Think about you’re the visitor on some form of frenetic, software-engineering-themed recreation present. The host is consistently flipping over new playing cards with questions that it’s important to reply as quick as potential:
Is that this adjustment to the database schema proper?
Do these bits of knowledge look believable?
Do these 5 paragraphs of textual content describe an precise collection of handbook assessments that occurred?
Does this prompt structure move the odor check?
Is that this implementation higher than the present code? Or this one? Or this one?
Working in 2026 feels a bit like this. When frontier AI fashions can do a lot of the duties in your queue, essentially the most environment friendly method to work is usually spinning off duties for an AI agent and regularly context-switching between the outcomes. This isn’t fairly senseless — in actual fact, it requires numerous talent to skim the AI response and quickly determine what to do with it — nevertheless it actually entails much less time for gradual, cautious reflection.
Why not decelerate?
Why does it need to be frenetic? Why not simply decelerate? I suppose you may, however I don’t suggest it. It’s simply such a depressing expertise to spend your day close-reading LLM output: fastidiously chewing and savoring every morsel of slop. It’s far much less disagreeable to skim by means of rapidly and select the helpful nuggets of content material.
Couldn’t you merely do extra of the work by hand? It’s sadly true that tech is high-pressure nowadays. When you’ve obtained the time and area to work extra slowly, that’s nice! However when your organization provides you a “resolve this process ten instances extra rapidly” button, you might be closely incentivized to make use of it as a lot as potential, or danger being outcompeted by your friends.
I typically fear that working with LLMs is making me dumber. Not within the “actually melting your mind” sense that some papers suggest, however within the sense that it’s biasing me in the direction of the fast “skimming and judging” elements of my psychological toolkit and away from the gradual “hammock time” wanted for deep thought and actual creativity. I don’t wish to attribute this shift completely to LLMs, because the post-2010s tech business has develop into extra frenetic for broader financial causes. However both method, it’s obtained me questioning how I can hold pondering slowly.
To maintain on pondering, learn and write
The principle factor that’s labored for me is to jot down extra. Particularly, I imply writing in my very own phrases. Writing with an LLM doesn’t work for this in any respect, even if you happen to’re going to some effort to iterate on the content material and description the belongings you wish to say. Why? Having to place the phrases collectively your self forces you to articulate your ideas. In a really actual sense, it forces you to suppose.
When you have got an thought in your head for one thing to jot down, you don’t actually have an thought. What you have got is a form of directional sense of the place an thought is perhaps, or a fraction of the form of factor which may ultimately develop into an thought. You assemble the concept itself whereas writing. By the way, because of this I don’t actually agree with “concepts are straightforward, execution is all the pieces”: most “concepts” will not be actually even concepts.
The opposite factor I like to recommend is to learn precise books. Books — notably dense non-fiction books — are the antithesis of AI slop. The slower you’ll be able to learn them, the higher. I’ve been studying increasingly more non-fiction in the previous few years, and I don’t suppose it’s a coincidence. I feel my mind is of course craving information-dense content material, in the identical method that sodium-deficient folks begin to crave salt.
Actually, I’ve been combining the 2 approaches: studying a e-book after which writing about it. This course of is strictly what I’ve been craving since I began programming with LLMs. I get to fastidiously learn a e-book, suppose onerous about it, typically go and skim one other e-book or two on the identical matter, then sit and attempt to articulate what I’ve discovered. It’s nice! I can really feel elements of my mind stretching once more.
Don’t lose the behavior
It was fairly good once I obtained paid to make use of these elements of my mind all day. Sadly, I feel these instances are coming to an finish. There’ll all the time be room for some quantity of cautious, gradual reflection in software program engineering, however (for a minimum of a short time) we’ll be anticipated to be quickly switching between LLM outputs. We might have to seek out methods outdoors of labor to proceed the behavior of pondering slowly.
Even simply when it comes to work, I feel dropping that behavior completely can be a giant mistake. There are nonetheless loads of odd issues which can be too onerous for present LLMs to resolve on their very own. The most typical instance I run into is “massive refactor on an advanced codebase”. Present-generation LLMs can do that with out (many) errors, however they’ll’t but do it tastefully. Generally you want to have the ability to suppose an issue by means of completely with your personal mind.
Here is a preview of a associated submit that shares tags with this one.
