Why AI Products Always Fall Short: The Key Is Deletion, Not Creation

A Veteran's Diagnosis: It's Not a Capability Problem, It's a Density Problem

Paul Bakaus may not be a household name, but his work is nearly ubiquitous. As the creator of jQuery UI, the code he wrote reportedly still runs on approximately 6% of all internet sites. jQuery UI is a collection of user interface interaction components built on top of the jQuery library, including drag-and-drop, date pickers, dialog boxes, and other common UI elements. It was born around 2007, during the Web 2.0 era when demand for web interactivity was exploding. Even today, when modern frameworks like React and Vue dominate, a massive number of legacy systems and WordPress plugins still use it—explaining its remarkable coverage. Bakaus sold a game engine startup to Zynga, spent nearly a decade at Google, and is now an independent founder who has secured investment from a16z—Andreessen Horowitz, one of Silicon Valley's most influential venture capital firms, known for betting on platform-level technological shifts. Their investments are often seen as a signal that a direction has gained industry validation. Bakaus is building an open-source "design skills" project called Impeccable, aimed squarely at a problem every AI practitioner knows all too well—preventing AI coding agents from producing "slop" (low-quality, perfunctory output).

The word "slop" became a high-frequency term in the tech community in 2024, referring specifically to AI-generated content that is low-quality, lacking personality, and mass-produced. Its harm extends beyond degrading individual product quality—it's systematically lowering the baseline content quality of the entire internet. Search engine results are being flooded with AI-generated SEO content, open-source repositories are seeing waves of low-quality AI-generated commits, and product interfaces are converging toward a uniform "AI style." This is essentially a tragedy of the commons: each individual using AI to reduce costs is rational, but the collective effect is value dilution.

In this widely shared interview clip, Bakaus offered a thought-provoking core argument: he wasn't talking about AI's capability boundaries, but about output density. This concept can be analogized to the signal-to-noise ratio in information theory—how much irreplaceable, effective information or functionality is contained per unit of output. In traditional software engineering, similar ideas have long existed: Robert C. Martin emphasized in Clean Code that functions should be as short as possible, with every line justified; Dieter Rams' ten principles of design express the same philosophy with "less is more." His assessment was concise and sharp—most code generated by AI agents is "too much code"; most AI-written articles are "too long"; most AI-generated designs are "too cluttered, too verbose." When output density is low, users must extract value from excessive redundancy, which itself constitutes cognitive burden and maintenance cost.

What's Scarce Isn't Generative Ability, But Judgment

Over the past two years, the industry's attention has been almost entirely focused on "generation"—faster models, longer contexts, stronger multimodal capabilities. But Bakaus identified a widely overlooked truth: the truly scarce human skill right now is not generating anything, but judgment—knowing exactly what to cut.

This is a counterintuitive yet deeply penetrating observation. When the cost of generation approaches zero, producing "more" is no longer value—it may become a burden. The training mechanisms of large language models inherently drive their output toward redundancy: models based on the Transformer architecture generate content by predicting the next token, essentially searching for the most probable sequence within the probability distribution of training data. RLHF (Reinforcement Learning from Human Feedback) further reinforces the model's tendency to be "comprehensive"—because during annotation, more complete and safer responses typically receive higher scores. This training paradigm makes models inclined to cover all potentially relevant information rather than make bold trade-offs. Models lack deep understanding of specific usage contexts and cannot judge "for this particular user, at this particular moment, which information is redundant"—they can only provide statistically optimal average solutions.

Therefore, AI can easily write five hundred lines of code to solve a problem that should take fifty; it can lay out two thousand words to express something that could be said in a paragraph; it can pile on a dozen interactive elements for an interface that should be minimalist. These outputs are often "technically correct" but lack the wisdom of trade-offs.

The Trap of Being "Technically Correct" Yet Mediocre

Bakaus offered a highly practical diagnostic criterion: if you're using AI agents to deliver products, and everything reads as "technically fine," but the whole thing inexplicably feels generic, mediocre, and soulless, then this is the root cause.

He emphasized that this is not a tooling gap, but an editing gap. This distinction is crucial—it touches on a long-standing misconception in software engineering and creative industries: people tend to use technical solutions for problems that are fundamentally about judgment. In software, this parallels Fred Brooks' "No Silver Bullet" argument in The Mythical Man-Month—essential complexity cannot be eliminated through tools. "Editing" here means far more than polishing text; it represents a systematic quality-review capability: at the code level, it's refactoring and simplification; at the design level, it's trade-offs in information hierarchy; at the product level, it's feature prioritization and the courage to cut requirements. This capability requires deep domain experience and empathetic understanding of end-user scenarios.

Many teams, when encountering poor AI output quality, instinctively respond by switching to better models, writing more clever prompts, or integrating more powerful frameworks. But Bakaus argues that the truly missing piece is human editing and deletion—the aesthetic judgment of knowing that "less is more."

Even the Creator Himself Had to Start Over

The most convincing evidence comes from himself. Bakaus admitted that when writing the launch announcement for his own product, the AI-generated first draft forced him to rewrite from scratch. This detail is quite ironic: a person specifically building tools to combat AI "slop" couldn't get satisfactory copy from AI in one shot.

This echoes the classic principle of "Kill your darlings" from literature and design. Hemingway said "all first drafts are garbage"—even human writing requires extensive revision. But the special problem with AI output is this: while human first drafts are rough, they usually contain the author's unique thought patterns and intuitive judgments; AI first drafts, though superficially smooth, lack this internal thinking skeleton. This means editing AI first drafts is often not "polishing" but "rebuilding the skeleton"—you need to inject your own narrative logic and value judgments, which essentially amounts to rewriting. This also explains why simple "polish this" prompts often fail to solve the slop problem.

What this reveals is the fundamental nature of creation. AI's first-version output is essentially an "average"—it synthesizes the median of countless similar expressions from training data, naturally trending toward safe, complete, and comprehensive, yet losing edge and point of view in the process. But good products, good writing, and good design precisely need those edges.

Design Is an Iterative Process That Must Carry a Human Perspective

Bakaus's closing argument elevated the discussion to a more universal level: design is an iterative process, and it must carry a human point of view. Whether it's code, design, or writing, no tool can "one-shot" this.

One-shot in machine learning originally refers to the ability to learn from just one or very few examples; Bakaus extends it here as a critique of AI workflows. Modern design theory—whether IDEO's Design Thinking, Lean Startup's MVP methodology, or Agile development's iterative cycles—all rest on the same premise: good output requires multiple rounds of feedback and correction. The human brain also relies on alternating "divergent-convergent" processes when handling complex design problems: first expanding the possibility space as much as possible, then gradually narrowing to the optimal solution through judgment. AI's one-shot output skips this cognitive process, directly providing an answer that "looks complete" but lacks the deeper understanding of the problem space accumulated through each iteration. Truly effective AI collaborative workflows should simulate this iterative rhythm rather than pursuing a single definitive solution.

What "human perspective" refers to is precisely what cannot be captured by averages—why you deleted this paragraph and kept that one, why you chose this interaction and abandoned that one, who your product is speaking to and what it's deliberately not saying. These trade-offs constitute a product's personality, and personality is exactly what AI's averaged output finds hardest to replicate.

Implications for Creators in the AI Era

From an industry perspective, Bakaus's observation provides a sobering reference point for today's "AI productivity anxiety."

First, generation is no longer the bottleneck—editing is. When everyone can use AI to rapidly produce vast amounts of content and code, the differentiating moat will shift from "can you build it" to "do you dare to delete, and do you know what to delete." This means teams should reallocate their efforts—devoting more energy to review, simplification, and trade-offs rather than purely pursuing output volume.

Second, AI's first draft should be treated as raw material, not a finished product. Whether it's code or copy, treating AI output as a deliverable final draft will almost inevitably produce the "slop" Bakaus describes. A healthier workflow is: AI expands possibilities, humans converge and refine.

Third, aesthetics and judgment will become high-value skills. In economic terms, when a production factor is fully commoditized, complementary scarce factors gain outsized value—this is called the "complements effect." When AI drives the marginal cost of content generation toward zero, the judgment to distinguish "good" from "mediocre" becomes the critical complementary scarce resource. Similar transitions have happened multiple times in history: the spread of printing made editors' and publishers' taste a core value; the democratization of digital photography made curation and photo editing vision more important; the democratization of music production tools increased the premium on A&R (Artists & Repertoire) judgment. The AI era's "editing premium" may far exceed any previous wave of technological democratization, because this time the explosion in generative capability is unprecedented in both speed and scale. This is a clear signal for content creators, designers, and engineers: cultivating the courage to delete and the sharpness of judgment matters more than learning more prompt engineering tricks.

Bakaus is productizing this philosophy through Impeccable, his open-source project—attempting to teach AI coding agents to "write less, write better." Regardless of whether this specific product ultimately succeeds or fails, the direction he's pointing to deserves serious consideration from everyone delivering products with AI:

What makes a product good is often not what you add, but what you dare to remove.

Key Takeaways