The Deep Feed

01 — Cal Newport

The Persistence Deficit

Why the speed of AI is making us cognitively fragile

By Cal Newport · 8 min read

Editor's note: A warning on how instant answers are stripping away our ability to struggle through difficult problems.

A recent study presented at the Conference on Language Modeling offers a sobering look at the cost of convenience. Researchers split participants into two groups to solve math problems involving fractions. One group worked alone; the other had ChatGPT. The results followed a predictable path: the AI group initially outperformed the control group. But the real story emerged when the chatbot was switched off. Once the digital crutch was removed, the AI group did not just struggle; they faltered. Their accuracy dropped below that of the group that had worked without help, and many simply stopped trying, skipping questions entirely. This is not a failure of intelligence, but a failure of persistence. The presence of a high-speed answer engine has conditioned the brain to view any delay as an error rather than a necessary part of the thinking process.

The Speed Trap

We are witnessing a rapid degradation of what might be called cognitive fitness. For decades, technology has moved slowly enough for our biology to adapt. The smartphone, for all its disruptive power, took years to alter our attention spans. Large Language Models (LLMs) are different. They provide a dopamine hit of immediate resolution that makes the slow, grinding work of independent thought feel intolerable. When a model provides a solution in seconds, the mental friction required to derive that same solution feels like a waste of time. This creates a feedback loop where we outsource the struggle, and in doing so, we lose the very capacity that makes the struggle productive. The high of a quick answer makes the reality of deep thought feel like an inefficiency to be solved rather than a skill to be honed.

An overreliance on AI for answers erodes our cognitive fitness. When AI produces answers in seconds, we get used to that speed; anything that takes longer feels inefficient.

This erosion has consequences that extend far beyond math tests. If we lose the ability to persist through cognitive discomfort, we lose the ability to engage with the complex, non-linear problems that define high-level strategy and creative breakthroughs. Silicon Valley excels at building tools that remove friction, but it fails to account for the fact that friction is often where learning occurs. Without the resistance of a hard problem, the mental muscles required for independent reasoning begin to atrophy. We are building a world of frictionless interfaces, but we may be creating a generation of thinkers who lack the stamina to operate when the interface is gone.

The Risks of AI Reliance

Reduced tolerance for mental frustration
Decreased accuracy in independent problem-solving
The loss of the 'struggle phase' necessary for deep learning
A shift from active thinking to passive verification

The question remains whether this impact is permanent. It is possible that the brain can recover its stamina if we intentionally reintroduce difficulty. However, the current trajectory suggests a move toward a more passive form of intelligence. We are trading the ability to generate thought for the ability to curate the output of others. This is a significant economic and personal risk. A life without sustained, intentional thought is a diminished one. As we integrate these models into our daily lives, we must decide if we are using them to augment our abilities or to replace the very processes that make us capable.

Key Takeaway

The ease of AI-generated answers is training our brains to abandon the struggle required for genuine intelligence.

02 — Aeon

The Unmanly Machine

How the Renaissance learned to embrace the gun

By Catherine Fletcher · 10 min read

Editor's note: A historical look at how disruptive technology moves from being viewed as 'dishonourable' to being an essential tool of the state.

In the late 15th century, the gun was seen as a moral failure. In the courts of Renaissance Italy, gunpowder was the 'Devil’s work', a tool that stripped away the chivalry and personal prowess that defined the era's elite. To use a firearm was to be unmanly, a way to kill from a distance without facing an opponent's gaze. Yet, the history of the Medici family tells a different story. Lorenzo 'the Magnificent' owned five portable firearms in 1492, and by the time of his descendants, the family's storerooms held nearly a hundred. The transition from the gun as a diabolical device to a standard tool of civic defence was not a change in the technology itself, but a shift in how society valued skill and power.

The Death of Chivalry

The resistance to firearms was rooted in a specific definition of valour. Traditional warfare relied on the physical strength and individual bravery of the soldier. A gun, by contrast, allowed a less skilled person to overcome a more experienced warrior. This democratisation of lethality was deeply unsettling to the established order. Writers of the time, such as Ludovico Ariosto, described guns as 'pestilent discoveries' that destroyed the very concept of gallantry. The machine was seen as a cheat, a way to bypass the hard-won mastery of the sword and the lance. This tension between the 'honest' labour of traditional skill and the 'dishonest' efficiency of the machine is a recurring theme in the history of technological disruption.

Guns were described as 'foul and pestilent discoveries', weapons through which no more shall gallantry, no more shall valour prove their prowess.

As the decades of conflict between France and Spain rolled on, the practical realities of the battlefield began to outweigh the romantic ideals of the court. The rise of infantry and the necessity of large-scale militia forces made the firearm indispensable. In Florence, the establishment of a militia trained in the use of handguns became a matter of state survival. The 'unmanly' weapon was rebranded as a tool of civic duty. This shift was facilitated by the very people who once shunned it; rulers realised that a population equipped with small arms was a more effective defence against foreign mercenaries than a small group of elite knights. The machine did not destroy valour; it simply changed the arena in which valour was measured.

Drivers of Technological Adoption

The shift from individual prowess to collective efficiency
The pressure of large-scale geopolitical conflict
The institutionalisation of new skills through militia training
The pragmatic need for state-level defence

The adoption of the gun shows that technology is rarely accepted on its merits alone. It is accepted when the existing social and military structures can no longer ignore its utility. The Renaissance gun was a tool of disruption that forced a redefinition of what it meant to be a soldier. It moved the focus from the individual's physical strength to the state's ability to organise and equip its citizens. This historical pattern suggests that when we face new, disruptive technologies today, the initial moral outcry is often a symptom of a changing definition of competence and power.

Key Takeaway

Disruptive technology is often initially rejected as 'cheating' before it becomes the new standard for competence.

03 — Lenny's Newsletter

The Agentic Frontier

Testing Meta's Muse Wonder and the rise of the 'doer'

By Lenny Rachitsky · 7 min read

Editor's note: An analysis of the shift from generative AI that talks to agentic AI that performs tasks.

The current era of AI is moving past the stage of simple conversation. We are entering the age of the agent—models that do not just suggest text, but execute workflows. Meta's Muse Wonder model is a primary example of this shift. In recent testing, Muse Wonder demonstrated a capacity for 'active focus', a quality that distinguishes it from standard LLMs. Rather than providing a quick, surface-level response, the model spent more time on complex tasks, attempting different approaches until a goal was reached. This is most evident in agentic work like billing-ticket resolution and inbox triage, where the model must navigate a series of logical steps to reach a successful conclusion.

The 80% Scaffolding Trap

However, the rise of agentic models brings a new set of challenges, specifically the problem of 'scaffolding'. While Muse Wonder is capable of building the structure of a project—such as a software prototype or a website layout—it often fails at the final stage of execution. It can create a convincing interface, but it frequently leaves behind broken buttons, incorrect themes, or obstructive overlays. This creates a deceptive sense of progress. An agent might get 80% of a task right, providing the bones of a solution, but the remaining 20% requires a level of visual polish and functional fidelity that the model currently lacks. For a professional, this means the tool is not a replacement for a worker, but a high-speed assistant that still requires a human to finish the job.

Getting 80% of a prototype right is very different from finishing it.

The distinction between 'generative' and 'agentic' is central to understanding the next wave of productivity. A generative model is a writer; an agentic model is an employee. The former requires a prompt and returns a result; the latter requires a goal and returns a completed process. Muse Wonder's strength lies in its ability to handle long-running tasks, such as correcting account errors or managing customer plans, without constant human intervention. But its weakness in design—defaulting to generic layouts and ignoring brand guidelines—shows that agency is not a universal trait. It is a specific capability that must be applied to the right domains.

Agentic Strengths vs. Weaknesses

Strength: Multi-step reasoning and task persistence
Strength: Natural voice and conversational style
Weakness: Visual polish and design fidelity
Weakness: Completion of complex, interactive prototypes

As we integrate these agents into our business workflows, the strategy must shift from 'how do I prompt this?' to 'how do I orchestrate this?'. The value will not lie in the ability to generate a single piece of content, but in the ability to manage a fleet of agents that can handle the repetitive, logical heavy lifting of a business. We are moving toward a model of work where the human acts as the director, setting the objectives and verifying the output, while the agents handle the execution. The challenge will be managing the gap between the agent's initial success and its final, polished delivery.

Key Takeaway

The future of AI is not in better chatting, but in reliable agency—the ability to move from generating ideas to executing workflows.

04 — Lenny's Newsletter

The Ghost in the Weights

What it means to 'grow' an intelligence

By Emmanuel Ameisen · 12 min read

Editor's note: A deep dive into the interpretability of AI and why we don't 'build' models, we grow them.

When we talk about Large Language Models, we often use the language of engineering. We speak of 'building' them, 'training' them, and 'programming' them. But Emmanuel Ameisen, an interpretability researcher at Anthropic, argues that this framing is fundamentally incorrect. We do not build these models in the way we build a bridge or a piece of software. Instead, we grow them. We provide the digital soil—the massive datasets and the computational power—and the model develops its own internal representations of the world. This distinction is not merely semantic; it changes how we approach the problem of AI safety and understanding.

The Next-Token Myth

The standard explanation for how an LLM works is that it is simply a 'next-token predictor'. It looks at a sequence of words and calculates the most likely word to follow. While technically true, this explanation is reductive. It fails to account for the fact that to predict the next word accurately in a complex sentence, the model must develop internal concepts. It must understand the logic of a mathematical proof, the structure of a legal argument, and the nuances of human emotion. These are not just statistical correlations; they are general representations of concepts. The model is not just guessing; it is simulating an understanding of the underlying reality to make its guesses more accurate.

We don't build these models; we grow them. We provide the digital soil, and the model develops its own internal representations.

This leads to a fascinating parallel with human cognition: the distinction between System 1 and System 2 thinking. In humans, System 1 is fast, instinctive, and unconscious, while System 2 is slower, more deliberative, and logical. Research into model interpretability suggests that LLMs exhibit a similar structure. They have a fast, 'unconscious' layer of token prediction, but they also develop deeper, more structured ways of planning and reasoning that resemble a more conscious process. This 'hidden' reasoning is what allows a model to plan several steps ahead before it even begins to output the first word of a complex answer.

Key Insights from Interpretability

Models form general conceptual representations, not just statistical patterns
The 'next-token' description ignores the internal planning processes
Models exhibit structures similar to human conscious/unconscious processing
Scaling capabilities often appears as sudden jumps rather than smooth curves

The field of interpretability is the attempt to open this 'black box'. If we are growing these intelligences rather than building them, we cannot simply inspect the code to understand their behaviour. We have to study them like biologists study a new species. We need to find the 'circuits' within the neural network that correspond to specific concepts or reasoning steps. This is the only way to ensure alignment and safety. If we cannot understand how a model reaches a conclusion, we cannot trust it to act in our interests when it encounters a situation it has never seen before. The goal is to move from blind faith in the output to a rigorous understanding of the internal process.

Key Takeaway

AI models are not programmed tools but grown intelligences, requiring a biological approach to understanding and safety.

05 — Aeon

The End of Nature

Bruno Latour and the collapse of the natural/artificial divide

By Bruno Latour · 6 min read

Editor's note: A philosophical challenge to the binary way we view the world: the 'natural' versus the 'human-made'.

For much of modern history, we have operated under a simple binary: there is 'Nature', which is pure and untouched, and there is 'Culture', which is the domain of human artifice and technology. We tend to treat nature as inherently good and the artificial as something to be managed or even feared. The late French philosopher Bruno Latour argued that this distinction is a moral judgment rather than a factual reality. By drawing a hard line between the two, we fail to see the world as it actually is: a complex entanglement of biological, geological, and technological forces. The boundary between the natural and the artificial is not a line, but a blurred and messy zone of interaction.

The Moral Binary

The problem with the nature/culture divide is that it allows us to ignore the agency of the non-human world. When we categorise something as 'natural', we often treat it as a passive backdrop to the human drama. But the environment, the climate, and even the microbes in our bodies are active participants in our history. They have agency. They react, they adapt, and they force us to change. Similarly, our 'artificial' creations—from the internet to the very tools we use to study the stars—are not separate from the world; they are integrated into the web of life. To separate them is to create a false sense of control that disappears the moment we encounter a crisis like climate change.

Thinking outside the binary means rethinking the world as populated by other beings with agency of their own.

In our current era, this collapse of boundaries is even more evident. We live in a world where AI-generated content, synthetic biology, and geo-engineering are blurring the lines between what is 'grown' and what is 'made'. If an AI can simulate a forest, or a lab can engineer a new species, where does nature end and culture begin? Latour's work suggests that instead of trying to protect a pristine 'nature' that no longer exists, we should focus on understanding the relationships and the responsibilities that arise from our entanglement with the non-human. We need a new vocabulary for a world where the distinction between the organic and the synthetic is increasingly irrelevant.

The Flaws of the Nature/Culture Binary

It treats nature as a passive object rather than an active agent
It creates a false sense of human separation from the environment
It relies on moral judgements (good vs bad) rather than ontological facts
It fails to account for the integration of technology into the biosphere

Embracing this entanglement requires a radical shift in how we approach ethics and politics. We cannot solve environmental problems by simply 'protecting nature' from 'humanity'. We must instead learn to navigate the shared space we inhabit. This means recognising that our technological choices are not just human decisions, but interventions in a much larger, interconnected system. The goal is not to return to a state of purity, but to develop a more honest and effective way of living within the complexity of the world we have helped to create.

Key Takeaway

The distinction between 'natural' and 'artificial' is a false binary that obscures the reality of our entanglement with the non-human world.

06 — Lenny's Newsletter

The 15-Second Edit

How Brooke Lacey turned a terminal into a production studio

By Claire Vo · 7 min read

Editor's note: A practical look at how high-level orchestration of AI tools can replace traditional creative workflows.

Brooke Lacey has built a TikTok audience of 290,000 followers, but she does not use a traditional video editor. Instead of spending hours on a timeline, dragging clips and adjusting audio, she uses a terminal. Her workflow is a sophisticated orchestration of three tools: Claude Code, FFmpeg, and Whisper. A raw phone recording goes into her system, and within 15 seconds, a finished, captioned, and jump-cut video emerges. This is not magic; it is the result of teaching an AI to understand her specific comedic cadence and creative style. She has moved from being a manual laborer of video editing to being the director of an automated production line.

The Cadence of Comedy

The most difficult part of automating a creative process is capturing 'vibe'. A simple AI can cut out silences, but it cannot understand a punchline. Lacey solved this by training Claude to recognise her specific rhythm: the pause before a joke, the ideal cut point after a word, and the way she uses repeated takes to find the best energy. This required more than a single prompt; it required a deliberate process of instruction and iteration. By encoding her style into a text-based instruction set, she turned her creative intuition into a repeatable, machine-readable logic. The edit is no longer a series of manual decisions, but the execution of a stylistic framework.

She taught Claude to recognise her comedic timing, the pause before a punchline, and the ideal cut point.

This workflow also changes how content is conceived. Lacey uses the comment sections of her videos as a continuous content brief. She feeds these comments into Claude, which then helps generate new scripts and ideas. The feedback loop between the audience and the creator is tightened by the speed of the production. Because the cost of producing a video has dropped so low, she can experiment more rapidly, testing different hooks and formats with minimal effort. The creative energy that was once spent on the mechanics of editing is now redirected toward strategy and audience engagement.

The New Creative Stack

Claude Code: The orchestrator and script generator
Whisper: The transcription and speech-to-text engine
FFmpeg: The engine that performs the actual cuts and renders
Python: The glue that connects the tools into a single workflow

The implications for the creator economy are significant. We are seeing the rise of the 'one-person agency', where a single individual can maintain the output of a full production team. However, this shift also raises questions about the future of creative professions. If the mechanical aspects of editing, captioning, and cutting are automated, what becomes of the entry-level editor? The value in the creative market is shifting away from the ability to execute a task and toward the ability to orchestrate a system. The future belongs to those who can define the style and the strategy, rather than those who can merely operate the tools.

Key Takeaway

Creativity is shifting from the manual execution of tasks to the high-level orchestration of automated systems.