Disclaimer: I spend most of this essay talking about Claude Fable 5. I have not had the chance to try Fable 5.1 or GPT-6 Astra yet, though I am sure that my point holds.
I am one of the lucky engineers who works at a very AI-forward company. I get nearly unlimited Claude tokens per month, and as such, I never have to spend any time worrying about token budgets, using small models, plans, subagents. I just spend my time talking to the best and brightest: Claude Fable 5.
I am also one of the lucky few junior engineers to be hired into a company in the AI era. This is not a scenario where AI is merely accelerating my existing software dev work, but a situation where my work is shaped by AI. I have never programmed in a production environment without the help of AI. My architecture decisions, the limits of my knowledge, my lack of past work experience… are all filled by AI.
This situation could take us down several tangents. Is my work better or worse because of this? Is the architecture being designed by AI really sound? How necessary am I, the human in the loop? (For the record, I would argue pretty necessary).
But the simple question I want to dig into today is: Is this good for me? Does talking to Fable all day make me smarter or dumber?
THE CASE FOR DUMBER
This is the case I think we’ve heard before. The studies show it. I’m doing “cognitive offloading”. I think this has two clear effects: for one, I literally am learning less, but more deeply, I am also offloading some important logical decisions to another being.
There is a lot of stuff I simply don’t have to learn anymore. I don’t need to know how our testing library works, or how we have type checking configured. I don’t have to remember specific commands to check for specific logs.
But, like, this doesn’t feel so bad. After all, I can just ask Claude to explain it to me. I ultimately end up with the same knowledge right?
Early into this AI thing Ezra Klein said something nice about this. He said that he used to think that learning was like it is in the Matrix. You “jack in” and you download the information, and you’ve read the book. But that’s just not at all how it works. He thinks that the creative insights, the connections you make, come from spending time with the material. In thinking through the problem yourself.
I think that this is certainly true when it comes to research and writing. And mostly true when it comes to software development.
To give a specific example, when it comes to learning a new programming language, I find that the only real way to learn is by doing. This is why CS curriculums are project-based. You cannot learn to code just by researching the syntax and patterns of a language.
But, when I code in a new language with Claude, this is essentially what I am doing. I may complete a project, but I don’t understand the language, not really. And it’s because, even when I am reading all of the code, I’m not sitting with it, I’m not physically typing it, and I’m not debugging like I used to. If I am curious, after Claude writes some code, I’ll sometimes ask: “why did you do this?” and it explains it to me. But “learning” this way is the same as reading a book about Python, not doing any of the exercises, and then claiming to know Python.
In this way it seems pretty clear that I am maybe not even learning anything at all. But there is a deeper case here: I am actually building up mental shortcuts that cut off my own ability to think through problems.
For all kinds of decisions I find myself relying on Claude, I find this even for things that I think I could do myself, but are only just too difficult to do by myself. This one is very tangible, it makes me feel dumber. I feel the short circuit happening, I feel my brain falling back to “just ask Claude”. It feels like a crutch.
Someone will ask me why I made a decision I made, and I have to ask Claude for the explanation. I am certainly not alone in this, and this is becoming endemic in the software industry, with AI being used to build whole projects and make high level architecture decisions. But I think it is potentially especially dangerous for someone early career like me. I am not simply using Claude to elaborate on an idea I had, or a pattern I know, but to literally make the decisions for me.
And perhaps a final way in which Claude is making me dumber, is that it is making me think that Claude is smart. I do find it hard to read Claude-written prose, but the ideas are often clever. Or at least they seem clever. In all honesty I don’t even know anymore. Claude certainly seems smart to me, but sometimes Claude makes basic errors that I wouldn’t make. I know deep down that it’s possible that Claude isn’t really that smart, that I’ve just slowly and unconsciously lowered my standard for intelligence. In some sense, it’s possible I’ve lost my ability to discern between smart and dumb, between good and bad thinking.
In short this sounds damning. I cannot write the languages I use all day. I am letting Claude make logical decisions on my behalf, strengthening my “ask Claude” muscle instead of my decision-making. And, potentially scariest, I’m not even smart enough to tell if Claude is actually smart or not.
THE CASE FOR SMARTER
I think this case probably would have been much harder to make before the release of Claude Fable. If you’re reading this and you haven’t had the chance to work with Fable and you’re thinking “yeah I mean AI is good but it makes mistakes and you have to watch it carefully”. Fable, in my experience, does not really make mistakes.
It does what it is told. Its thinking is clear. It pushes back when it feels something isn’t right. It double checks when something seems suspicious. It looks and asks for context instead of jumping to conclusions. I don’t know how to describe it. Fable is maybe the first AI model I’ve worked with that I feel like has good instincts.
With that said let me take a step back, though it should be clear where I am going. Have you ever talked with someone smarter than you? Someone who thinks faster than you? Someone for whom things that seem complicated to you seem simple to them? Even if you are very smart I hope that you have met someone like that.
I find that when I talk to someone smarter than me, two things happen. First, in the moment, after an initial feeling of dumbness, I find myself rising to the occasion. I would compare this to a team or a player “playing up” against a high level competition in a sport or a video game. When confronted with someone who knows so much more than me, or thinks so much faster, the conversation pulls me up. I feel myself thinking faster, and attempting to match their level. In some sense, talking to someone smarter than me makes me think better, if only briefly. And so then, second, after the conversion I become extremely interested in talking to them. In learning from them, in imitating them, and in learning to think like them. I come out of these conversations both feeling smarter and having learned from the person.
I felt this when I talked to some of my professors in college, or even some of my classmates. I feel this when I talk to senior engineers at my company. Every conversation is rewarding and I learn a little bit of what they know and how they think. If I’m able to challenge them on something it’s because I’m rising to their level just for a moment.
And so, I feel this with Claude Fable. Claude is an engineer who knows every programming language, every pattern, every best practice and every industry standard. In this way, Claude is easily the most knowledgeable being I’ve every spoken to. But Claude also thinks deeply about every problem you hand it and never gets tired of explaining itself. Yes Claude is fallible, yes Claude is unable to express these ideas succinctly, but it is undeniably one of the smartest beings I have ever talked to.
It seems obvious then. Spending so much time talking to and working with “someone” that smart certainly makes me smarter. If I had unlimited time to work with a staff engineer who had unlimited patience for me, I would undoubtedly learn a lot from them. And I would undoubtedly spend more of my time thinking better.
This is a little abstract, I will try to put it in a way that directly counters the “dumber” case.
First, I absolutely do learn when working with Claude. Now, I don’t learn mechanical things like how to write programming languages, and there are certainly gaps in my knowledge, but I am learning industry best practices, system designs, and smart solutions to already-solved problems. This is stuff that I would need to learn with or without Claude, but I am learning much faster working with Claude. This benefit may not be as pronounced in someone who has a lot of industry experience, but for me, I am able to learn to work with systems that would have usually taken me years of working to build familiarity with. In this way, I think that a major reason Claude makes me smarter is that it allows me to work on problems that I would otherwise need a lot more experience to even be able to start on.
Again, imagine that I had unlimited time to work with a staff engineer. The “dumber” case would frame this as a “crutch” I am building, that I am not learning and thinking for myself, but this is greatly oversimplified.
As I discussed at the start of this section, I am not just “learning systems and best practices” when I work with Claude in the same way that I could learn them by reading a book or downloading them into my brain. I am using them in practice with the help of Claude. I am doing higher-level engineering thinking all day long. Yes, Claude is making some decisions, but that is not the same as offloading, because I can learn from the decisions. Because I see the decisions play out in my work.
In a broader sense we can play the counterfactual here: If I worked this past year without Claude, would I be smarter or dumber? I think it is clear that the No-Claude Ron would be better at coding by hand, and have deeper knowledge of some systems, but the Claude Ron would be a better engineer with much more experience and better technical instincts.
CONCLUSION
It is hard to do an apples to apples comparison. If we are living in a world where AI exists and continues to get better, it is obvious to me that the higher-level engineering thinking skills are more valuable, and thus I am “smarter” in the ways that the time demands of me. If AI was never created, then perhaps the no-Claude Ron is “smarter” for that world, if only because he can actually write code.
My opinion on this is that, in the context of my work, Claude is making me smarter. I am building higher leverage skills and working on harder problems. I am no longer doing menial coding and research and debugging, no longer building skills for an era that is already over.
But I even will say that, if we tried to do this apples to apples, in some abstract sense, I still think that Claude is making me smarter overall. As the counterfactual makes clear, there is a tradeoff that happens when you use AI, that is to say I am getting smarter and dumber by using Claude all day. In terms of pure knowledge the difference comes down most to what I am learning, not how much. And in terms of my ability to think, I think that the amount of good thinking I am doing with Claude outweighs the risk of me relying on Claude as a crutch.
Finally, I will soften this by saying, it is hard to know. I don’t believe I have AI psychosis, but it is possible that my standards for intelligence are much lower than they used to be. I am also worried that building this “Claude crutch” has impacts that make me dumber outside of work. Do I think less critically now, and fall back to asking AI for other things in my life? Am I doing this on things where the process is more important, where the higher-level “learnings” are lower leverage? But that is a topic for another essay.
Appendix
Wanted to take a minute to add some caveats / questions that came up while I was writing this that didn’t fit cleanly into this essay. Some of these may turn into follow up essays.
It is possible that Claude is not really making me smarter or dumber. This case would say that Claude is simply letting me not waste a senior engineer’s time asking questions, and lets me “skip” some of the entry-level learning I would have needed to do anyway. This is more a statement about Claude replacing and accelerating a function that already exists (talking to a senior engineer, my own research), rather than changing anything meaningful about how I’m thinking. (I would describe this as making me slightly smarter)
There are good and bad ways of using Claude. I see many people becoming wholly reliant on Claude, and it seems to me that it is likely that they are not smarter because of it. The standard I hold myself to is essentially: If someone else will be reading it, I will make sure I understand it and can defend it myself. This means that any document I write, I write myself in my own words, and that I read every line of code that Claude writes on my behalf. I definitely think that this standard is critical to me learning from Claude, instead of offloading onto Claude.
Related: Claude certainly makes me feel smarter, but the models have a tendency to be sycophants. If it is true I’ve lost some of my ability to distinguish smart from dumb, then this might be biasing me in the wrong direction.
A lot of the “smarter” argument is predicated on me being early career. It is likely the case that there is a point in which Claude will stop making me smarter. If I already had a lot of industry experience, or was substantially smarter than Claude, then I think that the benefits might be smaller. The downsides too, are also likely much smaller. And in my opinion in this case, Claude probably mostly allows high level engineers to prioritize higher-leverage work.
I think that writing is probably one of the skills that should be cultivated without AI. Producing an AI generated piece is certainly not the same mentally as taking the time to think through and write your argument yourself. At least, I am not sure what the higher-leverage skill you are building by having AI do the writing, or even help you with writing. (though this probably depends on the type of writing you are doing). I sympathize with Ezra Klein on not making connections unless you sit with the material, but I am conflicted on using AI for research.
I find myself constantly writing code and working on projects that, in a pre-AI world, I would have no business working on, and need to rely on Claude to maintain that performance. Simultaneously, as mentioned in the essay, I often feel dumb when I have to fall back to asking Claude for something I should know. Am I losing the ability to really stretch my limits and work hard? Am I becoming just plain-old lazy? Am I being dishonest about my abilities?