The Deep Feed

01 — Lenny's Newsletter

The Engineering of Intent

Inside the rapid evolution of Anthropic's product philosophy

By Lenny Rachitsky · 12 min read

Editor's note: A look at how product management functions when the technology moves faster than the roadmap.

In the early days of Anthropic, the product team was not a department; it was a handful of engineers trying to make sense of a shifting frontier. Dianne Penn joined when the company was still finding its feet, stepping into a role that required more than just managing features. She had to manage the very nature of intelligence. As the first technical product manager, her task was to bridge the gap between raw research and the tools that people actually use. This wasn't about deciding which button to place where; it was about deciding how a model should behave when it encounters a problem it wasn't specifically trained for. The transition from Claude 2 to the current generation represents a shift from building a chatbot to building an agentic partner.

The Eval-Driven Loop

Development at this scale does not rely on intuition. It relies on 'evals'—rigorous, automated tests that measure how a model performs against specific benchmarks. Penn describes a loop where development is driven by these evaluations. You don't just 'release' a model; you iterate through a cycle of testing, identifying failure modes, and refining the training data. This process is what allowed Claude to develop its specific character. It is an iterative refinement of capability, where the goal is to move the needle on reasoning and tool use without breaking the safety guardrails that define the brand.

The success of a model lies in its willingness to push back when a user's request is illogical or harmful.

One of the most interesting developments is the move toward 'computer use' and tool use. We are seeing a transition where the model is no longer just predicting the next word in a sentence, but predicting the next action in a workflow. This requires a different kind of product thinking. You are no longer designing a conversation; you are designing a capability. When Claude can use a browser or write code to solve a geometry problem, the product is no longer the text it produces, but the work it completes.

Key stages of Anthropic's growth

The early research phase with a core engineering team
The transition to model-driven product development
The expansion into agentic capabilities like Claude Code
The integration of tool-use and reasoning as core features

The speed of this evolution creates a permanent state of instability. For a product manager, this means the roadmap is never finished. You are building on shifting sands. The challenge is to maintain a coherent product vision while the underlying technology changes every few months. It requires a focus on the fundamental qualities of the model—its reasoning, its reliability, and its utility—rather than just a list of superficial features.

Key Takeaway

Product management in AI is the art of managing capabilities rather than just features.

02 — Not Boring

The Oil Economy

Why the giants of tech are betting on open-weight models

By Packy McCormick · 8 min read

Editor's note: An analysis of the strategic alignment between competitors in the race for AI dominance.

There is a common misconception that the tech industry is a monolith of closed-door secrecy. In reality, the recent support for open-weight models by leaders like Satya Nadella and Jensen Huang suggests a much more complex game of strategic alignment. When the heads of Microsoft and Nvidia sign a letter supporting open access, it isn't necessarily out of a sense of altruism. It is a calculated move to commoditise the complements of their own products. In the economics of AI, the model is the engine, but the chips and the cloud infrastructure are the fuel. By making the engines cheaper and more plentiful, they ensure a higher demand for the fuel.

Commoditising the Complement

This is classic tech strategy. If you sell the shovels, you don't want the gold to be so rare that people stop digging. If high-end models remain locked behind expensive, proprietary gates, the total volume of AI implementation might slow down. But if capable models become a commodity—cheap, accessible, and easy to run on standard hardware—then the entire ecosystem expands. Nvidia sells more chips, Microsoft sells more Azure credits, and companies like Dell and ServiceNow find more use cases for their hardware and software. The goal is to prevent a duopoly of model labs from controlling the entire value chain.

AI is oil, not a god; it is an economically useful commodity that scales and multiplies everything we do.

The tension lies between those who see AI as a dangerous force requiring strict control and those who see it as a utility. The 'God' camp argues that open weights allow bad actors to bypass safety filters. The 'Oil' camp argues that the economic benefits of widespread access far outweigh the manageable risks. The current movement toward openness suggests that the market is leaning toward the latter, driven by the sheer economic gravity of the hardware and infrastructure layers.

The beneficiaries of the open-weight movement

Hardware providers like Nvidia who see increased compute demand
Cloud providers who host the models
Enterprise software companies that build on top of open models
Developers who need customisable, low-cost intelligence

This alignment of self-interest and consumer benefit is what defines modern capitalism. It isn't about a shared vision of the future, but about a shared desire for a larger market. As models become more commoditised, the real value shifts from the model itself to the specific application and the data used to fine-tune it. The battle for the frontier is being fought not just in the labs, but in the very structure of the market itself.

Key Takeaway

Tech giants support open models because commoditising intelligence increases the demand for the hardware that runs it.

03 — Simon Willison

The Proactive Engine

Claude Opus 5 and the end of the passive tool

By Simon Willison · 6 min read

Editor's note: A technical look at the shift from reactive models to proactive agents.

The release of Claude Opus 5 marks a subtle but significant shift in how we interact with large language models. For years, the paradigm has been reactive: a user provides a prompt, and the model provides a response. Opus 5, however, is being described as a proactive model. It doesn't just wait for instructions; it anticipates the necessary steps to complete a complex task. This is the difference between a calculator and a researcher. A calculator waits for numbers; a researcher looks for the context required to make sense of those numbers.

Solving the Vision Gap

One specific example of this proactivity involves a task where the model was asked to write code for a 3D machine part based on a drawing, but was denied direct access to the image. Instead of failing or asking for the file, Opus 5 wrote its own computer vision pipeline to extract the geometry from raw pixels. This is not just 'following instructions'; it is solving the problem of how to follow them. The model identified a gap in its own capability and built a bridge to cross it. This is the hallmark of true agency.

We are moving from models that answer questions to models that solve problems.

This proactivity brings new questions regarding safety and control. Anthropic has made a deliberate choice in its training: while Opus 5 is highly capable of finding cybersecurity vulnerabilities, it has been specifically prevented from learning how to exploit them. This is an attempt to create a model that is useful for defenders without being a weapon for attackers. It is a delicate balance, attempting to build a highly intelligent entity while strictly limiting its destructive potential.

What makes Opus 5 different

Proactive problem-solving instead of reactive response
The ability to build its own tool pipelines on the fly
High reasoning capability at a lower price point
Intentional training gaps in cyber-exploitation

As these models become more capable of independent action, the role of the human shifts from 'operator' to 'supervisor'. We are no longer typing every command; we are setting objectives and reviewing the work. This requires a new kind of literacy—one that focuses on verifying the logic of an agent's path rather than just checking its final output.

Key Takeaway

The next generation of AI is defined by proactivity: the ability to build the tools necessary to complete a task.

04 — The Marginalian

The 100-Millisecond Mirror

Oliver Sacks and the mechanics of being conscious

By Maria Popova · 10 min read

Editor's note: A philosophical and neurological exploration of how we experience reality.

We often assume that perception is a direct window into the world—that we see things as they are. But as the late neurologist Oliver Sacks explored, perception is not a window; it is a creative act. There is a gap between the raw data of the world and the experience of it. This gap is filled by the brain's ability to sieve the salient from the infinite. We do not experience the vast, chaotic stream of sensory input; we experience a curated, meaningful version of it, shaped by our history, our needs, and our biology.

The Perceptual Moment

Sacks identifies a specific temporal dimension to this experience: the 'perceptual moment'. For a sensation to reach consciousness, it must do more than just trigger a neuron. It must cross a threshold of intensity and be maintained for roughly a hundred milliseconds. During this brief window, billions of neurons across different parts of the brain form a 'coalition'. These neural coalitions talk to one another in a continuous back-and-forth, transforming objective cerebral computation into a subjective experience. This hundred-millisecond window is where the world becomes 'real' to us.

We do not merely calculate movement as a robot might; we perceive it.

This explains why two people can walk down the same street and see entirely different worlds. One person might notice the architecture, while another notices the movement of the crowd. Our attention is the aperture through which we narrow the infinite vista of the present into something we can bear. Perception is a mirror that reflects our own internal state back at us as much as it reflects the external world.

Requirements for conscious perception

Crossing a threshold of neuronal intensity
The formation of neural 'coalitions' across brain regions
Maintenance of the signal for approximately 100 milliseconds
The integration of sensory data with prior experience

Understanding this process changes how we view the relationship between mind and matter. It suggests that consciousness is not a passive recipient of information, but an active participant in the construction of reality. We are not just observers of the universe; we are the architects of the experience of it.

Key Takeaway

Consciousness is not a direct recording of reality, but a 100-millisecond creative construction.

05 — The Marginalian

The Beast in the Jungle

Henry James and the tragedy of the expectant life

By Maria Popova · 9 min read

Editor's note: An examination of the psychological trap of living for a future that may never arrive.

There is a specific kind of paralysis that comes from waiting for life to truly begin. Henry James explored this in his novella *The Beast in the Jungle*, through a protagonist who lives his entire life in anticipation of a 'great event'—something rare and strange that will either destroy him or remake him. He calls this event the 'beast', and he spends his years withholding his participation in the world, waiting for the moment that will justify his existence. He avoids love, avoids risk, and avoids the messy reality of the present, all to protect himself from the uncertainty of the future.

The If-Only Mind

This is the 'if-only' mind—the psychological habit of treating the present as a mere waiting room for a more significant future. We tell ourselves that we will be happy once we get the promotion, once we find the right partner, or once we reach a certain level of stability. In doing so, we vacate the only reality we actually possess. We treat the current moment as a placeholder, a period of tedium to be endured until the 'real' life commences. But as James shows, the beast is not an external event; the beast is the act of waiting itself.

As long as we are waiting, we are not living.

The tragedy of the protagonist is that by the time he realises that the transformative moment has arrived, he has lost the capacity to engage with it. He has spent so long bracing for impact that he has become brittle. The 'beast' finally strikes, but it does so in the form of missed opportunity and the realization that time has foreclosed all possibilities. He achieves his fate not through a grand catastrophe, but through the quiet failure of never having been anything at all.

The symptoms of the expectant life

Withholding emotional intimacy to avoid risk
Viewing current experiences as mere preparation for the future
A constant sense of being 'kept' for something greater
The belief that stability is a prerequisite for meaning

To escape this trap, one must accept that transformation does not require a grand stage. It happens in the small, often unglamorous acts of participation—in the decision to love despite the risk of loss, or to act despite the lack of certainty. The only way to defeat the beast is to stop waiting for it and start inhabiting the jungle of the present.

Key Takeaway

Waiting for a transformative moment is the very thing that prevents transformation from happening.

06 — Simon Willison

The Token Shadow

The underground economy of LLM arbitrage

By Simon Willison · 5 min read

Editor's note: An investigation into the grey market of resold AI tokens and the risks it poses to developers.

As the demand for large language models has exploded, a shadow economy has emerged to exploit the gap between official API pricing and the actual cost of access. This is the 'relay market'—a system where resellers pool API keys from various sources to offer discounted tokens to users. Primarily concentrated in China, this market operates by abusing free trials, proxying through unprotected support bots, or using stolen credit cards. It is a digital arbitrage that thrives on the vulnerabilities of the current AI deployment model.

The Mechanics of Abuse

The software driving this market is often legitimate, such as 'one-api' or its forks, which are designed to load-balance requests across multiple credentials. The illegality lies in the source of those credentials. By aggregating thousands of compromised or trial accounts, resellers can offer prices that no legitimate provider can match. The buyers are often looking for three things: cheap access, a way to bypass geographic restrictions, or a source of data for model distillation—the process of using a large model to train a smaller, cheaper one.

The existence of this marketplace makes every developer more cautious about exposing their applications.

For developers, this creates a significant security and financial risk. If an application is public-facing, it becomes a target for these resellers. They can find unprotected endpoints and use them to fuel their proxy services, leaving the original developer with a massive, unexpected API bill. The current model of API management, which often lacks granular, real-time spending controls, is ill-equipped to handle this type of automated exploitation.

Risks of the relay market

Financial loss due to API key abuse
Data leakage through unauthorized proxying
Model distillation using stolen intelligence
Increased difficulty in managing geo-restrictions

To combat this, LLM vendors need to move beyond simple usage limits. They require tools that allow developers to set hard, real-time thresholds—for example, an application that automatically shuts down the moment it hits a specific dollar amount. Until the infrastructure catches up with the sophistication of the exploiters, the shadow economy of tokens will continue to grow.

Key Takeaway

The rise of AI has created a new class of digital arbitrage based on the exploitation of API vulnerabilities.