replaces technical jargon with a recognizable cultural archetype
It frames capital expenditure as a 'quest' rather than a business plan
It creates an emotional connection to the concept of space exploration

This approach raises a question about the nature of modern industry. When the technology becomes sufficiently complex, the barrier to entry isn't just engineering—it's the ability to sustain the collective belief required to fund it. Cowboy Space Corporation is betting that people will fund a dream more readily than they will fund a series of orbital deployment schedules. In the era of the billionaire-led moonshot, the story is often as important as the rocket.

Key Takeaway

In high-stakes industries, a compelling narrative can be as essential as the engineering itself.

02 — Dwarkesh Podcast

The Fragility of Scale: Why AI Training Fails

Causality, bias, and the invisible bugs in the machine

By Dwarkesh Patel · 12 min read

Editor's note: As we push toward larger models, the technical reasons for failure become increasingly esoteric and difficult to debug.

Training a large language model is not a smooth climb; it is a precarious operation where a single mathematical error can ruin months of work and millions of dollars in compute. The difficulty lies in the fact that as models scale, the bugs become more subtle. They are no longer simple coding errors, but emergent phenomena caused by the way data moves through massive clusters of hardware. When a training run fails, it is often due to a breakdown in causality or the introduction of compounding bias.

The Causality Trap

One of the most significant risks in modern training techniques like expert routing is the accidental breaking of causality. In an effort to balance the workload across different 'experts' within a model, researchers might use 'expert choice' methods. This ensures every expert gets an equal number of tokens, but it creates a problem: the allocation of a token to an expert might depend on information from future tokens that the model wouldn't actually have during real-world use. This creates a training environment that is fundamentally different from the deployment environment, leading to models that perform well in the lab but fail in practice.

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

Then there is the issue of numerical precision. During the training of GPT-4, developers encountered significant issues with FP16 precision. In large-scale operations where you are adding thousands of small gradients into a single accumulator, the way computers round numbers becomes a critical failure point. If the rounding is inconsistent, the calculated value can drift significantly from the true value. At scale, these tiny errors don't just stay small; they compound, creating a systemic bias that can steer the entire model toward nonsense.

Primary failure modes in pretraining

Breaking causality through improper token routing
Numerical drift caused by low-precision arithmetic
Token dropping that creates off-policy biases
Compounding bias in gradient accumulation

The implication for the future of AI is sobering. Some believe that as we scale, these issues will eventually be solved by better engineering. Others argue that we are entering a regime where new, bespoke problems will emerge at every new level of scale. There is no guarantee of a smooth path to AGI; instead, there may be a series of increasingly difficult technical hurdles that require not just more compute, but a deeper understanding of the physics of information.

Key Takeaway

Scaling AI is not just a matter of adding more hardware; it is a battle against the compounding errors of complex systems.

03 — Dwarkesh Podcast

The Intelligence-Power Fallacy

Why the smartest minds are rarely the ones in charge

By Dwarkesh Patel · 10 min read

Editor's note: A necessary distinction between the ability to solve a complex equation and the ability to command a nation.

There is a persistent mistake in how we discuss the rise of Artificial Superintelligence (ASI): we conflate intelligence with power. When people imagine a superintelligent machine, they often picture a digital tyrant—a singular, god-like mind that outmanoeuvres every human leader through sheer strategic brilliance. This intuition is flawed. It confuses the ability to process abstract concepts with the ability to exert influence over human systems.

The Mechanics of Influence

In the real world, power is rarely the result of a single person's 'galaxy brain' scheming. Instead, power is a social phenomenon. It is the product of authority, trust, and the ability to coordinate large groups of people. A dictator or a president is powerful not because their individual brain is the most efficient optimization engine on Earth, but because they sit atop institutions that millions of people accept as legitimate. Intelligence, in the sense of high-level reasoning and scientific discovery, is a different metric entirely.

The physicists are not running the world.

Consider the difference between an individual and a collective. A single genius like Richard Trevithick might die in poverty, but the collective intelligence of a nation—the ability to build steam engines, navies, and global trade networks—is what creates an empire. This is why the most significant impact of AI will likely not be a single sentient agent, but the way AI empowers existing human structures. A superintelligent coder doesn't need to be a leader; they just need to be a tool for a leader who already has the authority to command resources.

Distinguishing Intelligence from Power

Intelligence: The ability to solve complex problems and build abstract models
Power: The ability to command resources and coordinate human cooperation
The Gap: High intelligence does not require high social authority

The real geopolitical shift will not be caused by an AI 'taking over,' but by certain countries or companies gaining a massive advantage in the AI-driven economy. The winners will be those who can best integrate superintelligent tools into their existing social and economic frameworks. The threat is not a rogue mind, but a massive, asymmetric shift in the capacity of certain actors to execute their goals.

Key Takeaway

Power is a social coordination problem, while intelligence is a computational one; they are not the same.

04 — Dwarkesh Podcast

Rebuilding AlphaGo: The Primitives of Learning

What a 2017 masterpiece teaches us about the future of RL

By Dwarkesh Patel · 15 min read

Editor's note: To understand where AI is going, we must look at the elegant mechanisms that first made it feel truly intelligent.

AlphaGo remains one of the cleanest examples of how intelligence can be engineered. It wasn't just a massive neural network; it was a combination of search, learning from experience, and self-play. By using Monte Carlo Tree Search (MCTS), AlphaGo could simulate thousands of potential futures, allowing it to make decisions that were not just based on pattern recognition, but on a deep, calculated understanding of consequence. This combination is what allowed it to tackle a game that was previously thought to be computationally intractable.

The Credit Assignment Problem

One of the most significant insights from the AlphaGo era is how it sidesteps the 'credit assignment problem.' In standard Reinforcement Learning (RL) used for Large Language Models, the system has to figure out which specific token in a long sequence actually led to a successful outcome. This is incredibly inefficient. AlphaGo, however, uses MCTS to suggest a strictly better action at every single move. This provides a much clearer training target, essentially telling the model exactly what it should have done, rather than forcing it to guess based on a distant reward.

Sometimes you understand the future better by stepping backward.

The democratization of this kind of research is also accelerating. What once required a team of researchers at DeepMind and millions of dollars in compute can now be achieved by a single developer using modern LLM coding tools and a few thousand dollars of rented compute. The barrier to entry for building world-class AI agents is collapsing, shifting the focus from 'how do we build this?' to 'what should we build?'

Core components of AlphaGo's success

Deep Neural Networks for pattern recognition
Monte Carlo Tree Search for look-ahead capability
Self-play to generate high-quality training data

As we look toward the next generation of AI, the question is whether LLMs can adopt these more efficient learning structures. If we can move beyond simple next-token prediction and toward systems that can perform internal search and self-correction, we may see the next leap in reasoning capabilities. The transition from 'stochastic parrots' to 'reasoning agents' depends on our ability to implement these fundamental primitives of intelligence at scale.

Key Takeaway

True intelligence requires more than pattern matching; it requires the ability to simulate and evaluate potential futures.

05 — The Marginalian

The Art of Seeing: Escaping the Urban Trance

How expertise and attention can reclaim a lost world

By Maria Popova · 9 min read

Editor's note: A meditation on the cost of productivity and the profound beauty of intentional observation.

Modern urban life is a masterclass in adaptive ignorance. We move through our cities in a state of constant, high-speed filtering, using our attention to block out the hum of traffic, the texture of the pavement, and the lives of the people passing us by. We call this concentration, and we treat it as a virtue of the productive mind. But this efficiency comes at a steep price: we inhabit a world that is largely unlived and unremembered. We are sleepwalkers on our own streets.

The Expert Eye

The cognitive scientist Alexandra Horowitz suggests that the cure for this trance is not just 'focus,' but the cultivation of expertise. When we look at a city block through the eyes of a geologist, a dog, or an artist, the familiar becomes strange. An expert doesn't just see a street; they see a complex system of geological layers, olfactory trails, or light and shadow. Expertise acts as a lens that pulls specific, previously invisible details into our field of awareness, breaking the monotony of our conditioned perceptions.

Attention is an intentional, unapologetic discriminator.

This is not a call for mindless mindfulness, but for an active, investigative way of being. It is about acknowledging that our current perception of reality is incomplete. We are missing the vast majority of the information that surrounds us. By intentionally choosing what to attend to—by leaning into a specific way of seeing—we can transform a mundane walk into an encounter with a living, breathing entity. The city is not just a backdrop for our errands; it is a subject worthy of study.

Ways to reclaim attention

Adopt a specific lens (e.g., look for textures, sounds, or patterns)
Recognize the bias of your own current focus
Practice 'investigative' observation rather than passive walking

To see is to participate in the world. When we stop filtering everything out, we stop being mere observers of our own lives and start becoming participants in the environment. The art of seeing is, ultimately, the art of being present in a world that is constantly trying to distract us from itself.

Key Takeaway

True presence requires the courage to look past what we expect to see.

06 — Not Boring

The Biotech Frontier: Solving Disease with AI

From protein folding to space-based drug manufacturing

By Packy McCormick · 11 min read

Editor's note: The convergence of AI and biology is moving from theoretical models to tangible clinical candidates.

The promise of AI in biology has moved beyond the hype of protein folding into the realm of actual drug design. Companies like Isomorphic Labs, led by Demis Hassabis, are no longer just predicting structures; they are building unified engines designed to discover new molecules. With a recent $2.1 billion funding round, the mission to 'solve all disease' is being backed by the kind of capital required to move from digital models to human clinical trials. We are seeing the transition from AI as a research assistant to AI as a primary driver of drug discovery.

The Microgravity Advantage

While Isomorphic focuses on the computational side, companies like Varda are looking at the physical environment of manufacturing. The lack of gravity in space offers a unique laboratory for pharmaceutical development. In microgravity, molecular structures behave differently, allowing for the creation of drugs that might be impossible to manufacture on Earth. The partnership between Varda and United Therapeutics to explore space-based manufacturing for chronic diseases marks a significant step in proving that the 'space economy' can include high-value biological products.

The craziest part is that if he does it again this time, the thing that sounds crazy that he will actually pull off is solving all disease.

This dual approach—improving the design of molecules on Earth and improving the manufacturing of those molecules in orbit—represents a two-pronged attack on the limitations of modern medicine. One side addresses the intelligence gap in discovery, while the other addresses the physical constraints of production. Together, they suggest a future where the timeline from identifying a disease to delivering a cure is radically compressed.

Key shifts in biotech

From structure prediction to active molecule design
From terrestrial-only manufacturing to orbital production
From human-led discovery to AI-driven pipelines

The success of these ventures will depend on their ability to bridge the gap between digital success and biological reality. An AI can design a perfect molecule, but that molecule still has to pass the brutal, unpredictable tests of human biology. The next decade will be defined by whether these computational breakthroughs can survive the transition from the silicon chip to the human cell.

Key Takeaway

The most significant impact of AI may not be in digital intelligence, but in its ability to master the physical complexities of biology.