The End of Programming as We Know It

There’s a batch of chatter successful the media that bundle developers volition soon suffer their jobs to AI. I don’t bargain it.

It is not the extremity of programming. It is the extremity of programming arsenic we cognize it today. That is not new. The archetypal programmers connected carnal circuits to execute each calculation. They were succeeded by programmers penning instrumentality instructions arsenic binary codification to beryllium input 1 spot astatine a clip by flipping switches connected the beforehand of a computer. Assembly connection programming past enactment an extremity to that. It lets a programmer usage a human-like connection to archer the machine to determination information to locations successful representation and execute calculations connected it. Then, improvement of adjacent higher-level compiled languages similar Fortran, COBOL, and their successors C, C++, and Java meant that astir programmers nary longer wrote assembly code. Instead, they could explicit their wishes to the machine utilizing higher level abstractions.

Betty Jean Jennings and Frances Bilas (right) programme the ENIAC successful 1946. Via the Computer History Museum

Eventually, interpreted languages, which are overmuch easier to debug, became the norm. 

BASIC, 1 of the archetypal of these to deed the large time, was astatine archetypal seen arsenic a toy, but soon proved to beryllium the question of the future. Programming became accessible to kids and store entrepreneurs, not conscionable the backmost bureau priesthood astatine ample companies and authorities agencies.

Consumer operating systems were besides a large portion of the story. In the aboriginal days of the idiosyncratic computer, each machine shaper needed bundle engineers who could constitute low-level drivers that performed the enactment of speechmaking and penning to representation boards, hard disks, and peripherals specified arsenic modems and printers. Windows enactment an extremity to that. It didn’t conscionable win due to the fact that it provided a graphical idiosyncratic interface that made it acold easier for untrained individuals to usage computers. It besides provided what Marc Andreessen, whose institution Netscape was astir to beryllium steamrollered by Microsoft, dismissively (and wrongly) called “just a container of drivers.” That container of drivers, fronted by the Win32 APIs, meant that programmers nary longer needed to constitute low-level codification to power the machine. That occupation was efficaciously encapsulated successful the operating system. Windows and macOS, and for mobile, iOS and Android, mean that today, astir programmers nary longer request to cognize overmuch of what earlier generations of programmers knew.

There were much programmers, not fewer

This was acold from the extremity of programming, though. There were much programmers than ever. Users successful the hundreds of millions consumed the fruits of their creativity. In a classical objection of elasticity of demand, arsenic bundle was easier to create, its terms fell, allowing developers to make solutions that much radical were consenting to wage for.

The web was different “end of programming.” Suddenly, the idiosyncratic interface was made up of human-readable documents, shown successful a browser with links that could successful crook telephone programs connected distant servers. Anyone could physique a elemental “application” with minimal programming skill. “No code” became a buzzword. Soon enough, everyone needed a website. Tools similar WordPress made it imaginable for nonprogrammers to make those websites without coding. Yet arsenic the exertion grew successful capability, palmy websites became much and much complex. There was an expanding separation betwixt “frontend” and “backend” programming. New interpreted programming languages similar Python and JavaScript became dominant. Mobile devices added a new, ubiquitous beforehand end, requiring caller skills. And erstwhile again, the complexity was hidden down frameworks, relation libraries, and APIs that insulated programmers from having to cognize arsenic overmuch astir the debased level functionality that it was indispensable for them to larn lone a fewer years before.

Big data, web services, and unreality computing established a benignant of “internet operating system.” Services similar Apple Pay, Google Pay, and Stripe made it imaginable to bash formerly difficult, high-stakes endeavor tasks similar taking payments with minimal programming expertise. All kinds of heavy and almighty functionality was made disposable via elemental APIs. Yet this detonation of net sites and the web protocols and APIs connecting them ended up creating the request for much programmers.

Programmers were nary longer gathering static bundle artifacts updated each mates of years but continuously developing, integrating, and maintaining long-lived services. Even much importantly, overmuch of the enactment astatine these immense services, similar Google Search, Google Maps, Gmail, Amazon, Facebook, and Twitter, was automated astatine immense scale. Programs were designed and built by humans, not AI, but overmuch of the enactment itself was done by special-purpose predecessors to today’s wide intent AIs. The workers that bash the bulk of the dense lifting astatine these companies are already programs. The quality programmers are their managers. There are present hundreds of thousands of programmers doing this benignant of supervisory work. They are already surviving successful a satellite wherever the occupation is creating and managing integer co-workers.

“Google, Facebook, Amazon, oregon a big of much caller Silicon Valley startups…employ tens of thousands of workers. If you deliberation with a twentieth period mill mindset, those workers walk their days grinding retired products, conscionable similar their concern forebears, lone today, they are producing bundle alternatively than carnal goods. If, instead, you measurement backmost and presumption these companies with a 21st period mindset, you recognize that a ample portion of the enactment of these companies – delivering hunt results, quality and information, societal web presumption updates, and applicable products for acquisition – is done by bundle programs and algorithms. These are the existent workers, and the programmers who make them are their managers.”—Tim O’Reilly, “Managing the Bots That Are Managing the Business,” MIT Sloan Management Review, May 21, 2016

In each of these waves, aged skills became obsolescent—still utile but nary longer essential—and caller ones became the cardinal to success. There are inactive a fewer programmers who constitute compilers, thousands who constitute fashionable JavaScript frameworks and Python libraries, but tens of millions who constitute web and mobile applications and the backend bundle that enables them. Billions of users devour what they produce.

Might this clip beryllium different?

Suddenly, though, it is seemingly imaginable for a nonprogrammer to simply speech to an LLM oregon specialized bundle cause successful plain English (or the quality connection of your choice) and get backmost a utile prototype successful Python (or the programming connection of your choice). There’s adjacent a caller buzzword for this: CHOP, oregon “chat-oriented programming.” The emergence of precocious reasoning models is opening to show AI that tin make adjacent analyzable programs with a high-level punctual explaining the task to beryllium accomplished. As a result, determination are a batch of radical saying “this clip is different,” that AI volition wholly regenerate astir quality programmers, and successful fact, astir cognition workers. They accidental we look a question of pervasive quality unemployment.

I inactive don’t bargain it. When there’s a breakthrough that puts precocious computing powerfulness into the hands of a acold larger radical of people, yes, mean radical tin bash things that were erstwhile the domain of highly trained specialists. But that aforesaid breakthrough besides enables caller kinds of services and request for those services. It creates caller sources of heavy magic that lone a fewer understand.

The magic that’s coming present is the astir almighty yet. And that means that we’re opening a profound play of exploration and creativity, trying to recognize however to marque that magic enactment and to deduce caller advantages from its power. Smart developers who follow the exertion volition beryllium successful request due to the fact that they tin bash truthful overmuch more, focusing connected the higher-level creativity that adds value.

Learning by doing

AI volition not regenerate programmers, but it volition alteration their jobs. Eventually overmuch of what programmers bash contiguous whitethorn beryllium arsenic obsolete (for everyone but embedded strategy programmers) arsenic the aged accomplishment of debugging with an oscilloscope. Master programmer and prescient tech perceiver Steve Yegge observes that it is not inferior and mid-level programmers who volition beryllium replaced but those who cling to the past alternatively than embracing the caller programming tools and paradigms. Those who get oregon invent the caller skills volition beryllium successful precocious demand. Junior developers who maestro the tools of AI volition beryllium capable to outperform elder programmers who don’t. Yegge calls it “The Death of the Stubborn Developer.”

My ideas are shaped not lone by my ain past 40+ years of acquisition successful the machine manufacture and the observations of developers similar Yegge but besides by the enactment of economical historiographer James Bessen, who studied however the archetypal Industrial Revolution played retired successful the textile mills of Lowell, Massachusetts during the aboriginal 1800s. As skilled crafters were replaced by machines operated by “unskilled” labor, quality wages were so depressed. But Bessen noticed thing peculiar by comparing the wage records of workers successful the caller concern mills with those of the erstwhile home-based crafters. It took conscionable astir arsenic agelong for an apprentice craftsman to scope the afloat wages of a skilled journeyman arsenic it did for 1 of the caller entry-level unskilled mill workers to scope afloat wage and productivity. The workers successful some regimes were really skilled workers. But they had different kinds of skills.

There were 2 large reasons, Bessen found, wherefore wages remained level oregon depressed for astir of the archetypal 50 years of the Industrial Revolution earlier taking disconnected and starring to a wide summation of prosperity. The archetypal was that the mill owners hoarded the benefits of the caller productivity alternatively than sharing it with workers. But the 2nd was that the largest productivity gains took decades to get due to the fact that the cognition of however champion to usage the caller exertion wasn’t yet wide dispersed. It took decades for inventors to marque the machines much robust, for those utilizing them to travel up with caller kinds of workflows to marque them much effective, to make caller kinds of products that could beryllium made with them, for a wider scope of businesses to follow the caller technologies, and for workers to get the indispensable skills to instrumentality vantage of them. Workers needed caller skills not lone to usage the machines but to repair them, to amended them, to invent the aboriginal that they implied but had not yet made afloat possible. All of this happens done a process that Bessen calls “learning by doing.”

It’s not capable for a fewer individuals to beryllium up of the curve successful adopting the caller skills. Bessen explains that “what matters to a mill, an industry, and to nine mostly is not however agelong it takes to bid an idiosyncratic individual but what it takes to make a stable, trained workforce” (Learning by Doing, 36). Today, each institution that is going to beryllium touched by this gyration (which is to say, each company) needs to enactment its enarthrosis to the wheel. We request an AI-literate workforce. What is programming, aft all, but the mode that humans get computers to bash our bidding? The information that “programming” is getting person and person to quality language, that our machines tin recognize america alternatively than america having to talk to them successful their autochthonal lingua of 0s and 1s, oregon immoderate specialized programming connection pidgin, should beryllium origin for celebration.

People volition beryllium creating, using, and refining much programs, and caller industries volition beryllium calved to negociate and physique connected what we create. Lessons from past archer america that erstwhile automation makes it cheaper and easier to present products that radical privation oregon need, increases successful request often pb to increases successful employment. It is lone erstwhile request is satisfied that employment begins to fall. We are acold from that constituent erstwhile it comes to programming.

Jevons paradox strikes again! As AI gets much businesslike and accessible, we volition spot its usage skyrocket, turning it into a commodity we conscionable can't get capable of. https://t.co/omEcOPhdIz

— Satya Nadella (@satyanadella) January 27, 2025

Not unsurprisingly, Wharton School prof and AI evangelist Ethan Mollick is besides a instrumentality of Bessen’s work. This is wherefore helium argues truthful compellingly to “always bring AI to the table,” to impact it successful each facet of your job, and to research “the jagged edge” of what works and what doesn’t. It is besides wherefore helium urges companies to usage AI to empower their workers, not to regenerate them. There is truthful overmuch to larn astir however to use the caller technology. Businesses’ champion root of applied R&D is the explorations of the radical you have, arsenic they usage AI to lick their problems and question retired caller opportunities.

What programming is volition change

Sam Schillace, 1 of the lawman CTOs astatine Microsoft, agreed with my analysis. In a caller conversation, helium told me, “We’re successful the mediate of inventing a caller programming paradigm astir AI systems. When we went from the desktop into the net era, everything successful the stack changed, adjacent though each the levels of the stack were the same. We inactive person languages, but they went from compiled to interpreted. We inactive person teams, but they went from waterfall to Agile to CI/CD. We inactive person databases, but they went from ACID to NoSQL. We went from 1 user, 1 app, 1 thread, to multi distributed, whatever. We’re doing the aforesaid happening with AI close now.”

Here are immoderate of the technologies that are being assembled into a caller AI stack. And this doesn’t adjacent see the plethora of AI models, their APIs, and their unreality infrastructure. And it’s already retired of date!

“AI Engineering Landscape,” via Marie-Alice Blete connected GitHub

But the detonation of caller tools, frameworks, and practices is conscionable the opening of however programming is changing. One issue, Schillace noted, is that models don’t person representation the mode humans person memory. Even with ample discourse windows, they conflict to bash what helium calls “metacognition.” As a result, helium sees the request for humans to inactive supply a large woody of the discourse successful which their AI co-developers operate.

Schillace expanded connected this thought successful a caller post. “Large connection models (LLMs) and different AI systems are attempting to automate thought,” helium wrote. “The parallels to the automation of question during the concern gyration are striking. Today, the automation is inactive crude: we’re doing the cognitive equivalent of pumping h2o and hammering—basic tasks similar summarization, signifier recognition, and substance generation. We haven’t yet figured retired however to physique robust engines for this caller root of energy—we’re not adjacent astatine the locomotive signifier of AI yet.”

Even the locomotive signifier was mostly an enlargement of the brute unit humans were capable to bring to carnivore erstwhile moving carnal objects. The indispensable adjacent breakthrough was an summation successful the means of control implicit that power. Schillace asks, “What if accepted bundle engineering isn’t afloat applicable here? What if gathering AI requires fundamentally antithetic practices and power systems? We’re trying to make caller kinds of reasoning (our analog to motion): higher-level, metacognitive, adaptive systems that tin bash much than repetition pre-designed patterns. To usage these effectively, we’ll request to invent wholly caller ways of working, caller disciplines. Just arsenic the challenges of aboriginal steam powerfulness birthed metallurgy, the challenges of AI volition unit the emergence of caller sciences of cognition, reliability, and scalability—fields that don’t yet afloat exist.”

The situation of deploying AI technologies successful business

Bret Taylor, formerly co-CEO of Salesforce, one-time Chief Technology Officer astatine Meta, and agelong ago, person of the squad that created Google Maps, is present the CEO of AI cause developer Sierra, a institution astatine the bosom of processing and deploying AI exertion successful businesses. In a caller conversation, Bret told maine that helium believes that a company’s AI cause volition go its superior integer interface, arsenic important arsenic its website, arsenic important arsenic its mobile app, possibly adjacent much so. A company’s AI cause volition person to encode each of its cardinal concern policies and processes. This is thing that AI whitethorn yet beryllium capable to bash connected its own, but today, Sierra has to delegate each of its customers an engineering squad to assistance with the implementation.

“That past mile of taking a chill level and a clump of your concern processes and manifesting an cause is really beauteous hard to do,” Bret explained. “There’s a caller relation emerging present that we telephone an agent engineer, a bundle developer who looks a small spot similar a frontend web developer. That’s an archetype that’s the astir communal successful software. If you’re a React developer, you tin larn to marque AI agents. What a fantastic mode to reskill and marque your skills relevant.”

Who volition privation to wade done a lawsuit work telephone histrion erstwhile they could beryllium talking to an AI cause that tin really lick their problem? But getting those agents close is going to beryllium a existent challenge. It’s not the programming that’s truthful hard. It’s profoundly knowing the concern processes and reasoning however the caller capableness tin alteration them to instrumentality vantage of the caller capabilities. An cause that simply reproduces existing concern processes volition beryllium arsenic embarrassing arsenic a web leafage oregon mobile app that simply recreates a insubstantial form. (And yes, those bash inactive exist!)

Addy Osmani, the caput of idiosyncratic acquisition for Google Chrome, calls this the 70% problem: “While engineers study being dramatically much productive with AI, the existent bundle we usage regular doesn’t look similar it’s getting noticeably better.” He notes that nonprogrammers moving with AI codification procreation tools tin get retired a large demo oregon lick a elemental problem, but they get stuck connected the past 30% of a analyzable programme due to the fact that they don’t cognize capable to debug the codification and usher the AI to the close solution. Meanwhile:

When you ticker a elder technologist enactment with AI tools similar Cursor oregon Copilot, it looks similar magic. They tin scaffold full features successful minutes, implicit with tests and documentation. But ticker carefully, and you’ll announcement thing crucial: They’re not conscionable accepting what the AI suggests…. They’re applying years of hard-won engineering contented to signifier and constrain the AI’s output. The AI is accelerating their implementation, but their expertise is what keeps the codification maintainable. Junior engineers often miss these important steps. They judge the AI’s output much readily, starring to what I telephone “house of cards code” – it looks implicit but collapses nether real-world pressure.

In this regard, Chip Huyen, the writer of the caller publication AI Engineering, made an illuminating reflection successful an email to me:

When AI agents commencement talking to agents…

…precision successful articulating the occupation correctly gets adjacent much important. An cause arsenic a firm frontend that provides entree to each of a company’s concern processes volition beryllium talking not conscionable to consumers but besides to agents for those consumers and agents for different companies.

That full broadside of the cause equation is acold much speculative. We haven’t yet begun to physique retired the standards for practice betwixt autarkic AI agents! A caller insubstantial connected the request for cause infrastructure notes:

Current tools are mostly insufficient due to the fact that they are not designed to signifier however agents interact with existing institutions (e.g., ineligible and economical systems) oregon actors (e.g., integer work providers, humans, different AI agents). For example, alignment techniques by quality bash not guarantee counterparties that immoderate quality volition beryllium held accountable erstwhile a idiosyncratic instructs an cause to execute an amerciable action. To capable this gap, we suggest the conception of cause infrastructure: method systems and shared protocols outer to agents that are designed to mediate and power their interactions with and impacts connected their environments. Agent infrastructure comprises some caller tools and reconfigurations oregon extensions of existing tools. For example, to facilitate accountability, protocols that necktie users to agents could physique upon existing systems for idiosyncratic authentication, specified arsenic OpenID. Just arsenic the Internet relies connected infrastructure similar HTTPS, we reason that cause infrastructure volition beryllium likewise indispensable to ecosystems of agents. We place 3 functions for cause infrastructure: 1) attributing actions, properties, and different accusation to circumstantial agents, their users, oregon different actors; 2) shaping agents’ interactions; and 3) detecting and remedying harmful actions from agents.

There are immense coordination and plan problems to beryllium solved here. Even the champion AI agents we tin ideate volition not lick analyzable coordination problems similar this without quality direction. There is capable programming needed present to support adjacent AI-assisted programmers engaged for astatine slightest the adjacent decade.

In short, determination is simply a full satellite of caller bundle to beryllium invented, and it won’t beryllium invented by AI unsocial but by quality programmers utilizing AI arsenic a superpower. And those programmers request to get a lot of caller skills.

We are successful the aboriginal days of inventing the future

There is truthful overmuch caller to larn and do. So yes, let’s beryllium bold and presume that AI codevelopers marque programmers 10 times arsenic productive. (Your mileage whitethorn vary, depending connected however anxious your developers are to larn caller skills.) But let’s besides stipulate that erstwhile that happens, the “programmable aboveground area” of a business, of the sciences, of our built infrastructure volition emergence successful parallel. If determination are 20x the fig of opportunities for programming to marque a difference, we’ll inactive request doubly arsenic galore of those caller 10x programmers!

User expectations are besides going to rise. Businesses that simply usage the greater productivity to chopped costs volition suffer retired to companies that put successful harnessing the caller capabilities to physique amended services.

As Simon Willison, a longtime bundle developer who has been astatine the forefront of showing the satellite however programming tin beryllium easier and amended successful the AI era, notes, AI lets him “be much ambitious” with his projects.

Take a acquisition from different tract wherever capabilities exploded: It whitethorn instrumentality arsenic agelong to render a azygous framework of 1 of today’s Marvel superhero movies arsenic it did to render the entirety of the archetypal Pixar movie adjacent though CPU/GPU terms and show person benefited from Moore’s Law. It turns retired that the movie manufacture wasn’t contented to present low-res crude animation faster and much cheaply. The other cycles went into thousands of tiny improvements successful realistic fur, water, clouds, reflections, and galore many much pixels of resolution. The technological betterment resulted successful higher quality, not conscionable cheaper/faster delivery. There are immoderate industries made imaginable by choosing cheaper/faster implicit higher accumulation values (consider the detonation of user-created video online), truthful it won’t beryllium either-or. But prime volition person its spot successful the market. It ever does.

Imagine tens of millions of amateur AI-assisted programmers moving with AI tools similar Replit and Devin oregon endeavor solutions similar those provided by Salesforce, Palantir, oregon Sierra. What is the likelihood that they volition stumble implicit usage cases that volition entreaty to millions? Some of them volition go the entrepreneurs of this adjacent procreation of bundle created successful concern with AI. But galore of their ideas volition beryllium adopted, refined, and scaled by existing nonrecreational developers.

The Journey from Prototype to Production

In the enterprise, AI volition marque it overmuch much imaginable for solutions to beryllium built by those closest to immoderate problem. But the champion of those solutions volition inactive request to question the remainder of the mode connected what Shyam Sankar, the CTO of Palantir, has called “the travel from prototype to production.” Sankar noted that the worth of AI to the endeavor is “in automation, successful endeavor autonomy.” But arsenic helium besides pointed out, “Automation is constricted by borderline cases.” He recalled the lessons of Stanley, the self-driving car that won the DARPA Grand Challenge successful 2005: capable to bash thing singular but requiring different 20 years of improvement to afloat grip the borderline cases of driving successful a city.

“Workflow inactive matters,” Sankar argued, and the occupation of the programmer volition beryllium to recognize what tin beryllium done by accepted software, what tin beryllium done by AI, what inactive needs to beryllium done by people, and however you drawstring things unneurotic to really execute the workflow. He notes that “a toolchain that enables you to seizure feedback and larn the borderline cases to get determination arsenic rapidly arsenic imaginable is the winning instrumentality chain.” In the satellite Sankar envisions, AI is “actually going to liberate developers to determination into the concern overmuch much and beryllium overmuch much levered successful the interaction they deliver.” Meanwhile, the top-tier taxable substance experts volition go programmers with the assistance of AI assistants. It is not programmers who volition beryllium retired of work. It volition beryllium the people—in each occupation role—who don’t go AI-assisted programmers.

This is not the extremity of programming. It is the opening of its latest reinvention.

---

On May 8, O’Reilly Media volition beryllium hosting Coding with AI: The End of Software Development arsenic We Know It—a unrecorded virtual tech league spotlighting however AI is already supercharging developers, boosting productivity, and providing existent worth to their organizations. If you’re successful the trenches gathering tomorrow’s improvement practices contiguous and funny successful speaking astatine the event, we’d emotion to perceive from you by March 12. You tin find much accusation and our telephone for presentations here. Just privation to attend? Register for escaped here.