Construct to Final – O’Reilly

I’ve spent many years instructing folks to code, constructing instruments that assist builders work extra successfully, and championing the concept that programming ought to be accessible to everybody. By means of quick.ai, I’ve helped tens of millions study not simply to make use of AI however to grasp it deeply sufficient to construct issues that matter.

However these days, I’ve been deeply involved. The AI agent revolution guarantees to make everybody extra productive, but what I’m seeing is one thing completely different: builders abandoning the very practices that result in understanding, mastery, and software program that lasts. When CEOs brag about their groups producing 10,000 traces of AI-written code per day, when junior engineers inform me they’re “vibe-coding” their approach by way of issues with out understanding the options, are we racing towards a future the place nobody understands how something works, and competence craters?

I wanted to speak to somebody who embodies the alternative strategy: somebody whose code continues to run the world many years after he created it. That’s why I referred to as Chris Lattner, cofounder and CEO of Modular AI and creator of LLVM, the Clang compiler, the Swift programming language, and the MLIR compiler infrastructure.

Chris and I chatted on Oct 5, 2025, and he kindly let me file the dialog. I’m glad I did, as a result of it turned out to be considerate and galvanizing. Take a look at the video for the complete interview, or learn on for my abstract of what I realized.

Speaking with Chris Lattner

Chris Lattner builds infrastructure that turns into invisible by way of ubiquity.

Twenty-five years in the past, as a PhD scholar, he created LLVM: probably the most elementary system for translating human-written code into directions computer systems can execute. In 2025, LLVM sits on the basis of most main programming languages: the Rust that powers Firefox, the Swift operating in your iPhone, and even Clang, a C++ compiler created by Chris that Google and Apple now use to create their most crucial software program. He describes the Swift programming language he created as “Syntax sugar for LLVM”. At present it powers the complete iPhone/iPad ecosystem.

If you want one thing to final not simply years however many years, to be versatile sufficient that folks you’ll by no means meet can construct stuff you by no means imagined on high of it, you construct it the way in which Chris constructed LLVM, Clang, and Swift.

I first met Chris when he arrived at Google in 2017 to assist them with TensorFlow. As a substitute of simply tweaking it, he did what he at all times does: he rebuilt from first ideas. He created MLIR (consider it as LLVM for contemporary {hardware} and AI), after which left Google to create Mojo: a programming language designed to lastly give AI builders the type of basis that might final.

Chris architects methods that turn into the bedrock others construct on for many years, by being a real craftsman. He cares deeply in regards to the craft of software program growth.

I advised Chris about my considerations, and the pressures I used to be feeling as each a coder and a CEO:

“Everyone else world wide is doing this, ‘AGI is across the nook. When you’re not doing the whole lot with AI, you’re an fool.’ And truthfully, Chris, it does get to me. I query myself… I’m feeling this strain to say, ‘Screw craftsmanship, screw caring.’ We hear VCs say, ‘My founders are telling me they’re getting out 10,000 traces of code a day.’ Are we loopy, Chris? Are we previous males yelling on the clouds, being like, ‘Again in my day, we cared about craftsmanship’? Or what’s happening?”

Chris advised me he shares my considerations:

“Lots of people are saying, ‘My gosh, tomorrow all programmers are going to get replaced by AGI, and due to this fact we’d as effectively surrender and go dwelling. Why are we doing any of this anymore? When you’re studying code or taking pleasure in what you’re constructing, you then’re not doing it proper.’ That is one thing I’m fairly involved about…

However the query of the day is: how do you construct a system that may truly final greater than six months?”

He confirmed me that the reply to that query is timeless, and truly has little or no to do with AI.

Design from First Rules

Chris’s strategy has at all times been to ask elementary questions. “For me, my journey has at all times been about attempting to grasp the basics of what makes one thing work,” he advised me. “And once you do this, you begin to notice that a variety of the present methods are literally not that nice.”

When Chris began LLVM over Christmas break in 2000, he was asking: what does a compiler infrastructure should be, essentially, to help languages that don’t exist but? When he got here into the AI world he was desperate to study the issues I noticed with TensorFlow and different methods. He then zoomed into what AI infrastructure ought to seem like from the bottom up. Chris defined:

“The rationale that these methods had been elementary, scalable, profitable, and didn’t crumble below their very own weight is as a result of the structure of these methods truly labored effectively. They had been well-designed, they had been scalable. The those who labored on them had an engineering tradition that they rallied behind as a result of they needed to make them technically glorious.

Within the case of LLVM, for instance, it was by no means designed to help the Rust programming language or Julia and even Swift. However as a result of it was designed and architected for that, you possibly can construct programming languages, Snowflake might go construct a database optimizer—which is admittedly cool—and an entire bunch of different functions of the expertise got here out of that structure.”

Chris identified that he and I’ve a sure curiosity in widespread: “We prefer to construct issues, and we prefer to construct issues from the basics. We like to grasp them. We prefer to ask questions.” He has discovered (as have I!) that that is vital if you’d like your work to matter, and to final.

After all, constructing issues from the basics doesn’t at all times work. However as Chris stated, “if we’re going to make a mistake, let’s make a brand new mistake.” Doing the identical factor as everybody else in the identical approach as everybody else isn’t prone to do work that issues.

Craftsmanship and Structure

Chris identified that software program engineering isn’t nearly a person churning out code: “Loads of evolving a product isn’t just about getting the outcomes; it’s in regards to the workforce understanding the structure of the code.” And in reality it’s not even nearly understanding, however that he’s searching for one thing rather more than that. “For folks to truly give a rattling. For folks to care about what they’re doing, to be happy with their work.”

I’ve seen that it’s doable for groups that care and construct thoughtfully to realize one thing particular. I identified to him that “software program engineering has at all times been about attempting to get a product that will get higher and higher, and your capability to work on that product will get higher and higher. Issues get simpler and sooner since you’re constructing higher and higher abstractions and higher and higher understandings in your head.”

Chris agreed. He once more burdened the significance of considering long run:

“Basically, with most sorts of software program tasks, the software program lives for greater than six months or a yr. The sorts of issues I work on, and the sorts of methods you prefer to construct, are issues that you simply proceed to evolve. Take a look at the Linux kernel. The Linux kernel has existed for many years with tons of various folks engaged on it. That’s made doable by an architect, Linus, who’s driving consistency, abstractions, and enchancment in numerous completely different instructions. That longevity is made doable by that architectural focus.”

This type of deep work doesn’t simply profit the group, however advantages each particular person too. Chris stated:

“I believe the query is admittedly about progress. It’s about you as an engineer. What are you studying? How are you getting higher? How a lot mastery do you develop? Why is it that you simply’re in a position to clear up issues that different folks can’t?… The those who I see doing rather well of their careers, their lives, and their growth are the folks which are pushing. They’re not complacent. They’re not simply doing what everyone tells them to do. They’re truly asking laborious questions, they usually wish to get higher. So investing in your self, investing in your instruments and strategies, and actually pushing laborious so as to perceive issues at a deeper degree—I believe that’s actually what allows folks to develop and obtain issues that they possibly didn’t suppose had been doable just a few years earlier than.”

That is what I inform my workforce too. The factor I care most about is whether or not they’re at all times bettering at their capability to unravel these issues.

Dogfooding

However caring deeply and considering architecturally isn’t sufficient in the event you’re constructing in a vacuum.

I’m unsure it’s actually doable to create nice software program in the event you’re not utilizing it your self, or working proper subsequent to your customers. When Chris and his workforce had been constructing the Swift language, they needed to construct it in a vacuum of Apple secrecy. He shares:

“The utilizing your personal product piece is admittedly essential. One of many massive issues that brought on the IDE options and plenty of different issues to be an issue with Swift is that we didn’t actually have a consumer. We had been constructing it, however earlier than we launched, we had one take a look at app that was type of ‘dogfooded’ in air quotes, however probably not. We weren’t truly utilizing it in manufacturing in any respect. And by the point it launched, you possibly can inform. The instruments didn’t work, it was sluggish to compile, crashed on a regular basis, numerous lacking options.”

His new Mojo mission is taking a really completely different route:

“With Mojo, we contemplate ourselves to be the primary buyer. We’ve lots of of hundreds of traces of Mojo code, and it’s all open supply… That strategy may be very completely different. It’s a product of expertise, but it surely’s additionally a product of constructing Mojo to unravel our personal issues. We’re studying from the previous, taking finest ideas in.”

The result’s evident. Already at this early stage fashions constructed on Mojo are getting state-of-the-art outcomes. Most of Mojo is written in Mojo. So if one thing isn’t working effectively, they’re the primary ones to note.

We had an analogous purpose at quick.ai with our Solveit platform: we needed to achieve a degree the place most of our workers selected to do most of their work in Solveit, as a result of they most popular it. (Certainly, I’m writing this text in Solveit proper now!) Earlier than we reached that time, I typically needed to power myself to make use of Solveit as a way to expertise first hand the shortcomings of these early variations, in order that I might deeply perceive the problems. Having completed so, I now respect how clean the whole lot works much more!

However this type of deep, experiential understanding is precisely what we threat shedding after we delegate an excessive amount of to AI.

AI, Craftsmanship, and Studying

Chris makes use of AI: “I believe it’s an important software. I really feel like I get a ten to twenty% enchancment—some actually fancy code completion and autocomplete.” However with Chris’ deal with the significance of workmanship and continuous studying and enchancment, I puzzled if heavy AI (and notably agent) use (“vibe coding”) would possibly negatively affect organizations and people.

Chris: If you’re vibe-coding issues, all of a sudden… one other factor I’ve seen is that folks say, ‘Okay, effectively possibly it’ll work.’ It’s nearly like a take a look at. You go off and say, ‘Perhaps the agentic factor will go crank out some code,’ and also you spend all this time ready on it and training it. Then, it doesn’t work.

Jeremy: It’s like a playing machine, proper? Pull the lever once more, attempt once more, simply attempt once more.

Chris: Precisely. And once more, I’m not saying the instruments are ineffective or unhealthy, however once you take a step again and also you have a look at the place it’s including worth and the way, I believe there’s a little bit bit an excessive amount of enthusiasm of, ‘Properly, when AGI occurs, it’s going to unravel the issue. I’m simply ready and seeing… Right here’s one other side of it: the anxiousness piece. I see a variety of junior engineers popping out of faculty, they usually’re very frightened about whether or not they’ll be capable to get a job. Loads of issues are altering, and I don’t actually know what’s going to occur. However to your level earlier, a variety of them say, ’Okay, effectively, I’m simply going to vibe-code the whole lot,’ as a result of that is ‘productiveness’ in air quotes. I believe that’s additionally a major downside.

Jeremy: Looks like a profession killer to me.

Chris: …When you get sucked into, ‘Okay, effectively I want to determine make this factor make me a 10x programmer,’ it could be a path that doesn’t convey you to creating in any respect. It could truly imply that you simply’re throwing away your personal time, as a result of we solely have a lot time to stay on this earth. It might probably find yourself retarding your growth and stopping you from rising and truly getting stuff completed.

At its coronary heart, Chris’s concern is that AI-heavy coding and craftsmanship simply don’t seem like appropriate:

“Software program craftsmanship is the factor that AI code threatens. Not as a result of it’s inconceivable to make use of correctly—once more, I take advantage of it, and I really feel like I’m doing it effectively as a result of I care so much in regards to the high quality of the code. However as a result of it encourages of us to not take the craftsmanship, design, and structure severely. As a substitute, you simply devolve to getting your bug queue to be shallower and making the signs go away. I believe that’s the factor that I discover regarding.”

“What you wish to get to, notably as your profession evolves, is mastery. That’s the way you type of escape the factor that everyone can do and get extra differentiation… The priority I’ve is that this tradition of, ‘Properly, I’m not even going to attempt to perceive what’s happening. I’m simply going to spend some tokens, and possibly it’ll be nice.’”

I requested if he had some particular examples the place he’s seen issues go awry.

“I’ve seen a senior engineer, when a bug will get reported, let the agentic loop rip, go spend some tokens, and possibly it’ll give you a bug repair and create a PR. This PR, nonetheless, was utterly flawed. It made the symptom go away, so it ‘fastened’ the bug in air quotes, but it surely was so flawed that if it had been merged, it might have simply made the product approach worse. You’re changing one bug with an entire bunch of different bugs which are tougher to grasp, and a ton of code that’s simply within the flawed place doing the flawed factor. That’s deeply regarding. The precise concern just isn’t this explicit engineer as a result of, thankfully, they’re a senior engineer and good sufficient to not simply say, ‘Okay, move this take a look at, merge.’ We additionally do code overview, which is an important factor. However the concern I’ve is that this tradition of, ‘Properly, I’m not even going to attempt to perceive what’s happening. I’m simply going to spend some tokens, and possibly it’ll be nice. Now I don’t have to consider it.’ This can be a large concern as a result of a variety of evolving a product isn’t just about getting the outcomes; it’s in regards to the workforce understanding the structure of the code. When you’re delegating data to an AI, and also you’re simply reviewing the code with out excited about what you wish to obtain, I believe that’s very, very regarding.”

Some of us have advised me they suppose that unit checks are a very good place to take a look at utilizing AI extra closely. Chris urges warning, nonetheless:

“AI is admittedly nice at writing unit checks. This is likely one of the issues that no one likes to do. It feels tremendous productive to say, ‘Simply crank out an entire bunch of checks,’ and look, I’ve received all this code, wonderful. However there’s an issue, as a result of unit checks are their very own potential tech debt. The take a look at might not be testing the best factor, or they is likely to be testing a element of the factor moderately than the true thought of the factor… And in the event you’re utilizing mocking, now you get all these tremendous tightly certain implementation particulars in your checks, which make it very troublesome to vary the structure of your product as issues evolve. Exams are identical to the code in your foremost utility—it’s best to take into consideration them. Additionally, numerous checks take a very long time to run, and they also affect your future growth velocity.”

A part of the issue, Chris famous, is that many individuals are utilizing excessive traces of code written as a statistic to help the concept that AI is making a optimistic affect.

“To me, the query just isn’t how do you get probably the most code. I’m not a CEO bragging in regards to the variety of traces of code written by AI; I believe that’s a very ineffective metric. I don’t measure progress primarily based on the variety of traces of code written. In actual fact, I see verbose, redundant, not well-factored code as an enormous legal responsibility… The query is: how productive are folks at getting stuff completed and making the product higher? That is what I care about.”

Underlying all of those considerations is the assumption that AGI is imminent, and due to this fact conventional approaches to software program growth are out of date. Chris has seen this film earlier than. “In 2017, I used to be at Tesla engaged on self-driving vehicles, main the Autopilot software program workforce. I used to be satisfied that in 2020, autonomous vehicles could be in all places and could be solved. It was this determined race to go clear up autonomy… However on the time, no one even knew how laborious that was. However what was within the air was: trillions of {dollars} are at stake, job substitute, reworking transportation… I believe at the moment, precisely the identical factor is going on. It’s not about self-driving, though that’s making progress, just a bit bit much less gloriously and instantly than folks thought. However now it’s about programming.”

Chris thinks that, like all earlier applied sciences, AI progress isn’t truly exponential. “I consider that progress appears like S-curves. Pre-training was an enormous deal. It appeared exponential, but it surely truly S-curved out and received flat as issues went on. I believe that we’ve quite a few piled-up S-curves which are all driving ahead wonderful progress, however I not less than haven’t seen that spark.”

The hazard isn’t simply that folks is likely to be flawed about AGI’s timeline—it’s what occurs to their careers and codebases whereas they’re ready. “Know-how waves trigger huge hype cycles, overdrama, and overselling,” Chris famous. “Whether or not it’s object-oriented programming within the ’80s the place the whole lot’s an object, or the web wave within the 2000s the place the whole lot must be on-line in any other case you’ll be able to’t purchase a shirt or pet food. There’s reality to the expertise, however what finally ends up taking place is issues settle out, and it’s much less dramatic than initially promised. The query is, when issues settle out, the place do you as a programmer stand? Have you ever misplaced years of your personal growth since you’ve been spending it the flawed approach?”

Chris is cautious to make clear that he’s not anti-AI—removed from it. “I’m a maximalist. I need AI in all of our lives,” he advised me. “Nonetheless, the factor I don’t like is the folks which are making selections as if AGI or ASI had been right here tomorrow… Being paranoid, being anxious, being afraid of residing your life and of constructing a greater world looks like a really foolish and never very pragmatic factor to do.”

Software program Craftsmanship with AI

Chris sees the important thing as understanding the distinction between utilizing AI as a crutch versus utilizing it as a software that enhances your craftsmanship. He finds AI notably helpful for exploration and studying:

“It’s wonderful for studying a codebase you’re not conversant in, so it’s nice for discovery. The automation options of AI are tremendous essential. Getting us out of writing boilerplate, getting us out of memorizing APIs, getting us out of trying up that factor from Stack Overflow; I believe that is actually profound. This can be a good use. The factor that I get involved about is in the event you go as far as to not care about what you’re trying up on Stack Overflow and why it really works that approach and never studying from it.”

One precept Chris and I share is the vital significance of tight iteration loops. For Chris, engaged on methods programming, this implies “edit the code, compile, run it, get a take a look at that fails, after which debug it and iterate on that loop… Operating checks ought to take lower than a minute, ideally lower than 30 seconds.” He advised me that when engaged on Mojo, one of many first priorities was “constructing VS Code help early as a result of with out instruments that allow you to create fast iterations, your whole work goes to be slower, extra annoying, and extra flawed.”

My background is completely different—I’m a fan of the Smalltalk, Lisp, and APL custom the place you have got a stay workspace and each line of code manipulates objects in that atmosphere. When Chris and I first labored collectively on Swift for TensorFlow, the very first thing I advised him was “I’m going to wish a pocket book.” Inside per week, he had constructed me full Swift help for Jupyter. I might kind one thing, see the end result instantly, and watch my information remodel step-by-step by way of the method. That is the Brett Victor “Inventing on Precept” fashion of being near what you’re crafting.

If you wish to keep craftsmanship whereas utilizing AI, you want tight iteration loops so you’ll be able to see what’s taking place. You want a stay workspace the place you (and the AI) are manipulating precise state, not simply writing textual content information.

At quick.ai, we’ve been working to place this philosophy into observe with our Solveit platform. We found a key precept: the AI ought to be capable to see precisely what the human sees, and the human ought to be capable to see precisely what the AI sees always. No separate instruction information, no context home windows that don’t match your precise workspace—the AI is correct there with you, supporting you as you’re employed.

This creates what I consider as “a 3rd participant on this dialogue”—beforehand I had a dialog with my pc by way of a REPL, typing instructions and seeing outcomes. Now the AI is in that dialog too, in a position to see my code, my information, my outputs, and my thought course of as I work by way of issues. After I ask “does this align with what we mentioned earlier” or “have we dealt with this edge case,” the AI doesn’t want me to copy-paste context—it’s already there.

Considered one of our workforce members, Nate, constructed one thing referred to as ShellSage that demonstrates this fantastically. He realized that tmux already exhibits the whole lot that’s occurred in your shell session, so he simply added a command that talks to an LLM. That’s it—about 100 traces of code. The LLM can see all of your earlier instructions, questions, and output. By the subsequent day, all of us had been utilizing it continually. One other workforce member, Eric, constructed our Discord Buddy bot utilizing this identical strategy—he didn’t write code in an editor and deploy it. He typed instructions separately in a stay image desk, manipulating state instantly. When it labored, he wrapped these steps into features. No deployment, no construct course of—simply iterative refinement of a operating system.

Eric Ries has been writing his new ebook in Solveit and the AI can see precisely what he writes. He asks questions like “does this paragraph align with the mission we acknowledged earlier?” or “have we mentioned this case research earlier than?” or “are you able to verify my editor’s notes for feedback on this?” The AI doesn’t want particular directions or context administration—it’s within the trenches with him, watching the work unfold. (I’m writing this text in Solveit proper now, for a similar causes.)

I requested Chris about how he thinks in regards to the strategy we’re taking with Solveit: “as a substitute of bringing in a junior engineer that may simply crank out code, you’re bringing in a senior skilled, a senior engineer, an advisor—any person that may truly provide help to make higher code and educate you issues.”

How Do We Do One thing Significant?

Chris and I each see a bifurcation coming. “It appears like we’re going to have a bifurcation of abilities,” I advised him, “as a result of individuals who use AI the flawed approach are going to worsen and worse. And the individuals who use it to study extra and study sooner are going to outpace the velocity of development of AI capabilities as a result of they’re human with the good thing about that… There’s going to be this group of those who have realized helplessness and this possibly smaller group of individuals that everyone’s like, ‘How does this particular person know the whole lot? They’re so good.’”

The ideas that allowed LLVM to final 25 years—structure; understanding; craftsmanship—haven’t modified. “The query is, when issues settle out, the place do you as a programmer stand?” Chris requested. “Have you ever misplaced years of your personal growth since you’ve been spending it the flawed approach? And now all of a sudden everyone else is far additional forward of you by way of with the ability to create productive worth for the world.”

His recommendation is evident, particularly for these simply beginning out: “If I had been popping out of faculty, my recommendation could be don’t pursue that path. Significantly if everyone is zigging, it’s time to zag. What you wish to get to, notably as your profession evolves, is mastery. So that you may be the senior engineer. So you’ll be able to truly perceive issues to a depth that different folks don’t. That’s the way you escape the factor that everyone can do and get extra differentiation.”

The hype will settle. The instruments will enhance. However the query Chris poses stays: “How can we truly add worth to the world? How can we do one thing significant? How can we transfer the world ahead?” For each of us, the reply includes caring deeply about our craft, understanding what we’re constructing, and utilizing AI not as a substitute for considering however as a software to suppose extra successfully. If the purpose is to construct issues that final, you’re not going to have the ability to outsource that to AI. You’ll want to speculate deeply in your self.