The Deep Feed

The greatest threat to the modern creator is not a lack of ideas, but the friction of the blank page. For most, the process of moving from a thought to a published piece is a slog of decision fatigue and structural hesitation. Alex Lieberman, co-founder of Morning Brew, has solved this by treating content creation not as a creative act, but as a systems engineering problem. He has built what he calls a 'Content Machine' inside Claude, a workflow designed to extract his specific insights and package them through a rigorous editorial filter. This isn't about asking a chatbot to 'write a post about business'; it is about building a closed-loop system that functions as a digital extension of his own brain.

The Death of the Prompt

Most people approach AI with a shallow prompt: 'Write a LinkedIn post about X.' The result is invariably 'slop'—that generic, mid-wit prose that smells of silicon and lacks any real bite. Lieberman argues that slop is a human failure, not a model failure. If the input is boring, the output will be boring. His system bypasses the prompt-engineering trap by using an 'Oracle' to find interesting signals in his own life—scanning Slack, Notion, and Gmail to find the actual 'spikes' of insight worth sharing. He then uses an interview panel of AI personas to interrogate him, forcing him to verbalise the specifics that make a story real. The AI doesn't invent the ideas; it merely harvests them from the source.

AI slop is a people problem. If the clay is bad, the sculpture will be bad.

Once the ideas are extracted, the machine moves to the drafting phase, guided by a Markdown file that serves as a digital DNA. This file contains his specific linguistic patterns, his preferred structures, and even his 'self-deprecating confidence.' By feeding this to the model, he ensures the draft is calibrated to his actual register rather than the smoothed-over average of the internet. The machine is not writing *for* him; it is drafting *as* him, based on a strict set of stylistic constraints that prevent the drift into generic AI-speak.

The Writer's Council

The final safeguard is a 'Writer's Council'—a group of AI personas tasked with being hyper-critical. This isn't a polite feedback loop; it is a rigorous scoring mechanism. One persona is even designated as the 'AI slop allergist,' specifically looking for the tell-tale signs of machine-generated fluff. A draft only moves forward if it clears a high aggregate score. This removes the need for constant manual policing and ensures that even when the founder is busy, the standard remains high. It is a way to scale excellence without scaling the headcount.

Ultimately, this approach changes the role of the creator from a writer to an editor-in-chief. The heavy lifting of transcription, structuring, and first-drafting is handled by the machine, leaving the human to focus on the only thing that matters: the quality of the original thought. In an era where content volume is infinite, the only way to win is to ensure that your volume is backed by genuine, un-simulated insight.

For decades, the corporate ladder was built on specialization. You were a great coder, a great designer, or a great marketer. You mastered a narrow slice of the world and became indispensable within it. But as generative AI begins to commoditise specific technical skills, the value of the narrow specialist is eroding. Elizabeth Stone, the CPTO of Netflix, is seeing a shift in the talent requirements of the world's most sophisticated tech organisation. The new premium is being placed on 'systems thinking'—the ability to understand how disparate parts of a complex whole interact and influence one another.

Beyond the Skill Set

In the AI era, knowing how to write a perfect function or design a flawless interface is becoming a baseline requirement, not a competitive advantage. When a model can assist with the execution of these tasks, the bottleneck shifts from 'how do we do this?' to 'what should we be doing, and how does this affect everything else?' A systems thinker at Netflix doesn't just look at a product feature in isolation; they look at how that feature impacts the engineering load, the user experience, the data pipeline, and the long-term business model. They manage the connections, not just the components.

AI fluency is no longer a specialist skill; it is a universal expectation.

This shift requires a fundamental change in how we approach excellence. Stone describes 'excellence as an operating system.' This means that quality isn't a goal you reach at the end of a project; it is a continuous, integrated process that informs every decision. In a world of AI-generated output, the sheer volume of content and code will explode. The danger is a loss of signal—a drowning in mediocre, high-speed execution. To combat this, organisations must build systems that prioritise high-level reasoning and architectural integrity over raw output.

For the individual professional, this means the path to seniority is no longer about getting better at your specific tool. It is about getting better at understanding the business, the technology, and the user. You must move from being a component in the machine to being one of the people who understands how the machine is built and why it runs. The specialists who thrive will be those who use AI to automate their narrow tasks so they can spend their time on the broader, more complex problems of system design.

There is a prevailing myth in the AI industry that open-weights models are fundamentally 'free' because they eliminate the R&D costs for the user. While it is true that you don't have to spend millions training a model like Kimi K3, the economics of serving that model are far from zero. We are entering an era where the distinction between R&D (a fixed cost) and COGS (a variable cost) is the most important metric in the industry. Unlike traditional software, where the cost of serving an additional user is negligible, every single token generated by an AI model carries a direct, measurable cost in electricity and compute.

The Return of Marginal Costs

For the last two decades, the internet was defined by zero marginal costs. Once you built the software, distributing it to a billion people cost almost nothing. AI breaks this rule. If it costs 50 cents in compute to generate $1 of revenue, your business model is fundamentally different from a SaaS company with 5% COGS. This makes the efficiency of inference—the actual running of the model—the primary battlefield. This is why Nvidia's framing of GPUs as 'token factories' is so accurate. The industry is moving away from valuing 'intelligence' as an abstract concept and toward valuing the efficient production of tokens.

Tokens are not a commodity. Intelligence is what is fungible.

However, there is a catch that the 'token factory' metric misses: the reasoning era. In the first wave of LLMs, the goal was simple: predict the next token. In the current wave, models use 'chain-of-thought' reasoning, which involves generating massive amounts of internal tokens to arrive at a single correct answer. This changes the math entirely. A model that is cheaper per token might actually be more expensive in practice if it requires ten times as many tokens to solve the same problem. A model's value is not its price per million tokens, but the cost of the intelligence it delivers.

The Agentic Shift

This complexity is compounded by the rise of AI agents. An agent doesn't just answer a question; it executes a workflow. This might involve hundreds of loops, tool calls, and self-corrections. In an agentic workflow, the efficiency of the model is measured by how many tokens it takes to complete a task, not how many it can spit out per second. We are moving from a world of 'cheap tokens' to a world of 'efficient intelligence.' The winners won't just be the ones with the biggest clusters, but the ones who can deliver the highest quality reasoning with the lowest computational overhead.

Ultimately, the battle between US and Chinese models, or open vs. closed, will be decided by these underlying economics. If a model can provide the same level of intelligence with a fraction of the reasoning tokens, it wins—regardless of whether its weights are open or its R&D was subsidized. The commodity is not the token; the commodity is the result.

Reverse-engineering has always been a high-stakes game of patience and expertise. Whether you were trying to automate a smart home device or understand a proprietary protocol, the barrier to entry was the sheer amount of human time required. You had to write the code, test it, fail, debug, and repeat. For most, the Return on Investment (ROI) simply wasn't there. The effort required to build a custom automation was often greater than the value the automation provided, especially when you considered the inevitable maintenance required when the undocumented API eventually changed or broke.

The Psychological Shift

Coding agents are fundamentally altering this equation. By drastically reducing the cost of writing and testing code, they have lowered the threshold for experimentation. It is no longer a massive undertaking to attempt to automate a device; it is a low-cost curiosity. If the agent fails to reverse-engineer the API, the cost is negligible. This changes the psychological relationship developers have with code. When code is expensive to write, you are precious about it. You fear maintenance. You fear breaking things. When code is cheap, you treat it as disposable.

Since the code is so cheap, the idea of having to maintain it carries way less psychological baggage.

This shift enables a new kind of 'disposable engineering.' We are moving toward a world where we can write highly specific, highly complex scripts for single-use tasks without feeling the need to build a robust, permanent architecture. We can afford to be wrong. We can afford to throw things away and start again. This accelerates the pace of personal automation and hardware hacking, as the friction of 'what if this breaks?' is replaced by the ease of 'I can just rewrite it in five minutes.'