The Deep Feed

01 — Lenny's Newsletter

The Death of the Prompt: Why Talking is the New Coding

How voice interfaces and 'thread-forking' are turning AI from a search engine into a digital staff member.

By Lenny Rachitsky · 12 min read

Editor's note: A look at how the most effective AI users aren't prompt engineers, but people who simply talk to their machines.

For the last two years, the industry has been obsessed with the 'perfect prompt'. We have treated the text box as a sacred altar where precision is the only way to achieve competence. But the era of staring at a blinking cursor, trying to engineer the exact sequence of words to get a desired output, is ending. The real breakthrough isn't in better syntax; it is in the removal of the interface itself. When you stop typing and start talking, you stop being a programmer and start being a manager. This is what some are calling the 'yapper's API'—the idea that the more freely you speak, the more context you provide, and the better the machine performs.

The Rise of the Agentic Workflow

The shift from a chatbot to an agent is defined by the ability to act on the world, not just describe it. We are seeing the emergence of tools like ChatGPT Codex that do not merely answer questions but operate the computer. Imagine a workflow where you trigger a hotkey, an 'orb' reads your screen, and suddenly, separate threads are spinning up to book a flight, file an expense report, and check your calendar simultaneously. This isn't sequential processing; it is parallel execution. The machine is no longer waiting for your next command; it is managing the branches of a task while you move on to something else.

The most effective AI users aren't prompt engineers; they are people who simply talk to their machines.

One of the most significant technical shifts here is 'thread-forking'. In traditional software, a conversation is a straight line. In an agentic workflow, a task can become complicated, and the AI can decide to open a new thread, branch from an existing one, or pull context from a conversation you had three days ago. This removes the cognitive load of organisation. You don't need to prepare a perfect brief; you just need to provide enough context through natural speech for the machine to build its own structure.

The New Workflow Stack

Voice as the primary context provider
Parallel task execution via thread-forking
Screen-reading for real-time environmental awareness
Automated 'heartbeats' for asynchronous monitoring

As these tools mature, the bottleneck will shift from intelligence to latency. A model that is slightly less smart but responds instantly is often more useful than a brilliant model that makes you wait ten seconds. The goal is immediacy. When the AI can watch your screen, hear your rambling thoughts, and execute three tasks in the background, the computer ceases to be a tool and becomes a collaborator. The barrier to entry for complex digital work is collapsing, replaced by the ability to simply direct a capable subordinate.

Key Takeaway

Stop trying to write perfect prompts and start providing rich, verbal context.

02 — Lenny's Newsletter

The Product Manager's Regret

Why the traditional role of PM is being dismantled by high-agency teams and AI.

By Lenny Rachitsky · 10 min read

Editor's note: A provocative look at why the 'middleman' role of product management is becoming obsolete.

There is a growing sentiment among top-tier product leaders that the traditional role of the Product Manager (PM) is a mistake. At Whatnot, one of the fastest-growing marketplaces in history, the product team was founded on the premise that they 'regret that product management exists'. This sounds like heresy in Silicon Valley, where the PM is often seen as the glue holding engineering and design together. But the reality is that 'glue' often becomes 'friction'. When you insert a layer of management between the people building the product and the people using it, you risk creating a layer of bureaucracy that prioritises documentation over delivery.

The Shift to Senior ICs

The modern trend is moving away from large teams of junior PMs who spend their days writing tickets and managing stakeholders, and toward senior Individual Contributors (ICs) who actually build. In an era where AI can handle the heavy lifting of data analysis, documentation, and even basic coding, the need for a 'translator' between departments diminishes. If an engineer can use an LLM to query a database, understand user sentiment, and draft a technical spec, the traditional PM's primary value proposition—information synthesis—evaporates.

The PM role often becomes a layer of friction that prioritises documentation over delivery.

Tom Verrilli, CPO of Whatnot, suggests a different model: hiring great people and getting out of their way. This is the opposite of the 'command and control' style of product management. Instead of a central authority dictating the roadmap, the focus shifts to creating systems where high-agency engineers and designers can move at speed. This requires a move toward 'systems thinking'—understanding how different parts of a complex machine interact, rather than just managing a list of features.

Why Traditional PM is Fading

AI handles data synthesis and documentation
High-agency engineers are becoming more autonomous
The cost of communication overhead is too high for fast-moving startups
The rise of the 'builder' over the 'coordinator'

This doesn't mean product thinking is dead; it means the *role* of the PM is being absorbed into other disciplines. The best products are built by people who understand the user, the business, and the code simultaneously. The era of the specialist coordinator is giving way to the era of the generalist builder. For agency owners and leaders, the lesson is clear: stop hiring people to manage processes, and start hiring people to solve problems.

Key Takeaway

The future belongs to builders who can think like product managers, not managers who pretend to build.

03 — Simon Willison

The Open Source Mandate

Why the future of developer tools depends on transparency and LLM-assisted hacking.

By Simon Willison · 7 min read

Editor's note: A technical argument for why closed-source developer tools are a strategic risk.

The traditional argument for open-source software has always been about freedom: the ability to inspect, modify, and repair the tools we rely on. For most people, however, that freedom was an illusion. Even expert programmers rarely have the time to dive into the source code of their daily drivers. The friction of setting up an environment, compiling code, and understanding a foreign architecture was simply too high. We accepted closed-source tools because they were easier to use, even if they were black boxes. But the arrival of Large Language Models has fundamentally changed the cost of curiosity.

LLMs as the Great Equaliser

We are entering an era where the 'zero-time investment' challenge is becoming a reality. Instead of spending hours trying to figure out how a piece of software works, a developer can now simply ask an LLM to 'clone this repository and explain the logic of this specific module'. The barrier to entry for hacking on existing software has collapsed. If you can prompt a model to build the environment and walk you through the logic, the primary excuse for using closed-source tools—the time cost of understanding them—disappears.

LLMs have turned the friction of code comprehension into a zero-time investment challenge.

This shift creates a massive strategic imperative for developer tools. If the cost of understanding open-source code drops to near zero, then closed-source tools lose their primary competitive advantage: ease of use through abstraction. Developers will naturally gravitate toward open-source tools that they can instantly adapt to their specific needs using AI. A tool that is 'closed' is now a tool that is 'unimprovable', and in a fast-moving technical environment, being unimprovable is a death sentence.

The New Open Source Advantage

Instant comprehension via LLM explanation
Automated environment setup and compilation
Rapid prototyping of custom modifications
Reduced reliance on vendor-specific abstractions

For the industry at large, this means the 'black box' model of software is becoming increasingly untenable. As developers gain the ability to manipulate any tool they touch, the value of transparency increases. The winners in the next decade of software will not be those who build the most impenetrable walls around their code, but those who build the most extensible and understandable foundations for others to build upon.

Key Takeaway

When AI makes code easy to understand, closed-source software becomes a liability.

04 — Simon Willison

The Safety Paradox

The tension between open-weight models and the push for centralized AI control.

By Simon Willison · 9 min read

Editor's note: An analysis of the political and technical battle over who controls the weights of frontier models.

A quiet but intense war is being waged over the future of AI development. On one side, a coalition of tech giants and companies like NVIDIA and Amazon are advocating for 'open weights' and American leadership. On the other, companies like Anthropic are sounding the alarm on the existential risks of uncontrolled, open-source AI. The debate is framed as a choice between safety and progress, but the reality is much more about power and the concentration of capability. The question isn't just whether AI is safe, but who gets to decide what 'safe' looks like.

The Argument for Open Weights

Proponents of open-weight models argue that relying on a handful of closed-source providers creates a dangerous single point of failure. If a single company's model is breached or fails, the impact is systemic. Furthermore, closed models prevent the global research community from identifying vulnerabilities and developing safeguards. Open weights allow for a distributed, democratic approach to safety, where thousands of researchers can stress-test the technology. It is an argument for resilience through transparency rather than security through obscurity.

Concentrating advanced AI capabilities behind a small number of closed models compounds systemic risk.

Conversely, the 'pacing' movement—supported by thousands of employees at frontier AI companies—argues that the speed of development is outstripping our ability to govern it. They fear that automated AI research will create a feedback loop that moves faster than any human institution can manage. There is a genuine concern that the ability to build highly capable models could be used for cyberattacks or biological warfare, and that once the 'weights' are out in the wild, they cannot be taken back.

The Two Sides of the AI Debate

Open Weights: Resilience, competition, and distributed safety
Closed Models: Controlled deployment, mitigation of misuse, and centralized governance
The Risk: Automated research accelerating beyond human oversight

The tension here is fundamental. We are trying to balance the benefits of a widely distributed, innovative ecosystem against the risks of a technology that can be weaponised. The outcome of this debate will determine whether the AI era is defined by a few massive, centralized platforms or by a sprawling, decentralized landscape of specialized tools. For the agency owner, this is not just a policy debate; it is a question of which infrastructure your business will eventually sit upon.

Key Takeaway

The fight over open weights is a fight over whether AI becomes a public utility or a private monopoly.

05 — Stratechery

Meta's Timing Problem

Why even massive earnings can't mask the uncertainty of the AI transition.

By Stratechery · 8 min read

Editor's note: A look at the financial reality of pivoting a giant toward a new technological era.

Meta's recent earnings report presents a classic dilemma for the incumbent giant. On the surface, the numbers are functional, but beneath the surface, there is a growing sense of unease regarding the company's timing. The transition from a social media company to an AI-first company is not just a matter of changing a mission statement; it is a massive, capital-intensive pivot that requires everything to go right at exactly the right moment. The market is no longer satisfied with steady growth in advertising; it wants to see the clear, monetisable path to AI dominance.

The Cost of the Pivot

The sheer scale of Meta's investment in AI infrastructure is staggering. This isn't just about software; it's about the physical reality of chips, data centres, and energy. The 'financial tail' of these investments means that even if the AI products are successful, the margins may look very different for a long time. There is a tension between the need to spend aggressively to stay ahead and the need to maintain the profitability that keeps shareholders happy. Meta is essentially running a race where the finish line keeps moving further away.

Meta is running a race where the finish line keeps moving further away.

The problem is one of timing and visibility. While Meta's AI integration into its existing apps (like Instagram and WhatsApp) is working well to improve engagement and ad targeting, the 'next big thing'—the standalone AI products—remains unproven. Investors are looking for the next revenue stream that doesn't rely on the legacy advertising model. Until Meta can demonstrate that AI is a driver of new, high-margin revenue rather than just a way to protect the old, the skepticism will persist.

Meta's Strategic Challenges

High capital expenditure on AI infrastructure
The need to prove AI as a new revenue driver, not just an ad optimizer
Managing shareholder expectations during a massive pivot
The risk of being outpaced by more agile, AI-native competitors

Ultimately, Meta's success depends on its ability to turn its massive user base into an AI-driven ecosystem. It has the scale, but scale can also be a burden. The company must navigate the transition without breaking the very engine that funds its future. It is a high-stakes gamble on the idea that the same social connections that built Facebook can be repurposed into the foundation of an artificial intelligence era.

Key Takeaway

Scale is a double-edged sword when you are trying to pivot an entire empire.

06 — Lenny's Newsletter

The UGC Revolution

How AI is turning raw footage into finished content overnight.

By Lenny Rachitsky · 6 min read

Editor's note: A practical look at how AI is automating the most tedious parts of content production.

Content production has always been a game of attrition. The difference between a professional creator and an amateur isn't just the quality of the idea, but the ability to endure the grueling process of editing. For years, the bottleneck has been the 'middle work': transcribing clips, scanning for the best takes, syncing audio, and assembling a coherent narrative. This is the work that eats up hours of a content team's day and drains the creativity out of the process. But a new workflow is emerging that treats video editing as an asynchronous task rather than a manual one.

From Raw Clips to Finished Video

The process is deceptively simple. A creator records dozens of raw, unpolished clips—often 50 or 60 in total. Instead of sitting down with a timeline, they upload the footage to a custom AI plugin and simply describe the story they want to tell. The AI then takes over the heavy lifting: it transcribes the audio, scans the visual content, identifies the strongest takes based on the user's instructions, and assembles a finished vertical video. The work that used to take an entire day is now completed while the creator sleeps.

The bottleneck of content creation is shifting from the ability to edit to the ability to direct.

This is a fundamental shift in the economics of UGC (User Generated Content). When the cost of editing drops toward zero, the volume of content can explode. This doesn't necessarily mean the quality will drop, but it does mean that the *type* of quality we value will change. We are moving away from highly polished, over-produced content and toward a model where the value lies in the raw idea and the speed of execution. The 'editor' is becoming a 'director'.

The AI Content Workflow

Mass recording of raw, unedited footage
Verbal direction of the narrative arc
Automated transcription and visual scanning
Asynchronous assembly of finished assets

For agencies and content teams, this represents a massive opportunity for scale. The ability to produce high volumes of platform-specific content without a proportional increase in headcount is the holy grail of digital marketing. The challenge will be maintaining a distinct brand voice when the assembly is handled by an algorithm. The creative tension will be between the efficiency of the machine and the intentionality of the human director.

Key Takeaway

The future of content is not about how well you can edit, but how well you can direct.