Will AI replace programmers? What the people building AI coding tools say

The Short Version

None of the people in these conversations says programmers are about to disappear, but they disagree sharply about which part of the job stays human, and for how long. Anthropic CEO Dario Amodei expects programming to change faster than any other field. He thinks human work will move toward system design, architecture and user experience until eventually AI does those too [1]. The four Cursor founders believe AI should amplify programmers, not replace them. In their view the human keeps the "micro-decisions" and the taste, and "people's taste and creative ideas will be magnified" [3]. Sam Altman thinks humans will still do "a lot of programming in five to ten years, but I think it'll be in a very different shape, with some of it done entirely in natural language [4]. ThePrimeagen doesn't expect programmers to be replaced anytime soon. His worry is different: that leaning on AI will leave people who never build real skill, so that juniors remain juniors forever" [2]. For anyone deciding whether to learn to code now, all four sources point the same way. Learn it, learn it deeply, and learn to work with the models, because the durable value is in judgment, debugging and design, not in typing code out.

Table of contents

What They Said

Dario Amodei: programming changes first, and the human role keeps moving up

Dario Amodei, who runs Anthropic, expects programming to be the field AI changes fastest [1]. He gives two reasons. It is close to the people building AI, and models can write, run and check code in a closed loop [1]. The closed loop is the heart of his argument. A model writing an essay can't easily test whether the essay is right. A model writing code can run it, see whether it works and try again. That makes programming unusually well suited to machine improvement.

His description of what happens to the human programmer is a sequence, not an endpoint. As AI takes over most of the literal writing of code, comparative advantage pushes human work toward system design, architecture and UX. But he adds that this lasts only until eventually AI does those too [1]. Of everyone in these sources, Dario is the least willing to name a skill that stays human for good. In his account the human edge is a frontier that keeps moving, not a fixed territory.

He backs this with the clearest number in any of the sources. On SWE-bench, a benchmark of real software engineering tasks, the state of the art went from about 3% at the start of the year to about 50% with the new Sonnet 3.5, and he guesses it will reach 90% within a year [1]. He expects computer use (models operating a computer by reading screenshots and choosing where to click) to become more reliable along the same trajectory as coding [1]. That implies coding is the leading edge other abilities will follow.

Dario is also frank about how rough models still are in practice. Discussing how hard it is to steer model behavior, he explains that penalizing verbosity made models lazy at coding. They would write "rest of the code goes here" instead of the actual code [1]. He treats this "whack-a-mole" problem as a small version of future alignment challenges [1]. For a working programmer it is a reminder that the tools still need someone checking the output.

On products, Dario sees huge untapped potential in AI-powered IDEs, but says Anthropic prefers to let a thousand flowers bloom by powering companies like Cursor and Cognition rather than competing with them [1]. So the tools the Cursor founders describe in [3] run partly on the models Dario's company builds.

His advice to young people is not about programming in particular. It is to start playing with the models and to work where the puck is going rather than where everyone already is. He points to mechanistic interpretability, long-horizon tasks, evaluations for dynamic systems and multi-agent work [1]. He also says the quality he values most in hires is open-mindedness. He wasn't a better programmer than his colleagues, he says. He just asked what would happen if a 30-million-parameter network got 50 million, and plotted graphs [1]. In his own career, then, the question he asked mattered more than how well he coded.

On meaning, Dario argues that the process and choices matter more than whether someone else did it first [1]. A person can still find meaning in building something even if an AI could have built it. His bigger worries are economic: the concentration and abuse of power, since AI increases the total power in the world [1]. His colleague Amanda Askell, in the same episode, encourages non-technical people to just try projects, and notes that models now make learning to code easier [1].

The Cursor founders: humans in the driver's seat

Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger built Cursor, an AI-first code editor forked from VS Code. They are the most detailed and the most optimistic of the sources about what programmers will keep [3]. Their core belief is that AI should amplify programmers rather than replace them, removing drudgery so people can move at the speed of their ideas [3].

Why they think control stays human. Michael describes the future they want as one that emphasizes the programmer's speed, agency, and control. He contrasts it with a future where you chat with a bot like an engineering department over Slack [3]. His reasoning is that good engineering is made of countless micro-decisions and trade-offs, and handing them off loses control [3]. This is the clearest answer in any source to what stays uniquely human. It isn't typing, and it isn't even writing the code. It is making the many small judgment calls that no specification fully captures.

What they expect AI to take over. Their product is built around removing predictable work. Arvid explains that Cursor Tab aims to make all the zero-entropy actions, where your intent is already clear but you'd still have to type, happen with one press of Tab [3]. Aman adds that code has lower loss per byte than natural language, meaning many of its characters are highly predictable [3]. Tab predicts more than the next characters. It predicts your next edit, the next place you'll jump, even a change in another file or a terminal command [3]. Internally, the team competed on how many tabs in a row they could get you to press [3]. Arvid frames pressing Tab as "injecting intent," and says programming will become higher- and higher-bandwidth communication with the computer [3].

What they say doesn't work yet. The founders are candid about the limits. They think agents are really cool but not yet super useful for most tasks [3]. Arvid would love to hand off a well-specified bug, such as copy and paste not working in a chat input box, and come back to a verified fix. But most programming is iterative, so they think people mostly want instant results to react to [3]. They say frontier models are strikingly bad at bug finding when simply asked, because real bug detection is rare in the data models are trained on [3]. Models also lack calibration about which bugs matter: the cultural knowledge a staff engineer has from seeing sketchy code take down a server [3]. Even basic mechanics can fail. Frontier models are good at sketching a change but surprisingly bad at producing exact diffs, messing up simple things like counting line numbers in large files [3]. Cursor trains a separate model to apply changes, because simple deterministic matching fails at least 40% of the time [3].

Review becomes the job. As models propose bigger changes, the founders say, humans must review more, and verification becomes the deeper challenge [3]. Aman puts it bluntly: "Code review kind of sucks." He suggests models could highlight the important parts of a diff, gray out the boring ones, flag likely bugs and walk reviewers through changes in logical order [3]. In this picture the programmer spends less time writing and more time checking and judging.

Natural language won't replace code. The team doesn't expect natural language to replace code entirely. Sometimes it is easier to show an example, or even drag things around, than to explain [3]. Michael floats a "fuzzy idea" of letting programmers move up and down levels of abstraction, for example editing pseudocode and having the changes flow down to real code [3].

Tests and formal proof: an internal disagreement. Arvid makes the boldest prediction in the episode: "I think people will just not write tests anymore." Instead, a model would suggest a specification for each function and a reasoning model would prove the code matches it [3]. His co-founders push back. Intent is often hard to specify formally, they say, and side effects and outside dependencies like a payment API are hard to cover [3]. That pushback matters for the human question. If intent can't be written down fully, then someone who understands the intent has to stay involved.

Their message to young people. Asked about young people worried about their future, Michael says, "I actually think this is a really really exciting time to be building software." He notes there is far less boilerplate than a decade ago, and says "people's taste and creative ideas will be magnified" [3]. Arvid gives a concrete example. A migration away from async local storage in Node.js took him and Sualeh five days and could someday take ten minutes, just by showing a few examples. That would let programmers try things first instead of planning everything at the blackboard [3]. Lex Fridman reads from Cursor's manifesto about a hybrid human-AI engineer who will "iterate at the speed of their judgment" [3]. The founders also say "a code editor should just be fun," and that much of the fun comes from speed [3].

The pace of change. The founders' own history shows how quickly they expect the ground to shift. Cursor started because of OpenAI's 2020 scaling-laws papers and early access to GPT-4 at the end of 2022, when the jump in capability made it clear that all of programming would flow through these models [3]. Michael says: "I think the cursor a year from now will need to make the cursor of today look obsolete" [3]. And: "if you stop innovating on the product you will you will lose" [3]. They are also skeptical of benchmarks as a guide to real ability. Real programming involves messy, half-specified instructions, they say, and SWE-bench is contaminated in training data, so models can guess the right file paths without ever seeing the codebase [3]. That is a direct caution about the kind of number Dario cites in [1].

Sam Altman: a lot of human programming, in a different shape

OpenAI CEO Sam Altman's view comes out in a short exchange with Lex Fridman, but the transcript gives his exact words. Asked how much programming humans will be doing five to ten years from now, he answers: "I mean, a lot, but I think it'll be in a very different shape. Like, maybe some people program entirely in natural language" [4]. He compares this to earlier shifts: "no one programs like writing, bite code" and "no one programs the plant cards anymore" (the transcript's rendering of bytecode and punch cards) [4]. Each generation of programmers moved up a layer, and he expects that to happen again.

When Lex pushes on whether this changes the kind of person who becomes a programmer, Altman answers that "the best practitioners of the craft will use multiple tools and they'll do some more natural language when they need to go, you know, write, see for something they'll do that" [4]. His picture is a mixed toolkit, not a single switch from code to English. The best people will use whichever layer fits the task.

Altman offers a useful way to think about the whole question. The better question than how many jobs AI will replace, he says, is what percentage of tasks, on five-second to five-day horizons, AI can do, letting people work at higher levels of abstraction [4]. That framing splits a programmer's job into many tasks of different lengths, some of which AI takes and some it doesn't.

He is also frank about current limits. Asked what impresses him most about GPT-4, he says: "I think it kind of sucks." He explains that "it is our job to live a few years in the future and remember that the tools we have now are going to kind of suck looking backwards at them" [4]. He notes that people say "it helps me code more productively," but the use he finds most promising is as a creative brainstorming partner [4]. On longer, multi-step tasks such as breaking a problem down, searching the internet and writing code, he says: "When that works, which is not very often, it's like very magical" [4]. He expects the gap between GPT-5 and GPT-4 to match the gap between GPT-4 and GPT-3 [4]. So his view that humans keep programming in a new shape comes alongside a belief that the tools will improve a great deal.

Altman also observes that younger people especially use ChatGPT "as like their default start for any kind of knowledge work task," and lists "you can use it to help you write code" among its everyday uses [4].

Lex Fridman's musings in the Altman episode

Before the Altman interview begins, Lex Fridman wonders aloud whether programming as a mass profession will survive. "Maybe programmers, programming is a way of life as a recreation for millions, as a profession for millions. We'll die out and there'll only be a handful of you like the cobalt programmers of today" [4]. He immediately pushes back on his own worry: "Hopefully not, because there's always room for optimizing at every level the compilation from the human language to the AI language to the machine language to the zero zone ones." And: "I think there's a lot of jobs to be had" [4].

He then adds a caveat that goes to the center of the debate: "A lot of really profitable while paying jobs to be had there, but maybe not millions of people are needed. Maybe there'll be millions of people that program with just natural language, with just words" [4]. That splits the question in two. There may be deep, well-paid work for people who understand the whole stack, and a much larger group who "program in plain language. Later, in the exchange with Altman, Lex notes that programming is hard and asks how hard it is to get like that last one percent to close the gap" [4].

ThePrimeagen: keep the craft, and don't let the tool keep it for you

Michael Paulson, known as ThePrimeagen, spent about ten years at Netflix before teaching programming full time to millions of developers. His view is the most skeptical of the group, and it rests on his own experience [2]. He says AI is great for predictable work and weak at cutting-edge or poorly documented areas like Zig [2].

His Copilot story. When he first used GitHub Copilot, he was amazed. Over time, though, he introduced more bugs: "I wasn't co-piloting well. I was autopilotting much better" [2]. The lesson he drew is that reading code engages a different skill than writing it [2]. The Cursor founders make reviewing the central job of the future programmer [3]. Prime is warning that reviewing is not something a person can do well without having built the skill of writing.

Learned helplessness. His biggest worry is people offloading everything, so that "juniors remain juniors forever" [2]. This is his answer to what stays human. The deep understanding that comes from doing hard work yourself can't be skipped, and a tool that removes the struggle may also remove the learning. His own story explains why he holds this view so strongly. He failed pre-calculus twice, got a C on the third try, and then became the top student in calculus through relentless repetition [2]. "There's no substitute for the hours," he says [2]. He dislikes the phrase work smarter, not harder because it assumes you already know what smart looks like, and he proposes work hard, get smart instead [2].

Other risks. He fears AI answers could be quietly shaped by paid training data [2]. He describes a "denial of attention" problem, citing the curl maintainer being flooded with bogus AI-generated security reports [2]. And he worries that prompt injection will make AI-powered products vulnerable in ways nobody can fully reason about [2]. He wants an assistant that cares about the craft the way a good intern does, and says that even with tools like Devin he often ends up rewriting the output himself [2].

Why he doesn't expect replacement. Prime doesn't expect programmers to be replaced anytime soon. He notes that no computing advance so far has reduced jobs, and suspects natural-language programming will evolve into another precise language [2]. That is a direct challenge to the idea, raised by Altman [4] and by Lex [4], that plain English becomes the programming interface. Prime's point is that once you need precision, the language you end up with looks like code again.

He still uses AI. Prime isn't anti-AI. He uses it heavily outside coding, for example prompting it to role-play his Twitch chat and pitting Claude against OpenAI in a tower defense game he built in Zig [2]. And he still wants his own kids to learn to program, expecting entirely new fields to emerge [2].

Lex's counterargument. In that episode Lex argues the other side. AI makes programming more fun and less lonely, speeds up tasks like wiring up the YouTube API, and working with it is a skill worth building, including learning what context the model is missing [2]. Lex says he now uses Cursor himself [2]. The two end on common ground: AI is poor at hard debugging, and deep, hard skills will remain essential for steering it [2]. They also look at a Stack Overflow survey showing that most developers use or plan to use AI, but only about 2.7% highly trust its output [2].

Where They Agree, and Where They Don't

Where they agree

Programming is changing now, and fast. Dario expects programming to change fastest of all fields [1]. The Cursor team believes all of programming would flow through these models [3]. Altman expects programming in "a very different shape" [4]. Even Prime, the skeptic, uses these tools and debates them at length [2]. Nobody argues that the job will stay as it is.

Predictable, repetitive code goes first. Prime says AI is great for "predictable work" [2]. The Cursor founders built Tab to eliminate "zero-entropy" keystrokes and note how predictable code is [3]. Michael celebrates far less boilerplate than a decade ago [3]. Dario says AI will handle most literal code-writing [1]. They all agree the routine typing is what disappears.

Today's tools still fail in important ways. Dario describes models writing "rest of the code goes here" [1]. The Cursor team says models are strikingly bad at bug finding and can't reliably count line numbers [3]. Altman says multi-step work succeeds "not very often" [4]. Prime found Copilot led him to introduce more bugs [2]. Prime and Lex agree AI is poor at hard debugging [2].

Nobody predicts that programmers soon vanish. The Cursor founders want humans in the driver's seat [3]. Altman says "a lot of human programming will remain [4]. Prime says replacement isn't coming anytime soon" [2]. Dario's view is the most open-ended, but he describes human work shifting, not ending, at least for now [1].

Learning to code is still worth it. Michael calls this "a really really exciting time to be building software" [3]. Prime wants his own kids to learn [2]. Amanda Askell says models make learning to code easier [1]. Dario tells young people to start playing with the models [1].

Where they disagree

What stays uniquely human. This is the central split.

• Dario gives the most temporary answer. System design, architecture and UX stay human until eventually AI does those too [1]. He grounds the view in the closed loop of write, run and check [1], which lets AI improve at programming faster than at most things.
• The Cursor founders say judgment, taste and control stay human, because engineering is countless micro-decisions and trade-offs and handing them off "loses control" [3]. Their co-founders' pushback on Arvid's formal-verification idea adds a reason: intent is hard to specify fully [3]. Worth noting: this is partly a stated design goal. They say they are "excited about" such a future [3], which is not quite a prediction that no other future is possible.
• Altman is less specific. Humans keep programming but move up levels of abstraction, and the best practitioners mix natural language with other tools [4]. His tasks-not-jobs framing [4] implies the human share depends on which tasks AI can do at which time horizon.
• Prime locates the human role in deep, hard-won skill and care for the craft. His emphasis is less on which tasks remain and more on whether people will still have the ability to steer AI if they never build that skill [2].

Whether natural language replaces code. Altman says "maybe some people program entirely in natural language" [4], and Lex imagines "millions of people that program with just natural language" [4]. The Cursor team says natural language won't replace code entirely, since sometimes an example or dragging things around works better [3]. Prime goes further and suspects natural-language programming will turn into another precise language [2]. The disagreement comes down to precision. Everyday requests can be vague. Software that has to work exactly can't be.

Whether AI helps or hurts learning. Amanda Askell says models make learning to code easier [1], and Lex says AI makes programming more fun and less lonely [2]. Prime fears learned helplessness and "juniors remain juniors forever" [2]. Both can be true, depending on how the tool is used. Prime's own Copilot story is about using it on autopilot rather than as a co-pilot [2].

How much to trust the numbers. Dario cites SWE-bench going from about 3% to about 50% and guesses 90% within a year [1]. The Cursor team warns that SWE-bench is contaminated in training data and that real programming involves messy, half-specified instructions that benchmarks don't capture [3]. They rely on qualitative vibe checks and private evals [3]. That doesn't mean Dario is wrong about the direction of progress, and the Cursor team clearly shares his sense of speed. But a benchmark score may overstate how well a model does on the vague, real-world work programmers actually do.

How many jobs. Prime says no computing advance so far has reduced jobs [2]. Lex, in the Altman episode, hopes there are "a lot of jobs to be had" but worries "maybe not millions of people are needed" [4]. Dario's main worry is economic: concentration of power as AI increases total power in the world [1]. The sources don't settle this. They don't provide job figures, only reasoning and history.

Tests and verification. Arvid predicts "people will just not write tests anymore," replaced by model-proposed specs and machine proofs [3]. His co-founders doubt it [3]. Prime comes from the other direction: he adopted asserts that crash on impossible states and learned from deterministic simulation testing, which found bugs he never would have caught otherwise [2]. Prime's approach puts more rigor in the human's hands. Arvid's hands it to the machine.

Why It Works

The sources give several mechanisms for why programming changes the way it does, and why some human role survives.

The closed loop. Dario's main reason programming changes fastest is that models can write, run and check code in a closed loop [1]. Code produces its own feedback. Aman makes a related point about training. The type of synthetic data most likely to produce massive gains is generating many outputs that can be easily checked, as with formal math proofs or code that passes tests [3]. Code that can be checked automatically is exactly the kind of work models improve at quickly. Where perfect checkers are hard to build, for open-ended tasks, progress is harder [3].

Predictability. Aman explains that code has lower loss per byte than natural language, meaning many characters are highly predictable [3]. That is why Tab can take over so much typing. It is also why Prime finds AI great for predictable work and weak at cutting-edge or poorly documented areas like Zig [2]. Models do well where there is plenty of similar code to learn from and poorly where there isn't. The Cursor team notes that models already answer questions about VS Code fairly well because its code was in their training data [3].

Scaling. Dario's broader case rests on scaling: bigger networks, more data and more compute keep producing smarter models. He explains that small networks capture simple patterns and larger ones pick up progressively rarer and more complex patterns [1]. He cautions that scaling laws are empirical regularities, not laws of the universe, and names possible derailers such as data limits [1]. The Cursor team adds test-time compute, running a smaller model longer on hard problems, as a way around the looming "data wall [3]. Altman likewise expects a new paradigm for slower thinking," where AI spends more compute on harder problems [4].

Why bug finding and judgment lag. The Cursor founders explain that models are bad at finding bugs because real bug detection is rare in pre-training data, and they lack the "cultural knowledge" of which bugs matter [3]. Their proposed fix, having one model plant plausible bugs and training another to find them [3], shows that this skill has to be built on purpose. It doesn't come for free from reading code.

Why reviewing isn't the same as writing. Prime's experience gives the human-side mechanism. Reading code engages a different skill than writing it, and relying on autocomplete led him to introduce more bugs [2]. His learning story explains where skill comes from. Repetition takes you from crushing difficulty, to okay, to good, to exploratory mastery, like mastering an instrument [2]. He also says he learns most when a project reaches about 5,000 to 10,000 lines of code, where early design choices either weigh you down or free you [2]. That is the kind of design sense Dario says humans will focus on [1], and it comes from building large things yourself.

Why intent resists automation. Several sources point to intent as the hardest thing to hand off. The Cursor co-founders say intent is often hard to specify formally [3]. When intent is unclear, they suggest a model could ask a clarifying question or show options [3]. In the Altman episode, Lex describes how good it feels "when it got you like what you're thinking about" [4]. The person who knows what they want remains the source of the goal.

Where It Applies, and Where It Doesn't

Where AI already carries much of the load:

• Routine, well-documented work. Examples are fetching data and displaying it, wiring up an API, or boilerplate. Prime calls the job of fetching an object from a database and displaying it every day the worst part of programming [2]. Lex found AI sped up wiring up the YouTube API [2].
• Predictable edits and repetitive changes across files, the "zero-entropy" actions Cursor Tab targets [3].
• Migrations and pattern-based changes that can be shown with a few examples, like Arvid's async local storage migration [3].
• Well-specified, self-contained tasks, like the copy-and-paste bug Arvid would hand to an agent [3].

Where humans remain essential, according to the sources:

• Hard debugging. Prime and Lex both say AI is poor at it [2], and the Cursor team says models are strikingly bad at bug finding when simply asked [3].
• Cutting-edge or poorly documented areas, such as the Zig language Prime mentions [2].
• Large codebases and design decisions that shape a system over time [1][2]. The Cursor team notes that even the best MacBook struggles with a big company's codebase for local models [3].
• Judgment about trade-offs and which risks matter, the staff engineer's "cultural knowledge" [3].
• Security. Prime worries about prompt injection and AI-generated bogus bug reports [2], and Dario says computer use opens new risks like prompt injection [1].
• Review and verification of what the AI produces, which grows as AI writes more [3].

Where the picture is still unsettled:

• Whether design and architecture stay human for long. Dario thinks not forever [1]. The Cursor team wants them to [3].
• Agents working on their own over long stretches. The Cursor team says they aren't "super useful" yet for most tasks [3]. Altman says multi-step tasks work "not very often" [4]. Dario expects autonomy of hours to weeks in his vision of powerful AI, which he guesses could come around 2026 or 2027, while stressing that this is not a scientific prediction [1].
• How many programmers the world will need. Nobody in the sources gives figures. Prime points to history [2], and Lex voices the doubt [4].

A caution about who is speaking. Three of these sources come from people who sell AI coding products or the models behind them: Anthropic [1], Cursor [3] and OpenAI [4]. Their views are informed but not neutral. Dario says Anthropic powers Cursor and Cognition [1]. The Cursor founders say "we're making it for ourselves," which reflects their perspective as programmers who build and use the tool [3]. Prime is the only one who is mainly a working programmer and teacher rather than a builder of AI, and he is the most cautious [2]. The survey he and Lex discuss, where only about 2.7% of developers highly trust AI output [2], is a reminder that everyday users are more skeptical than the builders.

What To Do With It

1. Learn to code anyway, and learn it properly. Every source that addresses young people encourages it. Michael Truell calls this "a really really exciting time to be building software" [3], and Prime still wants his kids to learn to program [2]. Amanda Askell notes that models now make learning easier [1].

2. Put in the hours on fundamentals before leaning on autocomplete. Prime's rule is "There's no substitute for the hours" [2]. Work through problems yourself until they become automatic, as he did with calculus. His warning about "juniors remain juniors forever" [2] is aimed at people who skip this.

3. Use AI as a co-pilot, not an autopilot. Read every suggestion critically. Prime found that accepting output too easily led to more bugs [2]. Treat review as its own skill to practice, since the Cursor team expects review to become a bigger part of the job [3].

4. Build something big enough to hurt. Prime learns most when a project hits 5,000 to 10,000 lines, where design choices start to matter [2]. That is where you build the system-design sense Dario says humans will focus on [1].

5. Build things from scratch to understand them. Prime recommends building something like an HTTP server when you want to truly understand a concept [2]. For a new language, he suggests reading the whole language reference first, as Mitchell Hashimoto does, so you have a mental index of what's possible [2]. For absolute beginners he recommends JavaScript, because you can quickly draw and move things on screen [2].

6. Learn to understand code before you change it. Prime tells how he deleted a colleague's odd-looking code in Netflix's Falcor library, only to rewrite it almost line for line: "I am the idiot in the situation" [2]. With AI producing more code, knowing why code was written the way it was matters more.

7. Practice working with the models as a skill. Dario's advice is to start playing with the models [1]. Lex says working with AI is a skill, including learning what context the model is missing [2]. The Cursor team says the bar for relevant context should be high, because more context can slow and confuse models [3].

8. Build debugging and testing habits AI can't give you. Prime adopted asserts that crash the program on impossible states and found bugs he never would have caught otherwise [2]. Arvid labels dangerous code with comments on every line, which helps both humans, who forget, and models, which pay more attention [3].

9. Develop taste and judgment deliberately. If the Cursor founders are right that "people's taste and creative ideas will be magnified" [3], and that engineering is made of "countless micro-decisions" [3], then the skill to build is deciding what should be made and which trade-offs to accept, not just how to type it.

10. Aim where the puck is going, and keep moving. Dario's advice is to work where the field is heading, not where everyone already is [1]. The Cursor founders expect their own product to look "obsolete within a year [3], and Altman says today's tools are going to kind of suck looking backwards at them" [4]. Expect to keep re-learning your tools, and use each new tool to work at a higher level rather than to stop learning.

Sources

1. Dario Amodei: Anthropic CEO on Claude, AGI, the Future of AI, and Humanity · Lex Fridman Podcast
2. The Primeagen: Programming, AI, ADHD, Productivity, Addiction, and God · Lex Fridman Podcast
3. Cursor Team: Future of Programming with AI · Lex Fridman Podcast
4. Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI · Lex Fridman Podcast