Llms Are Not Fun

So does autocomplete. Why not treat LLM as next autocomplete iteration?

Yesterday I had semi-coherent idea for an essay. I told it to an LLM and asked for a list of authors and writings where similar thoughts have been expressed - and it provided a fantastic bibliography. To me, this is extremely fun. And, reading similar works to help articulate an idea is absolutely part of writing.

"LLMs" are like "screens" or "recording technology". They are not good or bad by themselves - they facilitate or inhibit certain behaviors and outcomes. They are good for some things, and they ruin some things. We, as their users, need to be deliberate and thoughtful about where we use them. Unfortunately, it's difficult to gain wisdom like this a priori.

> LLMs code for you. They write for you.

A director is the most important person to the creation of a film. The director delegates most work (cameras, sets, acting, costumes, makeup, lighting, etc.), but can dive in and take low-level/direct control of any part if they choose.

To get the LLM to code for me, I need to write.

have you actually done some projects with e.g. claude code? completely greenfield entirely up to yourself?

because ime, youre completely wrong.

I mean i get were youre coming from if you imagine it like the literal vibe coding how this started, but thats just a party trick and falls off quickly as the project gets more complex.

to be clear, simple features in an existing project can often be done simply - with a single prompt making changes across mutliple files - but that only works under _some circumstances_ and bigger features / more indepth architecture is still necessary to get the project to work according to your ideas

And that part needs you to tell the llm how it should do it - because otherwise youre rolling the dice wherever its gonna be a clusterfuck after the next 5 changes

I write what I want the LLM to do. Generating a satisfactory prompt is sometimes as much work as writing the code myself - it just separates the ideation from the implementation. LLMs are the realization of the decades-long search for natural language programming, dating at least as far back as COBOL. I personally think they are great - not 100% of the time, just as a tool.

You don't have to do any of that if you simply don't make mistakes in the first place FYI

I'd rather spend my time designing and writing code than spending it debugging and reformatting whatever an LLM cobbled together from stack overflow and github. 'Design, architecture, problem solving, etc' all takes a backseat when the LLM barfs out all the code and you have to either spend your time convincing it to output what you could have written yourself anyway or play QA fixing its slop all day long.

>usually a mile or so into a walk

My place for that is in the shower.

I had one of those shower epiphanies a couple mornings ago... And I fed it into a couple LLMs while I was playing a video game (taking some time over the holidays to do that), and by the afternoon I had that idea as working code: ~4500 LOC with that many more in tests.

People keep saying "I want LLMs to take out the laundry so I can do art, not doing the laundry while LLMs do art." This is an example of LLMs doing the coding, so I can rekindle a joy of gaming, which feels like it's leaning in the right direction.

Or on the toilet.

I can't wait for the machine to do all my work so I can finally do my personal projects without being interrupted.

> Expertise and craft and specialized knowledge can become irrelevant in a heartbeat, so your meaning and joy and purpose should be in higher principles.

My meaning could be in higher purposes; however I still need a job to be enable/pursue those things. If AI takes the meaning out of your craft it takes out the ability to use it to pursue higher order principles as well for most people, especially if you aren't in the US/big tech scene with significant equity to "make hay while the sun is still shining".

To you maybe, to someone else maybe not. It's really hard to pin down a universal framing for existence.

My family eats out at a nice steak restaurant every Christmas no one wants to cook. None of us like to cook.

[delayed]

My particular brain feels the fun is diminished if I can't just click undo.

I just had Claude Code finetune a reranker model to improve it significantly across a large set of evals. I chose the model to fine tune, the loss function, created the underlying training dataset for the re-ranking task, and designed the evals. What thinking did I outsource exactly?

I guess did not waste time learning the failure-prone arcana of how to schedule training jobs on HuggingFace, but that also seems to me like a net benefit.

Maybe "scissor statement" would be more apt, at least for the headline.

My overwhelming experience is that the sort of developers unironically using the phrase "vibe coding" are not interested in or care about good design and code quality.

I really doesn't. I just ditched my wordpress/woocommerce webstore for a custom webshop that I made in 3 days with Claude, in C# blazor. It is better in every single way than my old webshop, and I have control over every aspect of it. It's totally production ready.

The code is as good or even better than I would have written. I gave Claude the right guidelines and made sure it stayed in line. There are a bunch of playwright tests ensuring things don't break over time, and proving that things actually work.

I didn't have to mess with any of the HTML/css which is usually what makes me give up my personal projects. The result is really, really good, and I say that as someone who's been passionate about programming for about 15 years.

3 days for a complete webshop with Stripe integration, shipping labels and tracking automation, SMTP emails, admin dashboard, and all the custom features that I used to dream of.

The future is both really exciting and scary.

You can be as specific as you want with an LLM, you can literally tell it to do “clean code” or use a DI framework or whatever and it’ll do it. Is it still work? Yes. But once you start using them you’ll realize how much code you actually write is safely in the realm of boilerplate and the core aspect of software dev is architecture which you don’t lose when instructing an agent.

Perhaps for the inexperienced or timid. Code quality is it compiles and design is it performs to spec. Does properly formatted code matter when you no longer have to read it?

In the long run I think it's pretty unhealthy to make one's career a large part of one's identity. What happens during burnout or retirement or being laid off if a huge portion of one's self depends on career work?

Economically it's been a mistake to let wealth get stratified so unequally; we should have and need to reintroduce high progressive tax rates on income and potentially implement wealth taxes to reduce the necessity of guessing a high-paying career over 5 years in advance. That simply won't be possible to do accurately with coming automation. But it is possible to grow social safety nets and decrease wealth disparity so that pursuing any marginally productive career is sufficient.

Practically, once automation begins producing more value than 25% or so of human workers we'll have to transition to a collective ownership model and either pay dividends directly out of widget production, grant futures on the same with subsidized transport, or UBI. I tend to prefer a distribution-of-production model because it eliminates a lot of the rent-seeking risk of UBI; your landlord is not going to want 2X the number of burgers and couches you get distributed as they'd happily double rent in dollars.

Once full automation hits (if it ever does; I can see augmented humans still producing up to 50% of GDP indefinitely [so far as anyone can predict anything past human-level intelligence] especially in healthcare/wellness) it's obvious that some kind of direct goods distribution is the only reasonable outcome; markets will still exist on top of this but they'll basically be optional participation for people who want to do that.

    Education for educations sake isn't a thing.

[delayed]

You don't have to use LLMs the whole time. For example, I've gotten a lot done with AI and had the time to spend over the holidays on a long time side project... organically coding the big fun thing

Replacing Dockerfiles and Compose with CUE and Dagger

If you're using your YouTube and TikTok addictions to build things that other people can use, go nuts

I’m not writing good things about LLMs because I’m busy thinking about what to build next, since the things build faster than I can come up with ideas.

It’s amazing and scary. I was wondering how takeoff would look like and I’m living it for better or worse.

Ehhhhhh hating on AI is also extremely popular on social media.

That's a very broad statement, explicitly covering all types of code and all kinds of coders. Are you really confident enough to make such an assertion?

That's not true of all code bases.

Some projects are built by very small close-knit teams which get to drive everything.

In that scenario the developers have tight control over every little detail.

Well one of the inherent issues is assuming that text is the optimal modality for every thing we try to use an LLM for. LLMs are statistical engines designed to predict the most likely next token in a sequence of words. Any 'understanding' they do is ultimately incidental to that goal and once you look at them that way a lot of the shortcomings we see become more intuitive.

There's a lot of problems LLMs are really useful for because generating text is what you want to do. But there's tons of problems which we would want some sort of intelligent, learning behaviour that do not map to language at all. There's also a lot of problems that can "sort of" be mapped to a language problem but make pretty extraneous use of resources compared to a (existing or potential) domain specific solution. For purposes of AGI, you could argue that trying to express "general intelligence" via language alone is fundamentally flawed altogether -- although that quickly becomes a debate about what actually counts as intelligence.

I pay less attention to this space lately so I'm probably not the most informed. Everyone seems so hyped about LLMs that I feel like a lot of other progress gets buried, but I'm sure it's happening. There's some problem domains that are obviously solved better with other paradigms currently: self-driving tech, recommendation systems, robotics, game AIs, etc. Some of the exciting stuff that can likely solve some problems better in the future is some of the work on world models, graph neural nets, multi modality, reinforcement learning, alternatives to gradient descent, etc. I think it's a debate whether or not LLMs are a local maxima but many of the leading AI researchers seem to think so -- Yann Lecun recently for e.g. said LLMs 'are not a path to human-level AI'

Thanks, this seems interesting. I'll give it a read. I admittedly don't keep tabs as much as I should these days. I feel like every piece of AI news is about LLMs. I suppose I should know other people are still doing interesting things :)

Good for you it feels that way to you. But then you need to check how the LLM refactored stuff. Because (a) you never know if it's actually correct and (b) the code needs to be maintained into the future, so you need to know how it works and be able to fix bugs in it. And then you are suddenly spending a lot more time understanding someone else's code, someone you can't discuss the 'why' of the code over a coffee with. Of course the same goes for library code but to me understanding bugs in a library (and reporting/fixing them) feels more useful than understanding the one-off output of an LLM. And for a library the coffee part might not fly, but at least you can discuss stuff with the original author(s). I'm not saying my feeling is the absolute truth, it's very subjective.

I guess mechanics must feel the same about modern computerized cars, where suddenly the injection timing is no longer a mechanical gadget you can tweak by experience, but some locked down black box you don't have control over.

Also I really dislike that (for now) using an LLM means selling your soul to some dubious company. Even if you use only the free tier you still need to upload your code and let the LLM do whatever with it. If an LLM is an indispensible part of being a programmer, everybody will be held hostage by the large techfirms (even more...).

> don't think the AI boom would of really been that hyped up.

For one thing, LLMs aren't terrible at grammar.