adamddev1

There's this crazy narrative that everyone starts parroting, saying the days of writing code yourself are over, now we have to use agents for everything.

No, we don't have to. We can just write code ourselves.

(My condolences to people who have jobs where AI is mandated.)

Linters and tests help of course, but they cannot "fix" the problem since tests cannot prove the absence of bugs.

Human incompetence, or human-written algorithmic incompetence in code can still be identified, appealed to, and dealt with.

With LLMs you get incompetence cut off from human embodiment and any chance of empathy, baked into opaque black-boxes, and automated and scaled.

We should be arriving to build things that are correct, not saying "stuff sucks already, let's make more stuff that sucks."

I made one language learning app with a top end LLM backend and at first I thought it was magic. But as I and other people used it more I realized:

- This thing is very consistently lying and misleading people. Do I want to introduce more deception and confusion into the world?

- people don't actually want to use this.

- I don't actually want to use this.

- Something about this feels wrong.

I dropped it. I have another couple of big language learning projects made with 100% human blood sweat and tears, long projects over many years. Zero LLMs or voice models used for anything. Those continue to grow and are loved, and I feel great about them.

It seems by definition "below par" of whatever human activity it's training on. How can we achieve greater and greater heights of intelligence if we're just creeping closer and closer to a sub-par copy of human creative activity.

People absolutely would feel the difference if AI models started making 90% of the decisions.

You were just denied an application because some LLM decided you fit a certain statistical pattern. Where did the LLM go wrong? What were the steps taken to make that decision? Who can you appeal to? Nobody. AI bureaucracy would just keep chugging along, making decisions based on prejudices and patterns it finds but doesn't feel.

Rainer Mühlhoff makes some compelling arguments in trying to warn people of the dangers of an AI bureaucracy/state.

> "I think there's a big difference between jazz and corporate landing pages."

Hard agree. I had to laugh at that sentence. I realized it wasn't really a fair analogy but also just kind of going off the copywriting example above. It's interesting how helpful this kind of thing is for different disciplines.

Jazz musicians also copy each other's solos for learning and practice purposes, but they would never actually perform more than a couple well placed quotes or licks from another player.

This logic enables and encourages corruption in all spheres of life. "Things are messed up as they are, so why try so hard to do the right thing?" "Let's be realistic."

Yes. People don't seem to understand is that if we build our tools and libraries with models that do random dumb stuff 0.01% of the time, those bugs and leaky abstractions bubble up and grow exponentially into errors and undependable behavior.

People have been doing great research with formal methods, dependent types etc. Disciplining ourselves to truly write and understand code, using the best in math and PL theory (FP and type theory etc) is the only way forward in my opinion. We have to make correctness a value and goal, or else we will keep spinning off into psychosis and the corruption of truth.

How did we get so much better at writing compilers? Was it a better understanding of how to make syntax trees with ADTs etc?

Interestingly enough, linguists also use Lisp-like parentheses or brackets to annotate sentence structures. Trees and brackets are isomorphic, as both phrase structure grammarians and the original SICP lectures pointed out.

The brackets in the title sentence would look a lot different though. ;-)

One of the cool things about code is that you can build stuff out of thin air, basically for free. It's not like woodworking where you have to pay for the wood.

We are moving into a weird time where people are assuming that now we have to pay machines churn out code.

Somehow they packaged up our own ability to think and are selling it back to us. If they can get us to forget how to do it we'll be the perfect customers, dependent forever.

I shouldn't have used the term "vibe-coded" as that's an extreme charicature. But I would argue in general heavy use of LLMs just does lead to sloppier code, and I think there's a general feel for that. Has AI use helped Microsoft make better software lately?

People live to benefit from AI for their own work, but are understandably uneasy about depending on code written by AI, even if it's (mostly) human reviewed. 2-3% more bugs may not seem like a big deal, but when you start building stacks of things these bugs bubble up and increase exponentially.

We can't say "just like to photographers did with photography" or "just like programmers did with higher-level languages." These developments are not analogous to LLMs. The jump into probabilistic text-guessing machines a fundamentally different thing.

And there might be no way to prove you really wrote something Post-2022. I wrote a long article, all by hand. I never used any LLMs, even for searching. I checked it with a couple of AI detection tools and they confidently said that 60% of the article was written by AI.

Yes. People are losing the plot and embracing corruption with logic like this:

There will always be a few employees stealing. So why don't we just use this system that consistently and randomly introduces theft into every level of our cashflow. We can't expect perfection!

People are forgetting that "if the tests pass" is not a good measure of reliability. CS 101 from Dijkstra: tests cannot prove the absence of bugs.

Everyone's excited about using AI to make their quick end products. But no-one wants to actually build on or rely on vibe-coded frameworks, languages, or tools.

Back when you could just buy the software and have it.

I second this feedback. The demo video was also clicking on emails and having Claude research and answer them automatically? Is this really the inhumane kind of communication we want to be pushing and normalizing? The UI looks nice, and I would love a better e-mail client, but the egregious pushing of AI convinces me not to try it.

If the developer is suggesting answering emails with agents without reading it, how do I know this project is shoved full of unchecked vibe-coding? Don't want to be so harsh, but I wanted to second this feedback since you seem serious about hearing it. All the best with the project, aside from the immediate AI turnoff it looks great.

So true. The last 3-4 times I have checked the sources did NOT say what the Google AI summary said they did.

The AI summaries are of course another layer of lies. The lie about the sources (which may also be lies). Before the lie would have to be written by a human and published on some website. I'm not saying that was good. But it was more manageable than the exponential and automatic falsehood being baked into the fabric of the web now.

This is a similar logic to saying, so what if the LLMs introduce bugs, did humans not make mistakes before? Or, so what if I'm cheating, does no-one around me cheat? Corruption doesn't justify larger scale corruption.

The lack of consent is real. As one of those viral commencement speakers who got booed said:

"Find a way to say yes."

My greater concern is that we are trading away hard earned truths for drastically inferred best guesses and slop that is just good enough to work. Search engine AI summaries flat out lie and deceive people all the time. LLM agents fudge over and cut corners. What people don't understand is the exponential damage that will be done as we keep baking these errors and untruths into everything we build, and build with.

It's the definition of cutting corners. Using statistical inference to guess at what's right as fast as possible.

I remember learning to make sites back around 2019 and seeing so many blog posts and hype around JWTs. It seemed like "this was the way to do it!" But I couldn't understand why session cookies weren't the better, simpler solution. I just used session cookies. Nice to be vindicated in retrospect.

Delightful little experience. Very nice. What would be even cooler would be if the axe only went partway down sometimes and then you have to lift the log up with the axe inside a couple of times to finish it off with that satisfying full split.

I think this is a great idea, and I've wondered about something like this before.

I do find it sad though that the opening description has to be:

> Two agents edit different functions in the same file? Clean merge.

Why does EVERYTHING has to be geared towards agents? Humans can use this too. Why not just "two commits contain edits for different functions in the same file?"

> I’ve watched this industry absorb higher-level languages, IDEs, autocomplete, agile and devops, and there were always crusty hold-outs reminiscing about the good old days before X came along and ruined everything.

But all of those things were consciously built, deterministic and transparent tools. LLMs/AI are something fundamentally different.

This is quite a nice idea and works well, but I think I would rather spend the time listening to and imitating real Miles Davis solos etc.