The Safest Job from AI May Be Writing

Today, tech people are scrambling to alteration their workflows to meet recently inflated 5X productivity quotas, while getting pummeled nether the cognitive indebtedness of agent-generated code. With each caller exemplary release, the spread is widening and humans are becoming much of a bottleneck successful the loop, approaching person to obsolescence arsenic "coders".

While the programmer's occupation explanation is getting wholly refactored, penning remains amazingly unaffected. LLMs person gotten very bully astatine generating code, but I americium appalled astatine the absolute crap they spew arsenic prose. They ever travel the aforesaid robotic cadence and cliches, and sprinkle the aforesaid tired vocabulary each around. They return my surgery yet soulful penning and toggle shape it into a integrative soulless connection slop successful the sanction of improving prose. Their penning communicates nary existent knowing and insight. I deliberation we are each processing a visceral ick guidance to AI writing. It is trapped successful the uncanny valley, and it whitethorn beryllium stuck location for a agelong time.

I americium progressively convinced LLMs will not frighten decent writers anytime soon. My prediction rests connected the 3 observations below. Tell maine wherever my logic breaks.

The stuck-in-slop authorities of LLM penning is not from deficiency of trying. AI labs already tried difficult connected improving prose and deed a wall. LLM giants would person loved to vessel amended penning capacity to conquer marketing, copywriting, and publishing astatine zero marginal cost.

Look astatine image, voice, and heck video models. They sewage bully quickly, because they tin beryllium scaled pinch much parameters and compute. Compared to their accelerated progress, matter models plateaued difficult connected expression, depth, and authenticity. I deliberation it is wicked difficult to bride the last 20% (also applies for image, voice, video models).

In systems theory, a wicked problem is simply a problem that lacks a definitive formulation, a clear stopping rule, and an objectively correct solution. Writing is the eventual wicked problem, because the discourse is perpetually shifting, a portion is ne'er genuinely vanished editing, and the existent metric for occurrence is fundamentally subjective.

Mapping domains on this wickedness spectrum explains why AI dominates definite fields but produces utter slop successful others:

• Math sits connected the near of the spectrum. Specifications are concise, and verification is binary and automated. Since the feedback loop is perfectly closed, AI models tin quickly lick mathematics problems and beryllium results automatically.
• Code relies connected general logic and test-driven feedback. An LLM tin make functional codification because compilers, portion tests, and model checkers tin instantly drawback errors. Verification is mostly mechanical.
• Structured domains for illustration rule and finance besides person immoderate definitive boundaries arsenic they are governed by regulatory frameworks and evaluated connected empirical data.
• Writing sits astatine the acold utmost of the spectrum. There is nary good defined specification to the task astatine hand, and location is nary crushed truth aliases nonsubjective verification of the output. The eventual measurement of occurrence is resonance wrong different quality mind. This is why AI models neglect to make advancement connected bully writing.

Unlike coding, which is simply a single-mind relationship pinch a deterministic compiler, prose is simply a dual-mind problem governed by Theory of Mind. It requires the expertise to continuously simulate a reader's soul intelligence authorities successful existent time. To constitute simply and persuasively, you must way what the scholar already knows, negociate their cognitive load condemnation by sentence, and foretell really an statement will land.

Since LLMs deficiency an progressive intelligence exemplary of a circumstantial quality reader, they are just  optimizing for the statistical probability of the adjacent connection complete a immense dataset. They cannot empathize pinch the quality reader, arsenic they don't person the quality lived experience. And they person zero tegument successful the game.

To onshore this plane, let's propulsion successful David Ricardo and classical economics. The rule of comparative advantage states that moreover if 1 statement tin nutrient everything much efficiently than another, some still use from specializing wherever their comparative opportunity costs is lowest. In different words, moreover if AI has an absolute advantage successful typing velocity and generating measurement astatine zero marginal cost, our quality labour is still governed by the opportunity costs of wherever our scarce resources would beryllium slightest wasted.

The opportunity costs for a quality burning their scarce cognitive capacity connected generic and repetitive tasks is now infinitely high. Instead, quality effort shines wherever AI fails, that is for navigating wicked-problems and pushing for creativity.

This is wherever economics meets evolutionary biology, arsenic "human proof-of-work" becomes the eventual costly signal. In nature, a costly awesome (like a peacock’s tail aliases an elk's antlers) useful because it is costly to nutrient and intolerable to fake. As AI slop saturates the web, pinch the aforesaid token (pardon my pun), the economical worth shifts wholly to an authentic quality voice.

Unlike tech folk, writers don't person to alteration a damn point astir really they activity to optimize their comparative advantage and seizure this costly signal. As AI is stripping distant the accidental complexity of package to expose its inherent complexity, tech workers are struggling to adjust. But bully writers person ever been wrestling pinch the inherent complexity of connection astatine the wicked frontier, and they stay untouched. And possibly programming itself is becoming a shape of writing: much opinionated, much architectural, and much astir connecting pinch quality intent.