The Human Is the Loop
brentfitzgerald.com
brentfitzgerald.com
The problem with writing, and something I am noticing lately, is that AI robs your ability to create prose in your style. Even if you use AI just a little bit to touch up your writing - you start being influenced - And over time it adds up - you start writing like AI just a little bit every now and then. It seems to be robbing me of my ability to create original prose.
Strategic thought is the other thing - like if I am deciding what we, as a company should focus on long term, it is easy for LLM's to give you a wrong direction. I think this is more because Strategic directions require a ton of context - and it is almost impossible for you to feed the ai EVERYTHING. There might be some things which are easy to miss out, which might make all the difference, and the LLM would assume that the 'requirements' are complete and would go to town with it. I suspect, this is largely because the current SOTA LLM's are being optimised for code and there isnt enough data in the training set for them to learn long term thought.
Coding is I think fine for me. But even there, I dont let the agent use memory at all. I create the context - and it runs from that only. The biggest unlock here is that I can now program in any language. And it frees up sometime to think in systems and come up with more robust architectures.
I feel like LLMs can be very helpful to double-check a bunch of stuff here, and evaluate alternative-paths and so on, but the "direction" really has to come from you, then you can use the LLM to double-check and what not.
But if you let the LLM set the "direction", then all bets are off, as they don't have a single care about anything in the world, so they'll just do whatever essentially.
I use it very sparingly for stuff like this - and even when I do - I discount its outputs heavily and maybe just use it let me see some blind spots and sanity testing.
IMHO the fix is to read more books. Obviously I mean human-written books, and specifically with prose that you appreciate.
It's quite different than just reading a lot of text from other authors and passively absorbing diction and style. Instead, it means getting detailed reinforcement signals in the context of your own active writing effort.
This can start an ironic reinforcement learning pattern where the writer gets tuned to write in the LLM's "corrected" form. Imagine a traditional professional writer, in a long-term partnership with a particular copy editor who gates their output.
After a hiatus of a few years, when the internet had me, I've gone back to prioritising reading books once more. It's not only great for my writing, it's also just good for the soul. And it's the absolute antithesis of the infinite scroll.
The great poems, plays, novels, stories teach us how to go on living, even when submerged under forty fathoms of bother and distress. If you live ninety years you will be a battered survivor. Your own mistakes, accidents, failures at otherness beat you down. Rise up at dawn and read something that matters as soon as you can."
-- Harold Bloom
I only ever have 2-4 opencode/claude tabs open at once. And I use them for consistent tasks. There’s no way I’d be able to keep track otherwise. I have an “email reviewer” terminal, and a “doc maker” terminal, and 1-2 “coding” terminals.
But I don’t start a half dozen things simultaneously (which is ironic, because I have ADHD). I don’t feel the need to maximize my efficiency with AI. I also don’t lament the ideas / things I start but don’t finish; that has been my life, lol - see the aforementioned ADHD. So why should I start caring about my half-finished projects now, just because I sunk some tokens into them?
I spent an entire day the other day trying to use an LLM to help me automate something manual and tedious, and it was a failure. Did nothing but waste time in the end. Perhaps it’s those experiences (of which I’ve had plenty) that help me realize the limitations?
I don’t feel like it should make me more productive; it does, because I can offload random tasks that I do not have to be the one to do (“find the function where this happens” or “find all the laundromats in the cities I’m going to”) but I use it for research like that, and not so much “do my taxes for me.”
That’s why I paid a guy to do my taxes. I don’t feel this “need” for sending everything into AI. And just because I can use it to do anything myself… why should I?
There are two caveats:
1) when I’m directly working on something, I don’t want to leave it. I’m hyperfocused on it. But that’s true even without AI; I would have been obsessed with building doughflow.app were it for AI or not. The difference is I may never have finished it without AI.
2) I find myself using it as a crutch sometimes for simple things like sending a message to a colleague. I don’t like that, and have been trying to avoid it. I’ve spent all my life writing naturally, and I don’t need someone to proofread my chats. That’s a habit I am tossing, in general.
Perhaps it is the ADHD that actually helps me focus; that is, perhaps the coping mechanisms built for managing my countless disparate thoughts also work for managing multiple disparate agents? Just thinking out loud.
Maybe you're right in the fact that the coping mechanism help with that. Maybe using LLMs is similar in that it is easy to lose focus for neurotypicals so having lived with that we know how to handle it. Or maybe it's just us, who knows.
Now I want to do a study!
I wonder if the extremely fast dopamine hits of LLM use trigger the same sort of dopaminergic response that ADHD does when faced with a stimulus? If so, are we simply better at dealing with those swings, since we’ve literally dealt with them our entire lives?
It's obviously massively generalising, but I'd say someone without ADHD is more likely to feel like the number of tasks they can do at once is only limited by pesky things like physics and time, and LLMs "solve" this by allowing them to do more at once.
A bit like when a good manager takes on more and more reports and their ability to manage the larger team goes completely to shit, but without the useful feedback loop of the grumpy, mismanaged team.
Things like drawing small defined boundaries, understanding where and how to delegate, and forcing myself to “hyperfocus” on 1-2 tasks at once. My strategy with building revolves around offloading mental context as fast as possible and into sequential, modular pieces. It’s made me a much better systems thinker. However, because of that I also don’t really see the point in running more than 1-2 coding agents at once. Others I work with don’t relate and are trying to parallelize as much as they can.
Part of me is concerned that this way of thinking isn’t compatible with the new wave of agentic SWE and these skills will become outdated. Or maybe it’s actually all converging back and is more valuable than ever.
I have not found myself less productive than my peers, but I have found myself less scattered across a bunch of different things. I do not think this is worse.
Also, consider reading about and/or getting tested for ADHD. You may not have it, of course, but I was diagnosed at 33 and it changed my life. Not because of medication (though I did that for about a year and a half, with success) but for a dozen other reasons, including being able to accept parts of myself rather than fight them, recognize (and reinforce) the coping mechanisms I’d built, recognize the ways in which it made me literally better than other people at certain things (some refer to ADHD being a “superpower” which it can definitely feel like at times), etc.
If you haven’t read Mark Suster’s posts about his experience with ADHD, read them - there are 3. If it resonates, consider talking to someone.
My email is in my profile; I’m happy to chat if you ever want. :)
What I’m trying to solve right now is how to communicate this intuition with colleagues and management. It feels like fighting against the current and near impossible with the levels of hype. Hoping that we as an industry make this realization soon
(That is also an unproven statement, obviously, but that’s what I’d want to study/test)
People who keep dozens of agent plates spinning all the time must have a broken understanding of productivity where they think more code = better regardless if anyone needed the code. In the old days engineers burned out when they got a promotion and thought they now need to write more code to compensate, maybe burnout in the agentic age looks different.
Pre-AI, I worked on projects where people tended to multitask and juggle many tasks at once. They felt productive, they worked a lot and hard and felt they are treated unfairly when outsiders complained about slow speed. Except that, their overall speed was extremely slow, really.
I am much more able to use fragmented bits of time now. Whereas before if I had 20 minutes here and there I would not be able to even read code or notes in order to get context enough to make a change let alone do real work.
I suspect personality matters. Often in discussions with AI or anyone I tend to take opiate position and constantly challenge and dig in. I don't go with the flow. With people this can be problematic. With AI it merely takes more time but probably a avoids some pitfalls.
Funnily enough this is my biggest problem when using AI. The time it spends "thinking" is time I'm more likely to get distracted and switch to another window.
And I use the term thinking liberally, because a lot of the time it's just wasting time. Sometimes I ask it to edit a file, it does, and then it thinks for 2 or 3 minutes afterwards. During that time I've verified the edit, committed it, opened a pull request and sometimes even merged before it finishes thinking about that edit. What was it thinking about during that time?
Or maybe it’s just keeping them in the work box - if someone wants to pay me to use LLMs and generate code, so be it. But keeping them away from hobbies, plans, art. Noble maybe, but then again there are a lot of projects I’m unlikely to do if I can’t find some way to make the level of effort required fit in with my lifestyle. With young kids, side project time is so incredibly limited.
Sometimes I feel like the main thing I’m learning from the AI wave is how modern life has a way of becoming more frustrating as more technology enters the picture. There’s all sorts of good in it, sure. But I do feel like so many more people are disgruntled by new tech compared to years past. I don’t think our cave man dopaminergic system is well suited for dealing with the expectations or availability of modern productivity and information.
I wonder about the transcendence take as well. I think a lot of us are strapped into this ride regardless. If there’s a fresh new paradigm awaiting, I hope it’s a good one.
But my main observation is mostly personal: I get taken in a lot by technology waves, hoping things will get better, but I haven’t ever crisply defined “better.” Or my criteria has been too loose and I conflate novelty with progress. If advanced models eventually accelerate science and medicine so that more people can lead healthy fulfilling lives, that sounds a lot better. But if a model helps me use an undocumented API of a random IoT device tomorrow, I don’t think anything is much better, and I didn’t even get much out of the process. When efficiency isn’t paramount, I’m often not clear on when to take the shortcut versus when to do the work.
Yeah, I think this is the issue more than anything having to do with LLMs. It’s also why I stopped switching to the latest JS library du jour many years ago. The one I had worked fine. I don’t need to learn React because it just came out after learning fourteen other paradigms / frameworks.
> Or my criteria has been too loose and I conflate novelty with progress.
Yup.
> If advanced models eventually accelerate science and medicine so that more people can lead healthy fulfilling lives, that sounds a lot better.
Agreed.
> But if a model helps me use an undocumented API of a random IoT device tomorrow, I don’t think anything is much better, and I didn’t even get much out of the process.
Disagree, but it depends on what you’re looking to get out of it. Is the interesting part the RE / discovery? Or do you just want to control your car from Home Assistant?
That is, once you’ve RE’d a React native app, you don’t really learn anything new by doing it again. I have no problem with an LLM helping me with that.
But if I want to learn to RE a binary or app? Then I’ll do that!
Just because you can doesn’t mean you must, even if everyone else is.
LLMs provide you with the illusion of “free concurrency,” and it sucks people in.
Also, especially in work, it undermines skills and experience you then just lack. That also will only be fine as long as you are doing the work somewhere where they provide/allow use of LLMs.
Not really a requirement, if you did fine without the tools before you would do fine again. The difference is you would know which ones are worth manually maintaining
Sometimes I'll get a neat little feature out while I feel inspired. This is great for doing something cool on a Sunday afternoon when priorities don't matter.
Sometimes I'll start something bigger and get halfway there, then I realise it's not what I should be doing, and it becomes another open tab that does not fit in my list of priorities.
Success depends on focus and saying no to most things. AI makes it easy to say yes to everything. Instead of making my todo list shorter, it just makes me add more stuff to it. It allows me to act on impulses instead of thinking carefully.
Well said.
And all my colleagues that whole heartily agree, just to go 'but i use it ...', are not much different from smokers & day traders that think they are the exception.
Would you also say that any message worth sending is worth writing by hand and mailing via the postal service, instead of emailing? And if you tried to replace your work use of email (or Slack etc.) with handwritten letters, what do you think would happen? That’s essentially equivalent to the position you’re taking.
Choose any 2: Cheap, fast, reliable. You might have a faster delivery time, but you cannot cheat on the maintenance burden.
At the moment I'm getting most value out of: - Checking integrity of data through custom scripting - Controlling my smart home stuff via quick commands - Reconciling account software with my bank and uploaded / emails receipts - Researching info across disparate sources - Managing pipelines of work (eg sales cycles or content schedules) reminding me when I need to do something - Parsing interesting articles / videos and pulling out the essentials - Reworking text I've written to work across multiple social accounts - Asking me great questions about whether I'm on track with what I said I wanted to do - Pinging me when an email looks important so I don't have to continually check it. - Coding on projects that I want to make rapid progress on for work (not tinkering)
This list is quite different to the one I'd have written a year ago - that would have mostly been about writing, information retrieval and coding.
The most obvious use cases are I think the least useful ones. Diffusion of AI will take time!
This reflects my impression of working with LLM as well.
Some people reluctantly position themselves solely at the bottom or at the top.
More flour more water. More water more flour.
This part resonates. AI has been such an accelerator that just yesterday I told Claude to make a brand new "Christian male meal planner with grace quotes from the Bible", bought a domain name, had the whole thing ready to go then thought "wtf am I doing, I already have half a dozen other side projects in the go, this is insane!" and just stopped it.
Human focus is a precious, finite resource. The productivity boom enabled by AI tools is probably going to benefit those capable of laser-like focus the most. Scattered thinkers (like myself, or perhaps the original article author) are in for a struggle at times.
Some part of you realized that Bronze Age superstitions aren’t worth that kind of effort.
I’m not mocking you. I’m pointing out that you’re making an epistemic mistake, and speculating that at some level, you recognize that, based on your self-described choices.
I don't think human should only judge AI design decisions described by LLMs in English. Understanding the codebase is important to be able to judge further development. So reading code, correcting code, articulating on it is still an important skill imo.
1) Read more books, 2) Learn more vocab, 3) Just write more!
Nothing crystalizes the thoughts coursing through our heads than writing something out - often times refining it - and forcing yourself to use new vocab that you have independently learned.
That’s why I found the final paragraph to be a bit of a shock - “ More important: it’s freeing me up to do human stuff like write and reflect right now.”
What? How? This doesn’t follow from the rest of the post at all. Maybe your new less frenetic and more human scale AI usage would be a good subject for the next post.
That was my read, anyway.
That is, stepping away from the laptop is what has given them back some free time and brain space.
What I was hoping to convey in the last two paragraphs was that I was trying to be thoughtful with where I use AI (work project where I did the initial thinking and setup and the agent does a lot of review and aggregation and reports back) and where I don’t (reflecting and writing). And the hope is that if I pay more attention to my values and priorities, I will make better decisions about my goals and how to achieve them, and maybe I’ll get more value out of both the tech and the process.
None of this is strictly new or limited to effects of AI. There has always been a process vs outcome calculus. But some of the parameters have changed in the past few years and I think we need to adjust.
It's always looming in the mind of an overachiever-type person such as myself that 'AI could search and synthesize better', and so it has a very seductive siren song for almost all tasks.
The current moment with AI does very much feel like the early days of atomic energy where people had all kinds of now-crazy-seeming use-cases for the atom, and this showed up again with at-home microwaves, as seen in that "Microwave Cooking for One" book from the 1980s (see: https://www.lesswrong.com/posts/8m6AM5qtPMjgTkEeD/my-journey...).
I do think with new technologies it's important to try many different forms and applications of it before finding their true practical utility, and this is no different. Maybe we're collectively (at least within the HN crowd) finally feeling the "cool-down of the mania" and starting to shift to more realistic uses of AI (at least until there might be another big capabilities jump...)
But at the same time, I cannot deny that AI has also made me become a better person in other areas. I'd always had poor 'handiness' or 'around the house' type skills, and the number of things I've gotten comfortable doing because of AI assistance, like hanging art on the walls, taking on some basic at-home plumbing and fix-it tasks, trying new things in the kitchen with cooking and food prep and brewing coffee etc, is frankly much higher than pre-2022. And that's to say nothing of all the help it's provided getting more fit in the gym. Maybe my upbringing lacked these things that others never had an issue with, but I am grateful to this technology for helping 'level the playing field' in this regard. Just like I know how it's done the same for non-techie people who are now a little more well-versed in the matters that I'd been naturally more fluent in.
So, like most tech, we need to be mindful of our use, especially us HN-types that are drawn to information processing systems, optimization, efficiency, etc. It's really a double-edged sword, and I've never felt that more than lately. And this is doubly true for anyone with certain neurodiverse information habits (all the open tabs, 'research trails' with poor follow-up, etc). I often joke that I'll know we've reached AGI when my open browser tab count hovers below ~5!