The Deep Feed

01 — Aeon

The Optimization Trap

Why equating intelligence with efficiency is a dangerous error

By Sasha Mudd · 12 min read

Editor's note: As AI models become more efficient, we must distinguish between solving a problem and understanding why the problem matters.

In a crowded university lecture hall, a computer scientist recently offered a definition of intelligence that sent a chill through the room. He argued that intelligence is simply the ability to optimise—to take a set of inputs and refine them based on past experience to reach a goal. The audience agreed. It sounded logical. After all, large language models are built on the machinery of optimisation: loss functions, rewards, and feedback loops. They are designed to get better at a specific task by reducing error. But there is a massive gap between a machine that calculates the most efficient path to a destination and a mind that decides whether that destination is worth reaching.

Hume’s Slave

This tension is not new. It traces back to the 18th-century philosopher David Hume, who famously argued that reason is merely the 'slave of the passions'. In Hume’s view, reason has no agency of its own. It cannot tell you what to want; it can only tell you how to get it. If you want revenge, reason provides the blueprint for the weapon. If you want peace, it provides the strategy for negotiation. Reason is a tool, a calculator of means, but it is entirely subservient to the desires—the passions—that set the direction. To Hume, reason is a servant, never a master.

Reason does not command. It defers. It does not tell us what to want or what might be worth wanting.

The danger in our current technological moment is that we are building systems that are perfect servants. We are creating models that can optimise any given objective with terrifying speed. If we define intelligence solely as this capacity for optimisation, we inadvertently accept Hume’s lopsided view of existence. We risk creating a world where we have incredible tools to achieve our goals, but no capacity to question if those goals are actually good for us. We become masters of the 'how' while losing all grip on the 'why'.

The Kantian Correction

Immanuel Kant offered the necessary counterweight. He rejected the idea that reason is just a clever instrument for satisfying appetites. For Kant, true reason includes the power to judge, to revise, and to legislate ends for itself. It is the faculty that allows us to look at a desire and ask: 'Is this worthy of a rational being?' This is the difference between a machine that finds the fastest route to a cliff edge and a human who decides not to walk there. Intelligence, in the fullest sense, requires the ability to step back from the objective and evaluate the objective itself.

The two models of reason

The Servant Model (Hume): Reason as a calculator of efficient means to satisfy pre-set desires.
The Sovereign Model (Kant): Reason as a judge capable of determining the value of the ends themselves.

As we integrate AI into the core of our decision-making, we must ensure we do not accidentally outsource our sovereignty. If we treat intelligence as nothing more than optimisation, we are essentially building a world of highly efficient slaves with no moral compass. We need systems that don't just help us get what we want, but help us understand if what we want is actually what we need.

Key Takeaway

Efficiency is a technical metric, not a definition of wisdom.

02 — Lenny's Newsletter

The Decision Engines

Moving beyond the chatbot era into the age of agency

By Claire Vo · 15 min read

Editor's note: The next phase of AI isn't about talking to a box; it's about building systems that act.

For the last two years, the primary way we have interacted with artificial intelligence has been through the chat interface. We type a prompt, we wait, we get a response. It is a conversational loop that feels intuitive but is fundamentally limited. It is a medium for exploration, not for execution. However, a shift is occurring. Developers are moving away from the idea of the 'chatbot' and toward the 'decision engine'. This is the move from generative models that produce text to agentic models that produce actions.

Jev and the Death of the Prompt

Take Jev, a new model being developed by John Lindquist. Jev isn't designed to write you a poem or explain quantum physics. It is designed to be a router—a high-speed, low-cost decision maker. Instead of a long-winded conversation, Jev takes an input and immediately maps it to a function or a specific path within an application. It is less about 'thinking' and more about 'directing'. When you tell a system like this you need to add milk to your grocery list, it doesn't discuss the merits of milk; it executes the command in milliseconds.

The future of AI is not a conversation; it is a series of rapid, invisible decisions.

This distinction changes the architecture of software. In the old model, the human is the driver and the AI is the passenger. In the new model, the AI acts as the nervous system of the application. It handles the messy, high-frequency tasks—data deduplication, routing, and command execution—leaving the high-level reasoning to more expensive, slower models. This allows for 'real-time' experiences, such as voice-to-do apps that respond without the awkward pause of a traditional LLM processing a request.

Practical applications of decision engines

Real-time voice commands with zero perceived latency.
Automated data cleaning by merging messy records using confidence scores.
App routing where a single text input navigates a user through complex software.
Multi-agent coordination to prevent task collisions in automated workflows.

The Cost of Speed

The move toward decision engines is driven by a brutal economic reality: latency and cost. Using a massive, reasoning-heavy model like GPT-4 to decide if a user clicked a button is like using a sledgehammer to crack a nut. It is too slow and far too expensive. The real value lies in the 'System 1' models—fast, instinctive, and cheap. By using these for the bulk of the heavy lifting, developers can build applications that feel alive and responsive, rather than sluggish and detached.

We are approaching a point where the 'AI' part of an application will become invisible. You won't know you are interacting with a model; you will simply find that your software anticipates your needs and executes your intent. The challenge for developers is no longer just about how well a model can write, but how reliably it can decide.

Key Takeaway

Stop building chatbots and start building agents that act.

03 — Lenny's Newsletter

The DevDay Disruption

Sorting the signal from the noise in OpenAI's latest release

By Claire Vo · 14 min read

Editor's note: OpenAI is moving from providing a brain to providing an entire operating system for agents.

OpenAI's DevDay 2026 was not just another product launch; it was an attempt to redefine the boundaries of what a developer can build. The sheer volume of announcements was overwhelming, but beneath the hype, a clear strategy emerged. OpenAI is no longer content with being the provider of a single, powerful model. They are positioning themselves as the foundational layer for a new kind of interactive, agentic web.

The Rise of Collaborative Spaces

One of the most significant, yet under-discussed, announcements was 'Spaces'. While much of the industry is focused on individual productivity, Spaces focuses on collaboration between humans and agents. It provides a shared environment where an AI isn't just a sidekick in a chat window, but a participant in a workspace. This is a fundamental shift in how we think about team dynamics. It moves the AI from a tool you use to a colleague you work alongside.

The goal is no longer just to automate tasks, but to integrate intelligence into the very fabric of teamwork.

Then there are 'Sites' and 'Connectors'. These tools aim to solve the data permission problem that has long plagued enterprise AI. By allowing teams to share internal tools with specific data access, OpenAI is attempting to bridge the gap between the playground of a demo and the reality of a secure corporate environment. It is an attempt to make AI useful in a way that doesn't require a complete overhaul of a company's security architecture.

Key releases to watch

GPT-6.1 Sol: A model optimized for the balance of speed and reasoning.
Astra Ultrafast: Enabling real-time, interactive AI applications.
Decisions API: Bringing vision-based decision making to developers.
Dots: New ways to interact with and personalize AI agents.

The Economics of Interaction

The release of Astra Ultrafast highlights a growing tension in the industry: the cost of immersion. When testing a 3D world that can be redesigned in real time via prompts, the costs can climb rapidly. A single experiment can cost nearly £80. This is the frontier of AI—highly interactive, visually rich, and incredibly expensive. For developers, the challenge is to find the sweet spot where the magic of real-time interaction meets the reality of a sustainable business model.

Ultimately, DevDay 2026 signaled that the era of the 'smart search engine' is over. We are entering the era of the 'intelligent environment'. Whether these tools can be deployed at scale without breaking the bank remains to be seen, but the direction of travel is unmistakable.

Key Takeaway

OpenAI is building the infrastructure for a world where AI is a teammate, not just a tool.

04 — Aeon

The Mimicry of Survival

How the green tree ant's success created an evolutionary arms race

By Aeon · 8 min read

Editor's note: In nature, looking like the winner is often more effective than being the winner.

In the dense, humid forests of the tropics, the green tree ant (Oecophylla smaragdina) is a dominant force. They are fierce, organized, and highly successful. But in the natural world, dominance is a magnet for opportunists. Because the green tree ant is so prevalent and so formidable, it has created a massive evolutionary incentive for other creatures to pretend to be them. This is the world of mimicry, where survival depends on a convincing performance.

The Art of the Imposter

Mimicry is not a one-size-fits-all strategy. It manifests in several distinct ways. Some insects have evolved to look like the ants to avoid being eaten. By mimicking the appearance of a fierce fighter, they trick predators into thinking they are much more dangerous than they actually are. It is a bluff that works, provided the disguise holds up under scrutiny.

In the tropics, your identity is your most valuable survival tool.

Other mimics take a more predatory approach. Instead of using the ant's reputation as a shield, they use it as a cloak. Certain spiders have evolved to resemble the green tree ant so closely—in look, behaviour, and even scent—that they can infiltrate the ant colonies. Once inside, they don't fight; they feast, preying on the ants and their young while remaining undetected by the colony's defenders.

Levels of mimicry

Visual Mimicry: Looking like the target to confuse predators.
Behavioural Mimicry: Acting like the target to blend into a group.
Chemical Mimicry: Smelling like the target to bypass biological security.

The Biological Arms Race

This constant pressure to deceive drives rapid evolutionary change. As mimics get better at pretending, the ants must get better at detecting the frauds. It is a perpetual cycle of refinement. The success of the green tree ant has effectively subsidised the evolution of dozens of other species, creating a complex web of deception that defines the tropical ecosystem.

The study of these imposters reveals a fundamental truth about biology: success breeds imitation. The more a species masters its environment, the more it becomes a template for others to follow, whether for protection or for predation.

Key Takeaway

Dominance in any system creates a vacuum that mimics will inevitably fill.

05 — Aeon

The Myth of Zoran

When a life becomes a collection of irreconcilable stories

By Aeon · 7 min read

Editor's note: A study in how memory and reputation can construct a person who never truly existed.

In Belgrade, there was once a man named Zoran. To some, he was a local legend; to others, he was a man who destroyed everything he touched. When filmmaker Maja Penčič set out to find him, she wasn't just looking for a person, but for a truth. However, she quickly discovered that when a life is lived with enough intensity, the truth becomes secondary to the stories people tell about it.

A Rotating Cast of Contradictions

The people who knew Zoran could not agree on a single version of him. The accounts were wildly divergent. One acquaintance described him as warm, intelligent, and exceptional. Another, with equal conviction, described him as a man who felt no shame and was, quite simply, a 'dick'. These weren't just minor disagreements; they were fundamentally different human beings. The film captures this through a home-video aesthetic that feels as fragmented and imperfect as the memories themselves.

He is, in simple terms, a man who destroys everything.

As Penčič surveys the people who claim to have known him, the details of Zoran's life begin to rearrange themselves. The film doesn't try to settle on a single, definitive biography. Instead, it allows the contradictions to exist side by side. It suggests that a person's identity is not a fixed point, but a shifting collection of impressions left on others.

The components of a legend

Inconsistent testimony from multiple sources.
The emotional weight of personal memory over factual accuracy.
The tendency to turn complex individuals into archetypes.

The Cautionary Tale

What remains at the end of the search is less a portrait of a man and more a cautionary tale about the nature of reputation. Zoran becomes a vessel for the projections of those around him. He is the hero in one story and the villain in the next, a man who has been erased by the very legends that keep his name alive.

The film forces us to confront a difficult question: if everyone's version of you is different, which one is the real you? Or is the 'real you' simply the sum of all these conflicting stories?

Key Takeaway

Reputation is not a reflection of who you are, but a reflection of who others need you to be.

06 — Psyche

The Full-Service Grandad

Finding purpose in the routine labours of care

By Liam Heneghan · 10 min read

Editor's note: A reflection on the dignity and the ticking clock of grandparenting.

There is a biological theory known as the grandmother hypothesis, which suggests that grandmothers increase the survival of their descendants by providing care. For a long time, the grandfather's role has been viewed through a more cynical lens—the 'grandfather effect', where grandfathers are seen as a drain on resources. But as I have entered this new stage of life, I have found a different path: the 'full-service grandad'.

The Labour of Love

Being a full-service grandad is not about being a figurehead or a source of stories. It is about the unglamorous, exhausting, and essential work of care. It is about changing 686 diapers and pull-ups. It is about overseeing feedings, nap times, and the endless, repetitive jaunts to the park. It is a commitment to the routine that keeps a family functioning. It is a role that requires presence, even when that presence is physically demanding.

The full-service grandad does all the routine labours of care, but must also ask what distinctive contributions he brings.

There is a particular tension in this role. As an ecologist, I am tempted to observe my granddaughter as a subject of study, to catalogue every sparrow and puddle she discovers. But the true value of being a grandfather lies in the opposite direction. I am not here to observe her; I am here to let her help me see the world again. Through her eyes, the mundane becomes remarkable.

The duties of the full-service grandad

Physical care (diapers, feedings, nap times).
Environmental immersion (park visits, nature walks).
Emotional presence (being available for the long haul).

The Ticking Clock

However, there is a weight to this service. Unlike the permanent career or the long-term research project, the phase of intensive care is finite. There is a clock ticking on the ability to change a diaper or walk a toddler to the park. This realization brings a certain urgency to the role. You cannot wait for the 'right time' to be present; the window of opportunity is narrow and moves quickly.

Ultimately, being a full-service grandad is about more than just helping out. It is about finding a new way to be useful, to find a way to contribute that is distinct from being a parent, and to accept the beautiful, exhausting reality of being part of a generational cycle.

Key Takeaway

Care is not just a service; it is a way of seeing the world anew.