The Deep Feed

01 — Not Boring

The Silicon Body: Robotics Beyond the Tabletop

How foundation models are finally giving machines a sense of movement

By Packy McCormick · 8 min read

Editor's note: Robotics is moving from scripted movements to genuine embodied reasoning.

For years, the promise of robotics has been trapped in the laboratory or limited to the repetitive, predictable motions of a factory arm. We have seen machines that can sort bolts or weld car doors with terrifying precision, yet they remain brittle. If you move a nut two centimetres to the left, the system fails. The problem has never been the hardware; it has been the brain. We have lacked a way to translate the messy, unpredictable visual data of the real world into the precise motor commands required to interact with it. That is changing as foundation models move from the screen into the physical body.

The Democratisation of Control

A startup called Enigma is attempting to bypass the traditional robotics PhD route by opening up real machines to the public. By allowing anyone with a browser to control robot arms in facilities across Israel and California, they are collecting a specific kind of data: how humans actually want to interact with machines. This isn't just a marketing stunt; it is a massive data acquisition strategy. They are looking for the edge cases, the weird commands, and the intuitive gestures that a controlled lab environment would never produce. If the goal is to build a robot that can function in a human home, you need to understand human chaos.

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

While Enigma focuses on the interface, Google DeepMind is attacking the core logic. Their Gemini Robotics 2 family represents a shift toward 'embodied reasoning'. Instead of just mapping a picture to a movement, these models allow for long-term planning and multi-robot collaboration. We are seeing models that can control humanoid robots like Apptronik’s Apollo, enabling them to walk, crouch, and perform delicate tasks like tying a knot or sealing a bag. This is the transition from 'doing' to 'thinking while doing'.

The Three Pillars of Modern Robotics

Foundation Models: Generalised brains that understand physical space.
Embodied Reasoning: The ability to plan multi-step tasks in messy environments.
Hardware Agnosticism: Software that can control different types of limbs and grippers.

The hurdles remain significant. Speed and fine-motor dexterity, particularly with multi-fingered hands, are still lagging behind the fluid grace of biological organisms. However, the architecture is settling. We are moving toward a world where a single model family handles perception, planning, and safety across various physical forms. The machine is finally getting a brain that matches its body.

Key Takeaway

The bottleneck in robotics is no longer the motor, but the model's ability to reason through physical space.

02 — The Marginalian

The Cruel Contract of Aliveness

Jane Kenyon and the necessity of letting go

By Maria Popova · 6 min read

Editor's note: A meditation on the inevitability of loss and the beauty of the present.

Life operates on a fundamental tension between holding on and letting go. We spend much of our existence trying to build structures—relationships, careers, legacies—that will endure, yet the very nature of being alive requires us to accept that everything we touch is transient. To love is to accept the eventual loss of the beloved. To build is to accept the eventual decay of the structure. This is not a failure of design, but the core condition of existence.

The Miraculousness of the Ordinary

The poet Jane Kenyon, writing shortly before her death from leukemia, captured this reality not through grand lamentations, but through the small, sharp details of a life lived. In her poem 'Things', she observes the pebble flung by a hen or the hole a mouse chews in a quilt. These are moments of simple existence: things lasting, then failing to last. There is a quiet dignity in the way she describes the light passing between objects. She does not fight the inevitable; she observes it with a clear eye.

Into light all things must fall, glad at last to have fallen.

Her poem 'Otherwise' serves as a counterpoint to the anxiety of 'what if'. We often haunt ourselves with the versions of our lives that didn't happen—the different career, the different city, the different partner. Kenyon acknowledges these possibilities but grounds herself in the reality of the cereal she ate, the dog she walked, and the work she loved. By acknowledging that it *could* have been otherwise, she makes the current moment more solid, not less.

Lessons in Impermanence

Acceptance of loss as the price of connection.
Finding weight in small, sensory details.
The power of acknowledging alternative realities without being consumed by them.

To live well is to sign this contract willingly. It is to press ourselves against the world, knowing the world will eventually move on. There is a specific kind of luck in having existed at all, in having been the vessel through which light passed for a brief, bright moment.

Key Takeaway

The value of a moment is not found in its permanence, but in the fact that it happened at all.

03 — The Marginalian

The Architecture of Social Friction

Eric Berne and the psychological games we play

By Maria Popova · 9 min read

Editor's note: Understanding the hidden ego states that drive our most frustrating interactions.

Most human conflict does not stem from a simple disagreement over facts, but from a profound confusion about what we actually want. We often approach social interactions with a clumsy, unconscious set of motives, fumbling through relationships like a child with a complex tool. In the late 1950s, psychiatrist Eric Berne identified this clumsiness as 'games'—patterned, self-defeating sequences of behaviour that people use to avoid the vulnerability of honest communication.

The Three Ego States

Berne’s model of Transactional Analysis posits that every person operates from one of three ego states: the Child, the Parent, and the Adult. The Child is the source of spontaneity, creativity, and raw emotion. The Parent is the collection of behaviours and attitudes we unconsciously mimicked from our own caregivers. The Adult is the rational, decision-making part of the self that processes reality as it is. We move between these states constantly, but trouble arises when our transactions with others become mismatched.

A 'game' is an ongoing series of complementary ulterior transactions progressing to a well-defined, predictable outcome.

A 'game' occurs when a person issues a stimulus from one ego state—say, the Parent—while concealing an emotional need from another, like the Child. When the other person responds, they often respond to the surface message, leading to a mismatch. This mismatch creates a predictable cycle of frustration, anger, or guilt. We play these games not because we want to be unhappy, but because they provide a predictable, albeit toxic, form of social 'strokes'—the recognition and affirmation we crave to feel psychologically alive.

Identifying the Ego States

Parent: Judging, nurturing, or mimicking authority.
Adult: Objective, factual, and problem-solving.
Child: Emotional, creative, or rebellious.

The goal of understanding these games is not to become a master manipulator, but to achieve clarity. By recognising when we are slipping into a Parent or Child state, we can consciously return to the Adult state. This allows for 'complementary transactions'—interactions where two people engage on the same level, enabling true intimacy and effective communication instead of repetitive, exhausting drama.

Key Takeaway

Social friction is often a symptom of hidden motives being played out through mismatched ego states.

04 — Simon Willison

The Open Weight Rebellion

The shifting power dynamics of AI leadership

By Simon Willison · 5 min read

Editor's note: The battle between proprietary closed models and the open-weight movement.

The AI industry is currently locked in a struggle over the definition of 'leadership'. On one side are the proprietary giants, building massive, closed-box models that offer high performance but zero transparency. On the other is the open-weight movement, which argues that for AI to be safe, useful, and truly innovative, the underlying weights of these models must be accessible to the public. This isn't just a technical debate; it is a political one about who controls the most powerful technology in human history.

The Performance Gap Closes

Recent developments, such as the Kimi K3 model, have demonstrated that open-weight models can now stand toe-to-toe with the most expensive proprietary frontier models. This shatters the argument that openness necessitates inferiority. When the performance gap narrows, the economic argument for closed models weakens. Why pay a premium for a black box when an open model offers comparable intelligence and complete control over your data and implementation?

Open weights are the difference between being a tenant of AI and being an owner of it.

This tension has led to public letters and intense lobbying. Major players are attempting to frame open weights as a threat to American AI leadership, suggesting that uncontrolled access to powerful models is a security risk. However, proponents argue that the real risk lies in the concentration of power. If only three or four companies control the 'brains' of the future, the rest of the world becomes economically and technologically subservient to their whims.

The Open vs. Closed Debate

Closed Models: High convenience, high cost, zero transparency.
Open Weights: High control, lower cost, rapid community innovation.
The Security Argument: Centralised control vs. distributed resilience.

As we move forward, the 'open weight revolution' will likely be defined by how quickly the community can optimise these models for specific, local tasks. The winners won't just be those with the largest clusters, but those who build the most robust ecosystems around accessible intelligence.

Key Takeaway

The accessibility of model weights determines whether AI becomes a utility or a monopoly.

05 — Simon Willison

The Efficiency Frontier

How AI is rewriting its own code

By Simon Willison · 5 min read

Editor's note: OpenAI's latest price drop is a result of AI-driven hardware optimisation.

In the race for AI dominance, the most important battle isn't just about intelligence; it's about the cost of inference. Every time a model generates a token, it costs electricity, GPU time, and memory bandwidth. For a company to scale, it must move down the cost curve. OpenAI's recent massive price drop for its GPT-5.6 Luna model is not just a marketing move—it is a demonstration of a new kind of recursive improvement: using AI to make AI cheaper.

The Sol Model and Kernel Optimisation

The driver behind this efficiency is GPT-5.6 Sol, a specialised model used to optimise the production environment. Sol doesn't just manage traffic; it actually rewrites the core mathematical operations—the kernels—that run on the GPU. By identifying inefficient data layouts and finding ways to parallelise tasks that were previously sequential, Sol has autonomously optimised the model's forward pass. This is a level of self-optimisation that was previously the domain of highly specialised human engineers.

AI is no longer just the passenger; it is now the mechanic optimizing the engine while it runs.

The result is a dramatic shift in the economics of intelligence. Luna's input costs have dropped to a level that undercuts almost every major competitor, including Google's Gemini Flash-Lite and Anthropic's Claude Haiku. When the cost of intelligence drops by 80%, the entire landscape of what is possible changes. Developers can now build agents that run continuously, performing thousands of small tasks without the fear of a catastrophic bill at the end of the month.

The Impact of Cheap Inference

Agentic Workflows: AI can now 'think' for longer and more frequently.
Edge Computing: Lower costs make it viable to run complex models on smaller devices.
Market Disruption: Low-cost models force competitors to innovate on efficiency, not just size.

We are entering an era where the bottleneck is no longer the cost of the thought, but the quality of the prompt. As intelligence becomes a commodity, the value shifts from the ability to compute to the ability to direct that computation effectively.

Key Takeaway

The next frontier of AI growth is not just larger models, but the recursive optimisation of their own efficiency.

06 — Simon Willison

Measuring the Ghost

The difficulty of evaluating intelligence

By Simon Willison · 6 min read

Editor's note: Why standard benchmarks are failing and how small, custom evals provide the real truth.

How do you know if a model is actually getting better, or if it has simply memorised the test? This is the central crisis of AI evaluation. Standard benchmarks are increasingly becoming useless; they are static targets that models are inadvertently trained to hit. To truly understand a model's capability, we need something more granular, more specific, and more repeatable. We need 'evals' that test for reasoning rather than just retrieval.

The Rise of Small, Targeted Evals

The new approach, embodied by tools like 'smevals', is to move away from massive, monolithic tests and toward collections of small, highly specific challenges. Instead of asking a model to 'solve math', you ask it to 'generate an SVG of a pelican riding a bicycle'. These tasks are designed to test specific dimensions of capability: spatial reasoning, coding accuracy, or adherence to complex formatting rules. By running these small suites across different model configurations, we can see exactly where a model succeeds and where it breaks.

An eval is a collection of challenges designed to answer a specific question about a model's capability.

The process involves three distinct stages: the run, the grader, and the report. A 'run' executes a task against a specific model configuration. A 'grader' then evaluates the output, using either simple string checks or more complex methods—such as using another, more powerful model to judge the quality of the first. Finally, the results are compiled into a report that allows for direct comparison. This creates a rigorous, scientific method for testing what was previously a matter of anecdotal observation.

Components of a Modern Eval Suite

Tasks: The specific challenges or prompts.
Configs: The models and system prompts being tested.
Graders: The mechanism for determining success or failure.
Checkers: The logic used to validate the output.

As the industry moves toward autonomous agents, these custom evals will become the only way to ensure safety and reliability. We cannot rely on general intelligence scores when we are deploying models to manage our calendars, our code, or our finances. We need to know exactly how they handle the edge cases, and we need the data to prove it.

Key Takeaway

Standard benchmarks are becoming obsolete; the future of AI development lies in hyper-specific, task-oriented evaluation.