The Deep Feed

01 — Julian Goldie SEO · Video

The Ghost in the Machine: Gemini’s Leap into Computer Use

Why the transition from 'chatting' to 'doing' changes the nature of work

By Julian Goldie SEO · 12 min read

Editor's note: We are moving past the era of prompting and into the era of delegation.

For the last two years, interacting with AI has felt like talking to a very bright, very fast librarian. You ask a question, it provides an answer. You ask for a summary, it provides a list. You are still the one doing the clicking, the scrolling, and the heavy lifting of moving data from one tab to another. But Google’s latest move with Gemini 3.5 Flash changes the fundamental relationship between human and hardware. By integrating 'computer use' directly into the model, the AI is no longer just a brain in a box; it is a pair of hands. It can see the screen, understand the UI, and execute actions across browsers, phones, and desktops. This isn't just a feature; it is a shift in the architecture of agency.

The Loop of Autonomy

The technical mechanism is deceptively simple, yet its implications are massive. The process operates in a continuous loop: screenshot, analyse, act, repeat. The model takes a visual snapshot of your current workspace, identifies the buttons, text fields, and menus, and then decides on the next logical step. It might click a 'Submit' button or type a string of text into a search bar. Once that action is performed, the app takes a new screenshot and feeds it back to the model. This cycle continues until the goal—whether it is booking a flight or auditing a website—is achieved. Because Gemini 3.5 Flash is built for speed and low cost, it can handle these hundreds of micro-actions without the latency that would make such a process unusable.

You give it a goal, it figures out the rest. No clicking, no typing, no you.

This capability solves the 'integration tax' that has plagued automation for decades. In the past, if you wanted to automate a workflow, you needed APIs. You needed two pieces of software to speak the same language. But most of the world's software doesn't have an API; it has a user interface designed for humans. Computer use bypasses the need for formal integration by using the interface itself. The AI interacts with the world exactly as a human does, meaning it can use any legacy software, any obscure website, and any complex dashboard without needing a single line of custom code to connect them.

From Testing to Research

Practical applications for autonomous agents

Automated QA: Running through sign-up flows to find broken buttons or confusing UX.
Onboarding Audits: Simulating a new user's first day to ensure the experience is seamless.
Deep Research: Navigating multiple niche websites to compile data without manual searching.

The immediate value lies in the boring, repetitive tasks that consume professional bandwidth. An agency owner can point an agent at a client's website and demand a full audit of every link and form. A researcher can task an agent with scouring industry forums for specific sentiment. We are moving toward a world where the primary skill is not 'how to use software,' but 'how to define a goal.' The bottleneck is no longer the execution of the task, but the clarity of the instruction.

However, this autonomy brings a new category of risk. When an agent has the power to click, it has the power to err. A mistake in a chat response is a typo; a mistake in computer use is a wrong purchase, a deleted file, or a sent email. The safety of these systems will depend on how we build the guardrails—not just in the model's logic, but in the environments where they operate. We are handing over the keys to our digital lives, and we need to be sure the driver knows where the brakes are.

Key Takeaway

The future of work is not about learning new software, but about learning how to manage agents that use software for you.

02 — Dwarkesh Podcast

The Training Gap: Why AGI Requires More Than Just Data

The struggle to teach machines how to navigate the messy, unsimulated real world

By Dwarkesh Patel · 15 min read

Editor's note: The path to AGI might be blocked by a simple lack of playable environments.

The prevailing belief in AI labs is that scaling is the answer to everything. If we throw enough compute and enough data at a model, it will eventually develop general intelligence. This logic has held true for language models; the sheer volume of text allowed them to master grammar, logic, and even coding. But as we move from language to action—from predicting the next word to predicting the next click—we hit a wall. The problem isn't just data; it's the lack of 'grindable' environments. To learn effectively through reinforcement learning, an AI needs to fail millions of times in a way that is fast, cheap, and repeatable. You can do this in a chess game or a video game, but you cannot do it in the real world.

The Simulation Problem

Consider the difference between coding and computer use. If you want to train an AI to write better Python, you can create a thousand virtual containers, each with a different coding problem. The AI tries a solution, the compiler tells it if it worked, and it learns. This is a deterministic, replayable, and highly scalable loop. Now, try to do the same for a task like 'organize a corporate event on Amazon.' You cannot spin up a thousand clones of Amazon's checkout flow. You cannot bypass their bot detection, and you certainly cannot spend millions of dollars on fake transactions to train a model. The real world is not a simulator; it is a high-stakes, non-repeatable environment that actively resists being 'grinded'.

Unless you can build a replayable training target for a domain, the models will struggle to make progress.

This creates a massive bottleneck for the development of agents. We can train them to be brilliant mathematicians and poets because those domains are essentially closed loops of logic. But teaching an AI to build a business, win a court case, or navigate a complex political landscape is much harder. These tasks require interacting with a world that doesn't reset when you make a mistake. There is no 'undo' button in a market crash or a legal proceeding. To bridge this gap, researchers are looking at creating high-fidelity clones of common software—Slack, Gmail, etc.—to serve as training grounds. But this is a labor-intensive approach that scales poorly compared to the sheer brute force of current LLM training.

The Context Window as a Solution

Alternative paths to on-the-job learning

In-context learning: Using massive context windows to 'remember' everything that happened in a session.
Synthetic environments: Building high-fidelity software clones for reinforcement learning.
Multimodal pre-training: Exposing models to more video data to understand physical and digital spatiality.

One potential way around the need for constant weight updates is the expansion of the context window. If a model can hold an entire month's worth of interactions in its immediate memory, it doesn't need to 'learn' in the traditional sense; it just needs to 'remember.' This would allow for a form of on-the-job competence that mimics human experience without the architectural complexity of continual learning. If we can fit the equivalent of six months of professional training into a single context window, the distinction between 'trained' and 'experienced' begins to blur. The race is no longer just about more parameters, but about more efficient ways to manage the history of an interaction.

Ultimately, the leap to AGI may depend on our ability to manufacture reality. We need to build digital playgrounds that are indistinguishable from the real thing—places where an AI can fail, learn, and evolve without costing a fortune or breaking the actual internet. Until then, the 'agentic' revolution will likely remain confined to the digital silos we can control.

Key Takeaway

The limit of AI intelligence is not just data, but the availability of safe, repeatable environments in which to practice.

03 — Stratechery

The Designer's Tailwind: Dylan Field on the Future of Creativity

Why the market's fear of AI might be missing the point for professional tools

By Stratechery · 18 min read

Editor's note: Figma's market cap has plummeted, but its CEO sees a different reality.

In the current market narrative, AI is a threat to any company that builds tools for creators. The logic is simple: if an AI can generate a website, a logo, or a user interface from a single prompt, why would anyone pay for a complex design tool like Figma? This fear has had a tangible impact on Figma's valuation. After a failed merger with Adobe and a subsequent IPO that saw its market cap crater, the company is being judged by a metric that many of its leaders find misguided. To the market, Figma is a legacy tool in an era of generative magic. To Dylan Field, it is the foundation upon which that magic will be built.

The Operating System of Design

Field's perspective rests on a fundamental distinction between 'design' and 'creativity.' Generative AI is excellent at the former—the execution of visual styles, the arrangement of elements, and the creation of assets. But professional design is a collaborative, iterative, and highly structured process. Figma succeeded not because it was a better drawing tool, but because it was a collaborative operating system. It leveraged WebGL to bring powerful graphical capabilities into the browser, making design a shared, real-time experience. This collaborative infrastructure is something a text-to-image prompt cannot replace; it is the environment where ideas are debated, refined, and turned into products.

AI gives the company a tailwind, not a headwind.

Field argues that AI will actually increase the demand for tools like Figma. As the cost of generating individual assets drops to near zero, the complexity of managing those assets within a cohesive product increases. The 'Canvas'—Figma's central workspace—is the natural intersection for this. Instead of a designer spending hours manually drawing a button, they can use AI to generate fifty variations, then use Figma's professional-grade tools to select, refine, and integrate the best one into a design system. The AI handles the grunt work, while the professional handles the intent and the architecture.

The Intent Gap

How AI changes the designer's role

From Pixel-Pushing to Curation: Moving from manual drawing to selecting and refining AI outputs.
Rapid Prototyping: Using generative tools to instantly turn ideas into high-fidelity mockups.
Systemic Design: Focusing on the rules and logic of a design system rather than individual components.

The real value of a professional tool in the AI age is its ability to manage 'intent.' A prompt is a blunt instrument; it is difficult to tell an AI to 'make this slightly more professional but keep the playful energy of the previous version.' Professional design tools provide the granular control necessary to steer the AI. They offer the precision that generative models currently lack. For Field, the goal is to integrate AI into the canvas so that the tool becomes more powerful, not less relevant. The designer becomes an orchestrator of intelligent components rather than a manual laborer of pixels.

This shift requires a move away from the idea of software as a static set of features and toward software as an intelligent partner. If Figma can successfully position itself as the interface through which human intent meets AI execution, it will not just survive the generative revolution—it will lead it. The market's current pessimism may well be a failure to recognize that the most valuable part of the design process is not the drawing, but the decision-making.

Key Takeaway

AI will automate the execution of design, but it will increase the value of the tools used to direct that execution.

04 — The Marginalian

The Internal Saboteur: Kafka and the Psychology of Creative Block

Why the most gifted often struggle to live up to their own potential

By Maria Popova · 14 min read

Editor's note: A look at the timeless struggle between talent and the self.

Creative work is a strange, dual-natured pursuit. It is a tool for self-liberation, a way to process the chaos of existence, yet it is also a primary source of psychological torment. For those with significant talent, the gap between their vision and their output can become a site of intense suffering. This is nowhere more evident than in the diaries of Franz Kafka. A man who lived a double life—an insurance salesman by day and a writer by night—Kafka’s journals offer a raw, unvarnished look at the internal saboteur that plagues the gifted. He did not just struggle with writer's block; he struggled with the weight of his own expectations.

The Four Perils of the Gifted

Through Kafka's writings, we can identify the recurring psychological hindrances that prevent talent from reaching fruition. These are not mere inconveniences; they are structural failures of the creative psyche. The first is the sense of inadequacy—the constant, gnawing feeling that one's work is 'bad' or 'insufficient.' Kafka frequently indicted himself, lamenting his failure to write even on days when he had the opportunity. This is not a lack of skill, but a lack of permission to be imperfect. The second is the fragmentation of attention, where the demands of a mundane life—the 'overflowing inbox' of the 20th century—strangle the deep focus required for meaningful creation.

I want to write, with a constant trembling on my forehead.

The third hindrance is time-anxiety, a paradoxical state where the creator feels they have too little time to achieve greatness, yet simultaneously falls into the paralysis of procrastination. Kafka described the 'shameful lowlands of writing,' where the fear of not meeting his own standards led him to avoid the desk entirely. He would spend hours in a state of 'dreamy' idleness, not because he was lazy, but because the prospect of failing was more painful than the act of not trying. Finally, there is the physical exhaustion that accompanies the mental strain—the way the body eventually rebels against the intensity of the creative impulse.

The Paradox of Procrastination

Understanding the mechanics of creative block

Time-Anxiety: The feeling of scarcity that leads to frantic, low-quality work.
Procrastination as Defense: Avoiding the task to avoid the judgment of the self.
The Perfectionism Trap: Refusing to begin because the first draft cannot match the vision.

Procrastination is often misunderstood as a failure of discipline. In the context of highly gifted individuals, it is more accurately described as a failure of emotional regulation. When the stakes of a task are tied to one's sense of self-worth, the brain treats the task as a threat. Avoiding the work is a survival mechanism to protect the ego from the possibility of being 'not enough.' Kafka recognized this, noting how his fear of writing would manifest in 'unusable, dry, broken-off' sentences. He was fighting a war against his own nervous system.

To overcome these hindrances, one must accept a certain level of 'trembling.' The goal is not to eliminate the anxiety or the doubt, but to build a discipline that functions in spite of them. Kafka's insistence on the diary was an attempt to create a metronome of discipline—a way to hold fast to the creative impulse even when no salvation seemed possible. The lesson for the modern professional is that the block is not an obstacle to the work; the block is a part of the work. The mastery lies in the ability to sit with the discomfort and produce something anyway.

Key Takeaway

Creative block is often not a lack of ability, but a defensive reaction to the fear of being imperfect.

05 — Not Boring

Optimism in the Age of Friction

From banking automation to the end of the common cold

By Packy McCormick · 10 min read

Editor's note: A weekly roundup of the technologies making life more efficient.

In a world often defined by its frictions—regulatory hurdles, biological limitations, and economic inefficiencies—there is a quiet, persistent movement toward a more seamless existence. This isn't the hollow optimism of a motivational speaker, but a technical optimism grounded in the deployment of new tools. We see it in the way fintech is moving from static dashboards to proactive command centers, and in the way philanthropic capital is being directed toward the most stubborn biological problems. The common thread is the reduction of the 'tax' we pay on being alive.

The Death of the Dashboard

Traditional banking has always been a chore of manual oversight. You log in, you export a CSV, you categorize a transaction, you check a forecast. It is a series of micro-frictions that aggregate into a significant cognitive load. Companies like Mercury are attempting to solve this by treating finance as a conversational command rather than a data entry task. By integrating AI directly into the banking interface, the user can move from 'viewing data' to 'executing intent.' You don't look at a dashboard to see if you can afford a new hire; you ask the system, and it provides the answer with full context. This is the transition from software as a tool to software as an agent.

No dashboards. No exports. No friction. Just total command.

Parallel to this digital streamlining is a massive effort to solve a biological friction: the common cold. While we treat respiratory infections as a minor nuisance, the economic and human cost is staggering. We spend roughly 5% of our lives sick, and the global productivity loss runs into the hundreds of billions. Stripe’s new initiative, Intercept, represents a new model of corporate philanthropy. Instead of funding broad, nebulous causes, they are making a concentrated, technical swing at a specific, underfunded problem. By combining broad-spectrum preventatives with air-cleaning technologies, they are treating the cold not as an inevitability, but as a technical challenge to be solved.

The Housing Bottleneck

Barriers to affordable living

Regulatory Complexity: Federal and local rules that make building slow and expensive.
Institutional Investment: Large-scale buyers driving up prices for individual homeowners.
NIMBYism: Local opposition to new density that keeps supply low.

Even in the political sphere, the fight against friction is evident in the bipartisan push for the 21st Century ROAD to Housing Act. The logic is simple: if you want to lower costs, you have to build more. The bill aims to loosen the regulations that make construction a bureaucratic nightmare and to set limits on the role of institutional investors in the market. While political maneuvering often threatens to stall such progress, the underlying consensus is growing. Whether through technological innovation in construction or regulatory reform in Washington, the goal is to remove the barriers that prevent people from securing a place to live.

These disparate developments—fintech, biotech, and housing policy—all point toward a single direction. We are learning how to use our intelligence and our capital to chip away at the inefficiencies that define the human experience. It is a slow process, often interrupted by political gridlock or biological setbacks, but the trajectory is clear. The future belongs to those who can turn friction into flow.

Key Takeaway

Progress is measured by the steady reduction of the unnecessary burdens—biological, financial, and regulatory—that slow us down.