The Deep Feed

01 — Lenny's Newsletter

The Mythos Reality Check

Testing Anthropic’s Claude Fable 5 against the promise of high-level reasoning

By Claire Vo · 8 min read

Editor's note: As models move into the 'Mythos' class, the gap between benchmark scores and actual utility becomes the new battleground.

Anthropic has released Claude Fable 5, the first model in its Mythos class to hit the general market. This isn't just another incremental update; it is a move toward a specific kind of reasoning capability that Anthropic claims sets a new ceiling for intelligence. For those building in the AI stack, the question isn't whether the model is 'smarter,' but whether that intelligence translates into reliable execution for complex, multi-step tasks. The marketing suggests a leap in reasoning, but the reality of using it on real-world product specs reveals a more complicated picture.

The Token Tax

One of the most immediate observations when working with Fable 5 is its appetite for tokens. The model is designed to be token-intensive, a choice that suggests Anthropic is prioritising depth of thought over cost-efficiency. This is a deliberate trade-off. By allowing the model more 'room' to process, they are aiming for a higher level of cognitive overhead. However, for an agency owner or a product lead, this translates to higher operational costs. You are paying for the ability to avoid the shallow logic that plagues smaller models, but that cost must be justified by the accuracy of the output.

The real test of a Mythos model isn't its ability to pass a benchmark, but its ability to manage a product graph without losing the thread.

In testing the model's ability to design a skills registry and a product graph spec, the results were mixed. Fable 5 shows a remarkable ability to grasp complex structures, yet it remains conservative in its execution. It often defaults to safer, more predictable paths rather than pushing the boundaries of a design. This conservatism might be a byproduct of the new safety classifiers, which act as a fallback mechanism. While this prevents the model from hallucinating wildly, it can also stifle the very creativity that high-level reasoning should enable.

Core observations from the Fable 5 testing:

High token consumption makes it a premium tool rather than a commodity.
Multi-agent orchestration shows promise but requires strict guardrails.
The model excels at structural logic but struggles with creative risk-taking.
Safety fallbacks can occasionally lead to overly cautious, unhelpful responses.

Ultimately, Fable 5 is a tool for high-stakes reasoning. It is not a replacement for a junior analyst, but rather a sophisticated engine for orchestrating complex workflows. If you are building managed agents, this model provides the backbone, but you will still need to provide the direction. The intelligence is there, but the agency still rests with the human architect.

Key Takeaway

High-level reasoning models are becoming more expensive and more cautious, trading speed and cost for structural accuracy.

02 — Lenny's Newsletter

The 15-Minute Avatar

Democratisation and the Uncanny Valley in AI Video

By Lenny Rachitsky · 7 min read

Editor's note: We are entering an era where creative production is no longer a matter of skill, but a matter of prompt engineering.

The barrier to high-quality video production has historically been a wall of expensive equipment and specialised skills. That wall is crumbling. Recent tests using Google Flow and Gemini Omni show that a non-professional can now clone themselves into an AI avatar and produce a complete hype reel in under fifteen minutes. This isn't just a neat trick; it is a fundamental shift in how content is created. The workflow—from scanning a face to generating scenes and stitching them together—is becoming a desk-bound activity rather than a studio-bound one.

The Consistency Problem

Despite the speed, the technology is not yet seamless. The most glaring issue is character consistency. An avatar might start a video with one hair length and end it with another, or find itself in a room where the background details shift from books to plants between cuts. This lack of environmental stability breaks the immersion. For professional branding, this is a significant hurdle. You cannot build a reliable identity if your digital twin changes its appearance every time the camera angle shifts.

AI video tools are turning creative production from a craft into a conversation.

Then there is the emotional problem. While side profiles and serious expressions can look remarkably real, any attempt at complex emotion often falls straight into the uncanny valley. A laugh can look forced, or a smile can look like a mask. The subtle muscle movements that signal genuine human feeling are difficult for current generative models to replicate. We are seeing the 'shape' of emotion without the 'soul' of it, which can be more unsettling than seeing no human at all.

What the current AI video workflow enables:

Rapid prototyping of visual ideas and storyboards.
High-speed content creation for social media and internal comms.
Lowering the cost of entry for solo entrepreneurs and small teams.
AI acting as a creative collaborator for scene brainstorming.

We are in a transitional phase. The tools are fast enough to be useful, but not stable enough to be perfect. For now, they are best used as collaborators—tools that help you brainstorm and structure a narrative—rather than final, hands-off production machines. The speed is the victory; the consistency is the next frontier.

Key Takeaway

AI video has solved the speed problem of production, but it has yet to solve the consistency and emotional authenticity problems.

03 — Lenny's Newsletter

The Anti-Amazon Strategy

Using Claude to Reclaim Quality in a World of Infinite Noise

By Claire Vo · 9 min read

Editor's note: In an era of drop-shipping and paid ads, AI can be used as a filter to find things that actually last.

The modern internet shopping experience is broken. For anyone seeking quality over mere convenience, the search for a durable product has become a battle against a deluge of drop-shippers, low-quality brands, and algorithmically boosted junk. Amazon and similar platforms have optimised for speed and volume, which is often the enemy of craftsmanship. To navigate this, some are turning to AI not to find more things, but to find better things. By using Claude Projects, users are building personal vetting systems that act as a shield against the noise.

Codifying Taste

The core of this approach is the creation of a 'Claude Project' populated with specific, reusable purchasing criteria. Instead of searching the entire internet, the user instructs the AI to prioritise a pre-vetted list of trusted vendors—brands with decades of history, strong return policies, and a commitment to natural materials. This flips the traditional shopping flow. You are no longer filtering through garbage to find a needle; you are searching within a curated haystack of quality.

AI allows us to automate the mental overhead of being a discerning consumer.

This system is particularly effective for small, heritage brands that often have terrible user experiences. These companies might have incredible products but outdated websites that are difficult to navigate. An AI agent can parse through their catalogues, find the specific item you need, and even help draft return requests or find receipts in your email history. It provides a layer of modern usability to old-world craftsmanship, allowing small artisans to compete with the infrastructure of giants.

How to build an AI-powered shopping system:

Create a list of 'trusted brands' based on history and materials.
Define strict criteria: repairability, return policies, and no 'DTC' hype.
Use AI to scan vendor sites that have poor search functionality.
Automate the administrative side of ownership, like returns and receipts.

This is a practical application of AI that moves beyond the hype. It is about using intelligence to reduce decision fatigue and reclaim time. By codifying our standards, we use technology to reinforce our values rather than letting algorithms dictate our consumption.

Key Takeaway

AI can be used as a sophisticated filter to bypass the noise of mass-market e-commerce and find high-quality, lasting goods.

04 — Cal Newport

The Job Apocalypse That Wasn't

Why AI is transforming roles rather than hollowing out workforces

By Study Hacks · 6 min read

Editor's note: The predicted 'jobs apocalypse' is being replaced by a more subtle, more complex reality of role transformation.

For the past two years, the dominant narrative around AI has been one of displacement. CEOs of major AI labs have frequently compared generative models to industrial automation, suggesting that just as machines replaced physical labour, AI would replace cognitive labour. The prediction was a massive hollowing out of white-collar roles. However, the reality on the ground is proving to be much less apocalyptic and much more interesting. The 'jobs apocalypse' is not happening; instead, we are seeing a fundamental shift in how work is structured.

From Power Looms to Hacking Tools

The industrial analogy—the power loom replacing the weaver—is fundamentally flawed when applied to intelligence. In practice, small businesses and professionals are not using AI to replace entire job functions. Instead, they are using it to 'vibe code'—hacking together quick, dirty, and highly specific tools to solve immediate problems. Much like a teenager writing a custom script to manage a company's inventory, professionals are using AI to automate the annoying, repetitive parts of their jobs, allowing them to focus on higher-leverage tasks.

AI isn't acting as a replacement for the worker, but as a highly efficient, albeit messy, assistant.

This leads to what is being called 'freestyle work.' Rather than following rigid, pre-defined workflows, employees are using AI to bend their roles to their own needs. They are building their own mini-systems, automating their own administrative burdens, and creating custom solutions for problems that their official software doesn't solve. This doesn't make the employee superfluous; it makes their efforts deeper and more focused.

Why the 'Job Apocalypse' narrative failed:

AI is better at assisting tasks than replacing entire roles.
The 'hacker' mentality of using AI for bespoke tools is more common than enterprise-wide automation.
Human agency and context remain difficult to automate.
Work is being transformed and deepened rather than eliminated.

The shift is subtle. We aren't seeing mass unemployment; we are seeing a shift in what it means to be a professional. The value is moving away from the ability to perform routine cognitive tasks and toward the ability to direct, orchestrate, and integrate AI-driven solutions into meaningful work.

Key Takeaway

AI is not a replacement for human workers, but a tool for hacking and augmenting the specific tasks that define their roles.

05 — Not Boring

Expanding the Radius

The Anthropological Invariant of Speed and Territory

By Packy McCormick · 7 min read

Editor's note: As we develop faster modes of transport, we don't save time; we simply expand the distance we are willing to travel.

There is a persistent question in technology: 'Where is my flying car?' For decades, this has been a shorthand for the failure of radical innovation to reach the daily lives of ordinary people. But the conversation is changing. With the rise of VTOL (vertical takeoff and landing) technology and electric aviation, the dream of expanded mobility is moving from science fiction toward engineering reality. The goal isn't just to get from point A to point B faster; it is to fundamentally change the scale of our daily lives.

Marchetti’s Constant

To understand why we need flying cars, we have to understand a strange human truth known as Marchetti’s Constant. The principle suggests that as transportation improves, humans don't use the saved time to rest; they use it to travel further. If a new technology makes a commute twice as fast, we don't spend half the time at home; we move further away and keep the commute at the same duration. We are, in a sense, biologically programmed to expand our territory in direct proportion to our speed.

We are born to spread. When we access a faster mode of transportation, we use it to travel further instead of traveling less.

This is an anthropological invariant. Our territorial behaviour dictates that as the cost of movement drops, the radius of our daily life expands. This has massive implications for urban planning, real estate, and the very concept of a 'city.' If VTOL technology becomes a reality, the distinction between urban and rural begins to blur. The 'radius' of where a person can live and work effectively expands, potentially decentralising human settlement in a way we haven't seen since the advent of the automobile.

The implications of expanded mobility:

Decentralisation of major urban hubs.
A fundamental shift in real estate value and geography.
Increased personal freedom in choosing living environments.
The expansion of the 'daily radius' for work and social connection.

The bet being made by frontier entrepreneurs is not just on a new vehicle, but on a new way of living. They are betting that the ability to move vertically and rapidly will unlock a new era of human expansion, allowing us to inhabit the world in a more distributed and less congested manner.

Key Takeaway

Technological speed does not save time; it expands the geographic boundaries of our lives.

06 — Lenny's Newsletter

The Product Builder's Library

Foundational wisdom for those who build things

By Lenny Rachitsky · 10 min read

Editor's note: In a world of rapid cycles, certain foundational principles remain the only reliable cheat codes.

In the fast-moving world of product development, it is easy to get caught up in the latest frameworks and tools. However, the most successful builders often rely on a set of foundational principles that have stood the test of time. Books, in this context, act as a way to access the distilled experience of a career in a few hours. They are the closest thing we have to cheat codes for real life. For a product builder, this library isn't about learning a specific software, but about understanding human psychology, design, and power.

The Mechanics of Design and Taste

Design is often treated as a subjective, 'squishy' discipline, but the best builders treat it as a science of usability. Works like 'Don't Make Me Think' and 'The Design of Everyday Things' move design away from aesthetics and toward objective functionality. They teach that when a user struggles, it is a failure of the design, not the user. Alongside this, there is the cultivation of taste. Taste is not an innate gift; it is a learned ability to recognise quality and manage the 'resistance' that comes with creating something new.

Books allow you to access the entire learnings of someone else's career in a few hours.

Beyond the interface, there is the psychology of influence. Understanding how people change their minds—through social proof, authority, and scarcity—is essential for anyone trying to drive growth or lead a team. Negotiation is another critical skill; moving from a 'you vs. me' mindset to a collaborative problem-solving approach is often the difference between a failed partnership and a successful product launch.

Essential categories for the builder's library:

Design: Focus on usability and objective UI principles.
Creativity: Learn to manage resistance and protect early ideas.
Influence: Master the fundamentals of persuasion and negotiation.
Strategy: Understand how to iterate, cross chasms, and solve problems.

The goal of reading these texts is not to memorise them, but to find a single 'nugget' of wisdom that can be applied immediately. In a career defined by constant change, these enduring ideas provide the stability required to build something that actually matters.

Key Takeaway

True expertise in building comes from mastering timeless principles of design, psychology, and strategy rather than chasing transient trends.