The Deep Feed

01 — Not Boring

The Silicon Brain and the Physical Body

How robotics is moving from tabletop tricks to real-world agency

By Packy McCormick · 12 min read

Editor's note: As AI moves from screens to limbs, the stakes of intelligence shift from text to physics.

For years, the conversation around artificial intelligence has been trapped behind glass. We interact with LLMs through text boxes and chat interfaces, treating intelligence as a linguistic phenomenon. But a shift is occurring. The frontier is moving from the digital ether into the physical world. We are witnessing the birth of physical AI—the marriage of advanced reasoning models with robotic hardware. This isn't just about better chatbots; it is about machines that can walk, grasp, and navigate the messy, unscripted chaos of a real kitchen or a warehouse floor.

The Democratisation of Control

Enigma is taking a strange, almost chaotic approach to this transition. Instead of keeping their hardware locked in a lab, they have opened up over 100 AI-powered robot arms to the public. Anyone with a browser can now command a physical machine in Israel or California to paint a picture or perform an experiment. This is a massive data play. By letting humans interact with robots through text and audio, Enigma is gathering the kind of messy, non-linear training data that a controlled laboratory environment could never replicate. They are betting that human unpredictability is the best teacher for robotic intelligence.

If the race in robotics is about accumulating real training data fast and cost-efficiently, this is a novel and orthogonal approach.

While Enigma experiments with the interface, Google DeepMind is attacking the core logic. Their Gemini Robotics 2 family represents a leap in how machines understand their own bodies. We are moving past simple 'if-this-then-that' movements toward vision-language-action models. These models allow a robot to see a cluttered room, understand a verbal command like 'tidy up the tools', and then execute the complex motor sequences required to do so. It is the difference between a remote-controlled car and a creature that understands its environment.

Key advancements in the robotics stack:

Gemini Robotics 2: Mapping vision and language directly to motor control.
Embodied Reasoning: Planning long-term tasks rather than just immediate reactions.
On-Device Adaptation: Models that learn new physical tasks in hours rather than months.

The ultimate goal is a universal model of movement. DeepMind's recent work shows a single model checkpoint controlling entirely different bodies, from humanoid walkers to different gripper configurations. If we succeed, we won't need to write new code for every new robot we build. We will simply give the new machine a 'brain' that already understands the physics of the world. The hardware becomes a commodity; the intelligence becomes the value.

Key Takeaway

The next era of AI will be defined not by how well a model speaks, but by how well it moves.

02 — Simon Willison

The Efficiency of Intelligence

DeepSeek's V4-Flash and the collapsing cost of reasoning

By Simon Willison · 6 min read

Editor's note: Intelligence is becoming a cheap commodity, and the math of value is changing.

The economics of intelligence are shifting beneath our feet. For a long time, the assumption was that higher intelligence required exponentially higher costs—more parameters, more compute, more electricity. But the release of DeepSeek-V4-Flash suggests a different trajectory. At 304 billion parameters, it is punching significantly above its weight class, outperforming much larger models in specific benchmarks while maintaining a price point that feels almost negligible.

The Value-per-Intelligence Metric

When we look at the cost of input and output, the numbers are startling. At $0.14 per million input tokens, we are entering a regime where the cost of 'thinking' is no longer a primary constraint for developers. This allows for a new kind of software architecture: one where an agent can perform thousands of reasoning steps to solve a single problem because the cost of those steps is trivial. We are moving from 'expensive, careful thought' to 'cheap, iterative reasoning'.

This may currently be the best value-per-intelligence model out there.

However, there is a catch. The quality of this cheap intelligence is highly dependent on how you steer it. Observations show that using default reasoning levels can lead to underwhelming results, but bumping the reasoning effort to 'high' unlocks a much more capable agent. This suggests that the 'intelligence' isn't just in the weights, but in the computational budget we are willing to allocate to a single prompt.

Why low-cost high-reasoning models matter:

Enables massive-scale agentic workflows.
Reduces the barrier to entry for complex AI applications.
Forces a shift in focus from 'model size' to 'reasoning efficiency'.

As these models become more efficient, the competitive advantage of big tech companies based on sheer scale may erode. If a smaller, more efficient model can provide 90% of the utility at 1% of the cost, the entire landscape of AI development will be forced to pivot from brute force to architectural elegance.

Key Takeaway

Intelligence is transitioning from a scarce resource to a utility.

03 — The Marginalian

The Geometry of the Soul

Carl Jung, Einstein, and the psychological meaning of death

By Maria Popova · 15 min read

Editor's note: A reflection on how our understanding of physics can reshape our understanding of mortality.

We tend to view death as a full stop—a punctuation mark that terminates the sentence of a life. This view is driven by our biological intuition, which perceives time as a linear progression from birth to decay. But Carl Jung, the Swiss psychiatrist, proposed something far more radical. He suggested that death is not an end, but a goal. To Jung, life is an energy process, and like all such processes, it is directed toward a specific state of rest. Death, in this light, is the fulfillment of life's aim rather than its cessation.

Spacetime and the Psyche

Jung’s insight was deeply influenced by his friendship with Albert Einstein. If Einstein had shown that space and time are not absolute, but are woven into a single, flexible fabric, then Jung applied a similar logic to the human psyche. He posited that while our conscious minds are bound by the constraints of space and time, the unconscious—the deeper layer of our being—might exist in a state of relative timelessness. In this view, death is not a disappearance, but a transmutation of consciousness into a different dimension of being.

Death is psychologically as important as birth and, like it, is an integral part of life.

This perspective challenges the rationalist's fear of the void. If we view death through the lens of physics, we see that energy is never truly lost; it only changes form. Jung argued that our terror of death is often a projection of our own psychological complexes—our fears and hopes lensed through a limited understanding of time. By recognizing the potential for the psyche to transcend the linear, we move from a state of fear to a state of integration.

Jung's view of death vs. the traditional view:

Traditional: Death is a termination; a period at the end of a sentence.
Jungian: Death is a goal; the completion of a biological and psychological process.
Physical: Death is a transmutation; a shift in the state of energy and consciousness.

To live with this understanding is to embrace the full scope of existence. It requires the courage to admit that our current perceptions of reality are incomplete. If death is the fulfillment of life, then the quality of our living is directly tied to how well we prepare for that eventual transition. We are not just walking toward an end; we are moving toward a completion.

Key Takeaway

Death may not be the end of the line, but the curvature of the journey.

04 — The Marginalian

The Loss of Wonder

Darwin's regret and the cost of scientific clarity

By Maria Popova · 14 min read

Editor's note: A study of how the very science that explains the world can sometimes strip it of its magic.

Charles Darwin is remembered as the man who dismantled the divine architecture of the world. His theory of evolution provided a mechanical explanation for the complexity of life, replacing mystery with natural selection. But there was a personal cost to this revelation. In his youth, Darwin was a man of intense aesthetic sensitivity. He was moved to physical shivering by the sound of a choir and found a sense of religious devotion in the grandeur of the rainforest. He saw the world through a lens of wonder.

The Dimming of the Light

As Darwin aged, a strange and tragic phenomenon occurred. The very scientific clarity he had worked so hard to achieve began to act as a veil. The 'higher feelings' of wonder and admiration that had once filled him began to dim. He remained intellectually brilliant, but he became, in a sense, emotionally deaf to the beauty of the natural world. The mechanisms he had uncovered—the processes of descent and adaptation—seemed to have replaced the ecstatic experience of simply being alive.

Darwin found himself mentally alert and active, but blind, deaf, dead to the life of feeling with which beauty inspirits us.

This was Darwin's greatest regret. He had explained the 'how' of life so thoroughly that he struggled to maintain the 'why' of its beauty. He had traded the sublime for the structural. His life serves as a warning for the modern age: that a purely reductive understanding of the world, while necessary for progress, can lead to a profound sense of alienation from the very subject we are studying.

The tension in Darwin's legacy:

The Scientific Achievement: Providing a rigorous, evidence-based framework for life.
The Aesthetic Loss: The erosion of the intuitive, emotional connection to nature.
The Human Paradox: The struggle to balance objective truth with subjective meaning.

Darwin's struggle suggests that science and wonder are not mutually exclusive, but they are in constant tension. The challenge for the modern thinker is to integrate the two—to use the tools of science to understand the world without letting them extinguish the capacity to be moved by it. We must learn to see the mechanism and the miracle simultaneously.

Key Takeaway

Understanding the mechanism of life should not come at the expense of feeling its magic.

05 — Simon Willison

Big Mathematics

OpenAI, Anthropic, and the new era of automated proof

By Simon Willison · 8 min read

Editor's note: AI is moving from writing poetry to solving mathematical problems that have stood for decades.

Mathematics has long been considered the final fortress of human intelligence—a domain of pure logic that was thought to be largely immune to the 'hallucinations' of large language models. That fortress is under siege. OpenAI has recently demonstrated that its next-generation models can tackle mathematical problems that have seen no progress for at least a decade. This is not just about solving arithmetic; it is about formalizing proofs and navigating the complex, abstract landscapes of higher mathematics.

The Rise of Big Mathematics

The mathematician Terence Tao has described this shift as the transition toward 'big mathematics.' In this new paradigm, the role of the human mathematician changes. Instead of spending years on the technical grunt work of verifying steps or performing tedious calculations, humans will act as architects. They will define the creative direction and the high-level conjectures, while AI handles the heavy lifting of formal verification and proof construction. It is a collaborative, decentralized approach to discovery.

AI is the catalyst for a fundamental shift in the discipline—a transition toward 'big mathematics'.

OpenAI's approach involves using models to generate Lean 4 formalizations of mathematical results. This provides a level of transparency that is essential for scientific progress. By reconstructing how a proof came together based on reasoning traces, the AI isn't just giving an answer; it is showing its work. This allows human mathematicians to audit the machine's logic, turning the AI from a 'black box' into a rigorous research partner.

What 'Big Mathematics' looks like:

Human-AI collaboration: Humans provide intuition; AI provides verification.
Scale: Solving problems that are too computationally or logically dense for a single human.
Formalization: Moving from informal, prose-based proofs to machine-verifiable code.

This shift is reminiscent of the impact of the computer on other sciences. Just as the calculator didn't end mathematics but expanded its reach, AI will not replace mathematicians. It will simply allow them to operate at a scale and complexity that was previously unimaginable. The question is no longer whether AI can 'do' math, but how much more math we can do once it does.

Key Takeaway

AI is not replacing the mathematician; it is expanding the scale of what can be thought.

06 — Simon Willison

The Stateless Protocol

Why MCP 2.0 is the backbone of the agentic future

By Simon Willison · 10 min read

Editor's note: The plumbing of the AI agent era is getting a much-needed upgrade.

In the world of software, the most important breakthroughs are often the ones you never see. They are the protocols, the standards, and the plumbing that allow different systems to talk to each other without friction. The latest evolution of the Model Context Protocol (MCP) is exactly that. With the rollout of the stateless specification, the way AI agents interact with external tools has become significantly more efficient, scalable, and secure.

From Stateful to Stateless

To understand why this matters, you have to understand the friction of the old way. The 'legacy' MCP was stateful, meaning it required a complex dance of multiple requests to initialize a session and maintain it. You had to keep track of session IDs and ensure that subsequent requests were routed to the same backend machine. This is a nightmare for scaling web applications. The new stateless MCP, however, allows for everything to happen in a single, clean HTTP request. It is simpler, faster, and much easier to implement.

The new stateless way is so much cleaner from both a client- and server-side implementation perspective.

This shift also addresses a major security concern. In the past, giving an agent a shell environment with internet access was incredibly risky. It was difficult to audit and control. MCP tools, being stateless and structured, are much easier to monitor. You can see exactly what tool is being called, what arguments are being passed, and what the output is. This makes it possible to build agents that are both powerful and safe enough to run in production environments.

Benefits of Stateless MCP:

Scalability: No need to maintain server-side session state.
Simplicity: Reduces the number of requests required for tool execution.
Auditability: Easier to monitor and control agent actions for security.

As we move toward a world populated by autonomous agents, the protocols they use will determine the stability of the entire ecosystem. A fragmented, stateful, and insecure protocol would lead to a chaotic and fragile web of agents. A clean, stateless standard like MCP 2.0 provides the foundation for a more robust and interoperable future. It is the invisible architecture that will allow the agentic revolution to actually scale.

Key Takeaway

The success of AI agents depends on the invisible standards that govern their interactions.