Coding expertise is going to collapse from AI reliance

"We spot a early wherever intelligence is simply a inferior for illustration energy aliases h2o and group bargain it from america connected a metre and usage it for immoderate they want to usage it for" - Sam Altman of OpenAI

In my erstwhile article, Agentic Coding is simply a Trap, I discussed the "skilled orchestrator paradox", wherever the skills required to negociate AI agents for coding are the aforesaid ones that tin beryllium diminished done the continued usage of said AI agents. Expertise was mostly the differentiator; the much knowledgeable a developer is, the less likely it is that they mightiness acquisition accomplishment atrophy, arsenic the knowledge has had a chance to ossify aft years of experience.

If you look astir correct now, you'll find the vast majority of those that are seeing the most benefits from these models are those that person had years, if not decades, of acquisition successful the section (which predates AI tooling, of course). And immoderate manufacture seasoned will show you the same: the bedrock of this knowledge comes from doing the work.

Developers who've entered the section astir the clip of LLMs are placed successful a position wherever they don't person the use of longevity, but they are being guided (and sometimes mandated) to accelerate their efforts utilizing coding assistants that require a history of expertise to wield efficaciously and responsibly.

It's an awkward spot to beryllium for that demographic, arsenic it creates a script wherever a novice needs expert-level skills to leverage the devices and support gait successful the industry.

The "Expert Novice"

We're presently sending very mixed signals to group crossed the industry. We're hammering successful that if you're not utilizing AI tools, you will beryllium "left behind" by your peers who are utilizing them. "AI won't switch you, personification utilizing AI will" has been connected repetition since 2023.

And successful the aforesaid breath, it's besides said that the measurement to get the best results from these models is to apply higher-order thinking; "vibe coding" is simply a dormant end; you request to "move up the stack" and create robust specs, designer pinch bully creation patterns, and always reappraisal the outputs diligently truthful you never vessel thing you don't understand.

The skills to do so, however, are a usability of personification who has knowledgeable the clash and challenges complete clip that culminate successful "good taste".

This leads to different situational paradox: If these devices demand expertise, yet the devices tin actively circumvent the clash that cultivates expertise, past what is the way for 1 to become an expert truthful they tin efficaciously usage these tools?

Confidence without Comprehension

One dream is that these models will extremity up accelerating learning arsenic they are utilized for codification generation. Junior developers tin activity pinch the aforesaid gravitas and assurance arsenic manufacture veterans pinch their "personal AI tutor". Knowing syntax is progressively little important, and immoderate knowledge aliases ambiguity gaps are filled by the AI tool. The deeper mechanics of the codification enactment abstracted away, since the developer sits higher successful the stack.

JetBrains, a awesome subordinate successful developer tools, recently completed a study of inferior and novice developers by painstakingly analyzing their individual behaviour successful unrecorded coding sessions, and testing their expertise to study coding pinch AI tooling successful varying degrees of assistance. Their main takeaway was stark and counterintuitive:

"Participants thought it was for illustration having a individual tutor. From the information successful our study ... we observed that they did not, successful fact, usage GenAI devices for illustration a individual tutor. In fact, it was rather the opposite."

The participants that leaned into heavier AI assistance:

• "Often skipped important readying stages, uncovering that because they hadn’t reasoned themselves into this position, Copilot had."
• "Finished pinch an 'illusion of competence' rather than existent understanding."

Counter to that, the participants that mitigated their usage of AI:

• "Succeeded because they had developed 'negative expertise'—which is 'the expertise to disregard incorrect aliases unhelpful GenAI suggestions'—allowing them to attraction connected penning their ain solutions alternatively than being led astray."
• "Were capable to usage GenAI to accelerate, creating codification they already intended to make."

The novice developers who were the astir unrestricted and assured successful their AI usage "had skipped important steps successful the programming problem-solving process, and were now lost."

Perhaps unsurprisingly, the novice developers who performed the champion were the ones that greatly mitigated aliases outright ignored the AI coding assistance.

Inverted Learning

Due to the self-directed quality of LLMs, the much acquisition you have, the much use they provide since you tin accurately steer, audit, and verify the outputs. The little knowledge you have, the much they tin mislead you. Interacting pinch LLMs for learning caller skills takes the style of an "inverted learning" model, a domiciled reversal wherever the student is initially guiding the mentor, the mentor responds, and past the student, again, steers the mentor.

The process is precarious; LLMs are incredibly delicate to the style of the prompt. When you're exploring caller domains, you don't cognize what you don't know, and the malleable and accommodating creation of an LLM tin lead you to judge you cognize much than you really do.

If you're exploring territory that is moreover somewhat unfamiliar, you often don't moreover cognize the questions that you request to inquire that could decently guideline the exemplary to providing the champion answers. It originates to consciousness for illustration a compass that ever points north, wherever you suggest northbound might be.

From the aforesaid JetBrains study, moreover the astir prepared students were derailed by the AI assistance owed to this type of learning model: One subordinate demonstrated bully basal readying and habits, but abruptly "skipped important problem-solving readying stages, jumping straight to coding and was enticed by Copilot into quickly producing code" and had to trust connected the LLM to hole the correction that the LLM introduced successful the first place.

AI models deficiency judgment, empathy, and pedagogical intent, and the solutions provided are not rooted successful experience but alternatively successful patterns successful the training information (LLMs are, astatine their core, incredibly analyzable shape interpolators).

The infinite reply instrumentality is tempting, and known to be addictive. It tin unwind alternatively quickly, particularly for inexperienced developers. Once you get heavy capable into a generated solution, you are often beholden to the AI instrumentality to besides decorativeness the job, circumventing the problem-solving clash that is required for the statement of a intelligence exemplary (and to beryllium fair, elder developers are prone to this phenomenon, arsenic well).

The Friction is simply a Feature

Expertise and mastery don't hap purely done study and dialogue, but done experience, repetition, and proceedings and error; you person to neglect to succeed. If I wanted to study really to cook, I could watch a Master Chef activity and make endless inquiries. After a month, I would beryllium capable to picture the perfectly medium-rare ribeye but ne'er cognize what it's for illustration to navigator one, and I'd almost surely overcook it connected my first attempt.

Coding has endless moments of tracing obscure errors pinch nary log record to help, experiencing the subtle capacity differences of definite methods, aliases having to rewrite an attack erstwhile it's clear it won't going to scale.

This applied clash is straight what builds "developer intuition" (or "taste"). The Germans person a awesome connection for this: Fingerspitzengefühl (fingertip feeling). It’s the musculus representation that triggers erstwhile a developer looks astatine thing and thinks, “yeah...this is astir apt going to origin problems.” By avoiding the mechanics of the struggle, this intuition is ne'er built.

In UPenn's large-scale 2025 study Generative AI without guardrails tin harm learning, they followed 1,000 students utilizing an LLM to study mathematics and recovered students utilized AI arsenic a crutch and ended up performing 17% worse than students pinch conscionable a textbook (and conscionable arsenic pinch the JetBrains study, the students utilizing the AI assistance thought they were excelling).

LLMs don't conscionable person to make code, though.

If leveraged arsenic Socratic sparring partners alternatively of reply generators, studies person shown that "dialogic AI systems tin meaningfully stimulate reflective, captious and independent thinking".

In that aforesaid UPenn study, they besides tested a "Tutor" type by having students inquire for thief and past independently lick the problem. The GPT Tutor group performed an astonishing 127% better successful the AI-assisted believe convention (although, interestingly, they scored astir the aforesaid connected the trial arsenic the textbook group).

This is effective because the exemplary is no longer being utilized arsenic a intends of production, and it shifts the cognitive activity backmost onto the individual. It's erstwhile the clash is still coming that it creates a lasting imprint that leads to expertise.

Anthropic's 2026 study "How AI assistance impacts the statement of coding skills" came to akin conclusions:

For novice workers successful package engineering aliases immoderate different industry, our study tin beryllium viewed arsenic a mini portion of grounds toward the worth of intentional accomplishment improvement pinch AI tools. Cognitive effort—and moreover getting painfully stuck—is apt important for fostering mastery.

There's a definite consciousness of irony here: the astir productive learning that tin hap pinch an AI coding instrumentality is erstwhile it isn't utilized to make overmuch of immoderate codification astatine all.

Pipeline Collapse

If LLMs tin constitute codification and debug code, and agentic workflows tin execute strategy creation from the abundance of patterns successful the training data, past what is the intent of this knowledge successful the first place? Programming will beryllium done wholly successful earthy language, and we tin dispense pinch the request to prosecute pinch the codification because the models proceed to amended and capable successful immoderate knowledge aliases ambiguity gaps. They will debug immoderate issues that originate and negociate immoderate complexity that they introduce.

The trillion-dollar stake that is being made is: this knowledge won't matter, because LLMs will return up the slack and efficaciously go the caller procreation of "developers". It starts springiness disconnected an aire of hubris that drove past no-code movements, and the fever dreams of CEOs, alternatively than the reality connected the ground.

Coding/programming/software is simply a unsocial intersection of logic, math, problem-solving, captious thinking, planning, communication, and creativity. LLMs tin observe patterns astatine a standard that nary quality ever could, but patterns only get you truthful far.

David Cramer, co-founder astatine Sentry (a capacity and correction search platform), put it succinctly in a caller interview:

Will the pipeline collapse, aliases conscionable change?

It really depends connected whether we make the needed displacement to a much pedagogical usage of these systems.

By continuing to attraction connected and beforehand AI coding workflows that prioritize codification procreation supra heavy understanding, we are not cultivating the adjacent procreation of expertise who will inherit the codification that is being created today.

My Approach: Friction First

Joel Spolsky presciently writes (in 2002, nary less) successful his Law of Leaky Abstractions:

Code procreation devices which dress to absurd retired something, for illustration each abstractions, leak. And the only measurement to woody pinch the leaks competently is to study astir really the abstractions activity ... the abstractions prevention america clip working, but they don’t prevention america clip learning.

If a developer wants to study Java, they should astir apt not commencement pinch Spring Boot. If they want to study JavaScript fundamentals, they should not commencement pinch React. If they want to go highly adept astatine CSS, they should not commencement pinch Tailwind. LLMs could beryllium considered the ultimate leaky abstraction.

My proposal present is very akin to my erstwhile prescription.

If a developer wants to go an master successful programming, they should mostly disregard the axenic codification procreation capabilities of these models, and alternatively usage them for interactive documentation, move tutorial generators, and Socratic exercises.

It's not a panacea, of course: Using an AI instrumentality arsenic a tutor carries its ain risks since it is susceptible to the aforesaid hallucinations arsenic immoderate different interactions, and it cannot beryllium relied upon solely arsenic a learning source. If you can't decently audit the accuracy of the generated code, they you can't audit the accuracy of the generated concept. If you usage AI arsenic a mentor, you must still verify its outputs against official documentation, human peers, and actual proceedings and error.

Choosing this slower, much deliberate way is the champion measurement to turn expertise, but I'm alert of really difficult that is erstwhile the surrounding ecosystem is actively moving against it. AI is being mandated (often recklessly) crossed companies, and baked into astir package improvement devices and IDEs arsenic they cater mostly to elder engineers (even pinch immoderate devices for illustration Cursor tucking distant the codification position unless the personification specifically seeks it out). Some companies are moreover forcing developers to only usage AI for all coding tasks, sloppy of acquisition level, and these companies will person to study their ain lessons.

However, for everyone other who is looking to onslaught a equilibrium betwixt heavy learning (no pun) and productivity, location are qualifying questions you tin inquire to guarantee your usage of these devices yields semipermanent benefits.

My AI-assistance checklist:

• If I did not person entree to an AI tool, could I still execute this task?
• Am I utilizing the exemplary to deepen my understanding, or expedite the answer?
• If I had to audit and verify the generated output, could I adequately explicate what was happening?
• If I'm learning a caller concept, have I done due investigation to cognize the correct questions to ask?
• Have I cross-referenced and verified the attack done different methods (reading documentation, modular hunt tools, StackOverflow, Reddit)?
• Is this a genuinely rote task that's been done 100 times before, aliases a task that requires executive decision-making location successful the process?

Even arsenic a developer pinch decades of acquisition nether my belt, I americium still perpetually referring to them passim my regular work, particularly erstwhile I americium attempting to study thing caller (which successful this field, is neverending).

The cardinal is to observe the quality betwixt cognitive debt and cognitive offloading: Cognitive indebtedness is abdicating your judgement and decisions, whereas cognitive offloading is delegating the mechanical aliases tedious.

As the Anthropic study mentioned, getting "painfully stuck" is simply a bully thing. It takes subject and effort to not drift backmost towards conscionable generating answers, which mightiness not moreover beryllium meticulous successful the first place. LLMs didn't abruptly rewrite the fundamentals of how we learn, but they did springiness america a caller measurement to do so.

Intelligence isn't a Commodity

The realignment I dream to spot complete the years is the knowing that skills don't create without progressive participation. You must prosecute directly and continously to acquisition the essential friction that culminates successful expertise (even if it intends moving much slowly).

If we enactment fixated connected lines of codification and tokens burned while the expertise pipeline dries up complete the years, Sam Altman's imagination of trading intelligence backmost to america connected a metre could go reality. Domain knowledge could go very difficult to travel by, and erstwhile 1 sits down to do immoderate type of improvement work, location will beryllium a pang of paralysis if that personification does not person an progressive AI instrumentality subscription astatine their side.