The Deep Feed

The capital markets are making a decisive bet on the idea that disease is a solvable engineering problem. This week, Isomorphic Labs secured a $2.1 billion Series B, a figure that signals more than just investor confidence; it marks the arrival of a new era in biotechnology. Led by Demis Hassabis, the company is moving beyond the predictive successes of AlphaFold to a more aggressive mission: the design of entirely new molecules. This is not merely about predicting how proteins fold, but about understanding the chemical logic required to intervene in biological processes. The scale of this investment, involving sovereign wealth funds and major tech players, suggests that the industry no longer views AI as a helpful assistant to biologists, but as the primary engine of discovery itself.

The Protein Problem Solved

The technical validation for this massive capital injection is already appearing in the data. Isomorphic’s unified AI drug design engine, IsoDDE, has demonstrated an ability to outperform existing benchmarks in protein-ligand interactions. In one instance, the system identified a second binding pocket for the protein cereblon using only sequence data—a task that took human researchers fifteen years of experimental work to confirm. This speed is the real disruption. When a machine can compress a decade of laboratory trial and error into a single computational cycle, the entire economic model of pharmaceutical research changes. We are moving from a model of serendipitous discovery to one of intentional design.

Solving all disease is no longer a slogan; it is a capital allocation strategy.

However, the true test of these models will not happen in a data centre, but in human clinical trials. The first wholly-owned drug candidates from Isomorphic are expected to enter human testing by the end of the year. This is the moment of truth for the 'AI-first' biotech thesis. If these molecules perform in the messy, unpredictable environment of the human body as they do in the clean simulations of a GPU, the implications for human longevity and healthcare economics are massive. The goal is to move from treating symptoms to engineering cures.

Manufacturing in Microgravity

While Isomorphic works on the design side, companies like Varda are tackling the manufacturing side by looking upward. The premise is simple but physically transformative: the absence of gravity allows for chemical processes that are impossible on Earth. In a microgravity environment, molecular crystals grow with a purity and structure that terrestrial labs cannot replicate. Varda’s recent partnership with United Therapeutics to explore drug manufacturing for chronic lung diseases marks a transition from experimental spaceflight to industrial application. We are seeing the beginning of a supply chain that extends beyond the atmosphere.

The synergy between these two fields—AI-driven design and space-based manufacturing—creates a powerful loop. AI can design the perfect molecule, and space can provide the perfect environment to build it. This represents a shift in how we view the limits of human biology. We are no longer restricted by what we can find in nature or what we can cook in a terrestrial lab; we are limited only by our ability to compute and our ability to reach orbit.

The success of these ventures will require more than just clever code. It requires a massive coordination of physical infrastructure, regulatory frameworks, and biological expertise. But the momentum is undeniable. The focus is shifting from the digital abstraction of life to the physical mastery of it.

For years, the dominant theory of digital competition was Aggregation Theory: the idea that platforms win by capturing user attention and aggregating supply. In the software-only era, the barriers to entry were relatively low, and the primary moat was the network effect. However, the rise of generative AI has introduced a new, physical constraint into this equation: the compute shortage. We are entering a period where the ability to aggregate is no longer just about code and users, but about access to massive clusters of high-end silicon. The digital economy is being re-anchored in the physical reality of hardware scarcity.

The Hardware Bottleneck

When compute becomes a scarce resource, the economics of software change. In a world of infinite compute, you can afford to be inefficient; you can run massive, unoptimised models to achieve better results. In a world of scarcity, efficiency becomes the primary competitive advantage. This creates a divide between the companies that own the hardware and the companies that merely rent it. The traditional software moat—the ability to build a superior user experience—is being challenged by a new moat: the ability to secure the training and inference capacity required to power that experience.

The winner is no longer the one with the best algorithm, but the one with the most silicon.

This scarcity also impacts how companies approach Aggregation Theory. If the cost of intelligence is high due to compute constraints, then the winners will be those who can most effectively capture the value generated by that intelligence. We may see a shift where the 'aggregators' are not just the platforms that hold the users, but the infrastructure providers that hold the compute. The power is moving up the stack, from the application layer to the physical layer.

The New Economic Moats

This shift suggests that the next generation of tech giants will look very different from the last. The software-only giants of the 2010s relied on marginal costs that approached zero. The AI giants of the 2020s will deal with marginal costs that are tied directly to electricity and silicon. This brings the tech industry closer to the industrial economy, where scale is achieved through massive capital expenditure and physical assets. The ability to manage these physical constraints will be as important as the ability to write elegant code.

Ultimately, the compute shortage is a reminder that even the most advanced digital intelligence is tethered to the material world. The struggle for dominance in the AI era will be fought on two fronts: the intellectual battle for better models and the industrial battle for the hardware that runs them.

The current obsession with Large Language Models (LLMs) often misses a fundamental distinction in how intelligence can be structured. While LLMs excel at predicting the next token in a sequence, they often struggle with deep, multi-step reasoning. To understand why, we have to look back at AlphaGo. AlphaGo was not just a pattern matcher; it was a system that combined deep learning with Monte Carlo Tree Search (MCTS). This combination allowed the system to not just react to the board, but to actively search through a tree of possible futures to find the most optimal path. It was a marriage of intuition and deliberate calculation.

The Mechanics of Search

In AlphaGo, the neural network provided the 'intuition'—it could look at a position and quickly estimate which moves were likely to be good. But the MCTS provided the 'reasoning'. It would simulate thousands of potential continuations of a move, effectively looking ahead to see if a seemingly good move led to a disaster ten steps later. This process of search allows a system to sidestep the 'credit assignment problem' that plagues many other forms of reinforcement learning. Instead of guessing which part of a long sequence of actions was correct, the search process provides a clear, mathematically grounded target for every single move.

Intelligence is not just learning; it is the ability to look ahead.

This is the primary gap in current LLM architectures. When an LLM generates text, it is essentially performing a 'policy gradient'—it is choosing the most likely next step based on its training. If it makes a mistake early in a reasoning chain, it has no internal mechanism to 'search' for a better alternative; it simply continues down the path of its own error. It lacks the ability to simulate the consequences of its words before it speaks them.

The Primitives of Intelligence

The future of AI research likely lies in integrating these search capabilities directly into the generative process. If we can move from 'answer inference'—where the model simply predicts the next word—to 'search-based inference'—where the model explores multiple reasoning paths before committing to an answer—we will see a leap in capability. This would move AI from being a sophisticated mimic to being a genuine problem-solver. We are looking for the moment when the model can 'think' before it 'speaks'.

The transition from pattern matching to systematic search is the difference between a student who has memorised a textbook and a mathematician who can derive a new proof. The former is impressive, but the latter is transformative. The goal is to build systems that do not just know what to say, but know why they are saying it.

We live in an era defined by the pursuit of optimization. Every app, every workflow, and every educational curriculum is designed to maximise some metric of efficiency or utility. We have become experts at asking how things can work better, but we have forgotten how to ask why they should exist at all. This relentless drive to pragmatise everything—to reduce human curiosity to a tool for economic gain—is a modern pathology. It treats knowledge not as a way to expand the soul, but as an ingredient in technical skill. It treats human beings not as ends in themselves, but as data points in a log of user statistics.

The Russell Prescription

A century ago, Bertrand Russell identified this trend. He observed that the modern world was increasingly obsessed with 'useful' knowledge—that which contributes directly to the economic life of the community. While he acknowledged the benefits of this technical progress, he warned of the psychological cost. When conscious activity is entirely concentrated on a single, productive purpose, the result is a loss of balance and a creeping sense of nervous disorder. We are losing the capacity for 'play'—the ability to engage in activities that have no purpose beyond present enjoyment.

A life lived only for utility is a life lived in a cage.

Russell argued for the necessity of 'useless' knowledge. He pointed to the arts and the pure sciences—fields that probe the workings of the universe or the heart without the immediate expectation of a profit. This knowledge is 'useless' in the most practical sense, but it is essential for maintaining a broad and humane outlook. It provides the counteracting force to the narrowness of professional competence. Without it, we become highly efficient machines, but we lose the very thing that makes us capable of empathy and self-compassion.

The Defense of the Irreducible

The defense of the 'useless' is a defense of our humanity. There is no economic value in admiring a sunset or studying the anatomy of a scallop, yet these acts mediate our capacity for despair and our capacity for war. By engaging with things that do not serve our productivity, we reclaim our autonomy from the cult of utility. We remind ourselves that we are more than just workers, consumers, or data points. We are creatures capable of wonder, and wonder is inherently inefficient.

In a world that demands we always be 'on' and always be 'improving', the most radical act may be to engage in something entirely, beautifully, and unapologetically useless.

The current interaction model with AI is based on 'answer inference'. A human asks a question, and the model provides an answer. This is a human-in-the-loop system where speed is important for user satisfaction, but the human remains the primary driver of intent. We are now approaching a shift toward 'agentic inference'. In this model, the AI is not just answering a question; it is executing a multi-step task. The human provides a high-level goal, and the agent navigates the complexities of the real world to achieve it. This shift moves the AI from a tool to a collaborator, and eventually, to an autonomous actor.

The Inference Divide

This distinction between answer and agentic inference has massive implications for the future of compute infrastructure. In the answer-based model, the bottleneck is latency—how fast can the model respond to the user? In the agentic model, the bottleneck shifts. When an agent is working autonomously, it might spend hours or even days iterating on a task. Speed is less critical than reliability, reasoning depth, and the ability to handle long-running processes. This will lead to a divergence in hardware requirements and architectural trade-offs.

The future of compute belongs to the agents, not the users.

The Infrastructure of Autonomy

As agents become the primary consumers of intelligence, the market for compute will move away from consumer-facing latency and toward massive, background-running workloads. This is why we are seeing interest in unconventional compute locations, such as space-based data centres. If an agent is performing a task that doesn't require millisecond responses, the cost and environmental constraints of terrestrial data centres become more pressing. The infrastructure of the future will be designed to support long-running, autonomous reasoning rather than just quick-fire chat.

This shift also changes the competitive landscape for companies like Nvidia. While they currently dominate the training and 'answer' inference markets, the rise of agentic workflows may favour different types of architectures—perhaps those that are more efficient at long-context reasoning rather than just raw throughput. The 'agentic era' will be defined by machines talking to machines, creating a new layer of digital economy that operates largely out of human sight.

We are moving from an era of digital tools to an era of digital employees. The infrastructure that supports them will be the most valuable real estate in the global economy.