There are free and offline options, like Llama.cpp, but you will have to pay by giving up your privacy to Meta (or similar large companies)

No it’s not much different. I grew up poor. It was a struggle to have internet access.

Small models and processors are going to continue improving to the point that you’ll be able to vibe code locally on your phone at some point.

When the iPhone came out, not everyone had a smartphone. Now 90% of the US has a smartphone, and many of these smartphones run generative local models.

"It's harder to convince other people to pay for this for me" is an insane criticism. Not every AI model needs a premium account, you can even run many excellent models locally if you don't want to pay for an internet connection.

At some point you just have to accept that yes things are easier if you have a little bit of spending money for things. That's not "sad" it's a basic fact of life.

This feels like picking a random thing against LLMs to complain about. These tools are not even necessary today I am not sure why they would be necessary tomorrow beyond efficiency. If that day does come though, you would have to assume open source models would also be coming a long way.

I guess hardware being able to run a local model will eventually get cheap enough, but for a lot of people even buying an Apple device or something with a good enough GPU is prohibitive.

Speaking of free models on OpenRouter, DeepSeek R1 0528 is also available for free.

https://openrouter.ai/deepseek/deepseek-r1-0528:free

Swift and swift playground might be a good introduction to programming, but it feels likely not to lead to as many opportunities as a more popular system. And I don’t just mean job opportunities.

12-year-old me wasn't either, but he was noodling around on a computer that cost $2500 (more like $5500 in today's dollars). I think our parents loved us very much and must have had some means to afford the capital cost of a computer back then.

I don't see my 7-year-old paying $5 for each hour he wants to program (and no way in hell would I give him my credit card), but I could easily envision paying $20/month for a Claude subscription and letting him use it. We pay more than that for Netflix & Disney+.

12-year-old me had (or rather, my family had) a Celeron 333 MHz and a Pentium III 550 MHz, both from Gateway, because that was the sole awesome perk my dad got from working there: literally free computers, with a required number of years of employment to pay them off. In 2000, the P3 was still pretty hot shit. I dual-booted them with every Linux distro under the sun. Since we had dial-up, the only way I had those distros was from 4-H [0], which at the time in Nebraska had a partnership with University of Nebraska to do tech instruction; once a quarter, we’d drive down to a campus (usually UNL) and spend a weekend learning something (LAMP stack, hardware troubleshooting, etc.), and having a LAN party at night. Also we had free access to their (at the time) screamingly fast internet, so I would download distros and packages to try out later.

My online upbringing was very much of the RTFM variety, and I am convinced that was and is a good method to learn. It’s not like the grumpy graybeards were cruel, they just didn’t want to waste their time answering the same “how do I…” questions from noobs. If you explained what you were experiencing, what you had read, and what you had tried, they were more than happy to help out. I don’t think that’s an unreasonable approach.

[0]: https://4-h.org/

You don't have to pay $100 to find out, you can do that for ~$5-20 by directly buying API credits.

I don't know for sure whether it's worth it yet. Further experimentation is needed, as well as giving it an honest shot and trying to learn the nuances of the tool. But the way I look at it - if this actually is a future career path, the net present value of its payoff is measured in the millions of dollars. It's worth spending ~$20 and a few nights of my time to figure that out, because the odds can be pretty damn low and still have the expected value pencil out. It's sorta like spending $200 on 1/4 of a Bitcoin in 2013 because I was curious about the technology - I fully expected it to be throwing money down the toilet, but it ended up being quite worth it. (I wish I'd had the same mindset when I could've bought into the Ethereum ICO at a penny or so an ETH.)

Are you spending every $100 absolutely efficiently? How do you know?

"Price dumping" isn't an economic term in common use.

"Dumping" in international trade is somewhat similar but the reasons that is illegal are very different: https://en.m.wikipedia.org/wiki/Dumping_(pricing_policy)

Pricing at a loss by VC funded companies is great for consumers. It rarely is at a loss though - they look at the lifetime value.

Pricing at a loss by big tech could be viewed as anticompetitive. Personally I like that Gemini keeps OpenAI prices lower but one could argue it has stopped OpenAIs growth.

My bill for LLMs is going up over time. The more capable, higher-context models dramatically increase my productivity.

The spend prices most of the developing world out -- an programmer earning $10k per year can't pay for a $200/month Claude Max subscription..

And it does better than $6k-$10k programmers in Africa, India, and Asia.

It's the mainframe era all over again, where access to computing is gated by $$$.

Software development costs hundreds of dollars in the 90s. My parents bought VB 6 for $600.

Only in tech are we shocked when things cost money. I don't know that any other industry expects such a reality.

Sure but that's not what the person I was replying to was talking about, nor what I was talking about.

Cost of access is absolutely a problem in tech.

The problem can certainly be multi-faceted though.

If the trend of the last four decades continues, this is the outcome.

> ... LLMs aren't very good at logical reasonsing.

I'm curious about what experiences led you to that conclusion. IME, LLMs are very good at the type of logical reasoning required for most programming tasks. E.g. I only have to say something like "find the entries with the lowest X and highest Y that have a common Z from these N lists / maps / tables / files / etc." and it spits out mostly correct code instantly. I then review it and for any involved logic, rely on tests (also AI-generated) for correctness, where I find myself reviewing and tweaking the test cases much more than the business logic.

But then I do all that for all code anyway, including my own. So just starting off with a fully-fleshed out chunk of code, which typically looks like what I'd pictured in my head, is a huge load off my cognitive shoulders.

You can view Claude Code as a non-deterministic compiler where you input english and get functioning code on the other end.

The non-determinism is not as much as a problem because you are reading over the results and validating that what it is created matches what you tell it to do.

I'm not talking about vibe-coding here, I'm grabbing the steering wheel with both hands because this car allows me to go faster than if I was driving myself, but sometimes you have to steer or brake. And the analogy favors Claude Code here because you don't have to react in milliseconds while programming.

TL;DR: if you do the commit you are responsible for the code it contains.

So here's a transcript of a recent Claude Code interaction I had, trying to build an MP3 organizer that relies on heuristics and fuzzy matching to make sense out of all the random MP3s I downloaded in the 90s. This all happened in-between prompts: I gave it a correct example and an incorrect example and asked it why it was giving the incorrect example.

● Let me trace through what happens with the "Jordan Rudess" example:

  1. Directory: Jordan Rudess/
  2. file_rel_path would be Path("Jordan Rudess") (after removing MP3 Disk prefix)
  3. If Jordan Rudess matches in path component 0, scope_path would be Path("Jordan Rudess")
  4. _path_is_in_scope checks if "Jordan Rudess".startswith("Jordan Rudess") → True

  This looks correct. Let me check the inference logic:

● The inference priority looks correct - path_component_match is indeed the highest priority. But I think I see the issue now. Let me look at the actual logic flow in the main organize method:

● Read(mp3_organizer_v2.py) ⎿ Read 10 lines (ctrl+r to expand)

● The flow looks right. Let me examine the actual problem more carefully. Looking back at the _apply_path_component_matching method, I think there might be an issue with when the matching happens. Let me check the logic:

● Update Todos ⎿ Check _path_is_in_scope logic for correctness Verify artist inference priority logic works with path_component_match Test and fix any remaining issues with artist propagation

● I think I see the issue now. The problem is that the current code only looks for matches in files that already have file_id3 tags, but it needs to propagate those matches to all files in the directory. However, the current implementation is still only applying matches to files that have matching candidates, not ALL files. Let me check the exact logic:

● Read(mp3_organizer_v2.py) ⎿ Read 8 lines (ctrl+r to expand)

● Actually, looking at this more carefully, the logic should work. The matched_artists dictionary contains the matches found from ANY file that had a file_id3 tag matching a path component, and then it applies that match to ALL files in the file_infos list via the loop at line 737.

  Wait, let me understand the scoping issue better. Let me check what file_rel_path and scope_path would actually be in your example:

● Ah! I think I see a potential issue. The file_rel_path is calculated as the directory path, not the individual file path. So for all files in the same directory, file_rel_path will be the same. This should actually work correctly.

  But let me double-check the scope matching logic. If the directory is "Jordan Rudess" and the scope_path is also "Jordan Rudess", then _path_is_in_scope should return
  True.

  Let me debug this by adding some debug output to see what's actually happening. But first, let me look at a potential issue: maybe the match isn't being found in the first place?

For things where I don't want creativity, I tell it to write a script.

For example, "write a comprehensive spec for a script that takes in the date and a location and computes when golden hour is." | "Implement this spec"

That variability is nice when you want some creativity, e.g. "write a beautiful, interactive boids simulation as a single file in html, css, and JavaScript."

Words like "beautiful" and interactive" are open to interpretation, and I've been happy with the different ways they are interpreted.

Seconding this. My work has had the same problem - by the time I've got things all hooked up, figured out the complicated stuff - my brain (and body) clock out and I have to drag myself through hell to get to 100%. Even with ADHD stimulant medication. It didn't make it emotionally easier, just _possible_ lol.

LLMs, particularly Claude 4 and now GPT-5 are fantastic at working through these todo lists of tiny details. Perfectionism + ADHD not a fun combo, but it's way more bearable. It will only get better.

We have a huge moat in front of us of ever-more interesting tasks as LLMs race to pick up the pieces. I've never been more excited about the future of tech

I'm kind of in this cohort. While in the groove, yea, things fly but, inevitably, my interest wanes. Either something too tedious, something too hard (or just a lot of work). Or, just something shinier shows up.

Bunch of 80% projects with, as you mentioned, the interesting parts finished (sorta -- you see the line at the end of the tunnel, it's bright, just don't bother finishing the journey).

However, at the same time, there's conflict.

Consider (one of) my current projects, I did the whole back end. I had ChatGPT help me stand up a web front end for it. I am not a "web person". GUIs and what not are a REAL struggle for me because on the one hand, I don't care how things look, but, on the other, "boy that sure looks better". But getting from "functional" to "looks better" is a bottomless chasm of yak shaving, bike shedding improvements. I'm even bad at copying styles.

My initial UI was time invested getting my UI to work, ugly as it was, with guidance from ChatGPT. Which means it gave me ways to do things, but mostly I coded up the actual work -- even if it was blindly typing it in vs just raw cut and paste. I understood how things were working, what it was doing, etc.

But then, I just got tired of it, and "this needs to be Better". So, I grabbed Claude and let it have its way.

And, its better! it certainly looks better, more features. It's head and shoulders better.

Claude wrote 2-3000 lines of javascript. In, like, 45m. It was very fast, very responsive. One thing Claude knows is boiler plate JS Web stuff. And the code looks OK to me. Imperfect, but absolutely functional.

But, I have zero investment in the code. No "ownership", certainly no pride. You know that little hit you get when you get Something Right, and it Works? None of that. Its amazing, its useful, its just not mine. And that's really weird.

I've been striving to finish projects, and, yea, for me, that's really hard. There is just SO MUCH necessary to ship. AI may be able to help polish stuff up, we'll see as I move forward. If nothing else it may help gathering up lists of stuff I miss to do.

Ironically, I find greenfield projects the least stimulating and the most rote, aside from thinking about system design.

I've always much preferred figuring out how to improve or build on existing messy systems and codebases, which is certainly aided by LLMs for big refactoring type stuff, but to be successful at it requires thinking about how some component of a system is already used and the complexity of that. Lots of edge cases and nuances, people problems, relative conservativeness.

Looks like the definition of boilerplate will continue to shift up the chain

What is described in that post is a skill issue. The attention issues spoke about can be worked through. Because attention is a skill like anything else. And if one does not learn to practice. One becomes victim to it

Aider actually has a mode that is called "copy-paste" where it basically gives you a context to paste in an LLM chat and then you copy-paste back the reply to aider

I can't tell how widespread any of is, to be honest.. mostly because it's anecdata, and impossible to determine if what I'm seeing is just ragebait, or shallow dunks by reply-guys in comment sections, or particularly-loud voices on social media that aren't representative of the majority opinion, etc

That said, the amount of sort-of-thoughtless, I'm-just-repeating-something-I-heard-but-don't-really-understand outrage towards AI that I'm seeing appears to be increasing -- "how many bottles of water did that slop image waste??", "Clanker"-adjacent memes and commentary (include self-driving + robots in this category), people ranting about broligarchs stealing art, music, movies, books to train their models (oddly often while also performatively parroting party lines about how Spotify rips artists off), all the way to refusing to interact with people on dating apps if they have anything AI in their profiles hahaha (file "AI" alongside men holding fish in their pics, and "crypto" lol)

It's all chronically-online nonsense that may well just be perception that's artificially amplified by "the algorithm".

Me, I have no fundamental issue with any of it -- LLMs, like anything else, aren't categorically good or bad. They can be used positively and negatively. Everything we use and consume has hidden downsides and unsavory circumstances.

Yes, people are refusing for those reasons. I don't know how many, but I'd say about half of the the people I know who do not work in tech are rejecting AI, with ethics being the primary reason. That is all just anecdata, but I suspect the tech bubble around AI is making people in tech underestimate how many people in the world simply are not interested in it being part of their lives.

In the case of programming is not quite as problematic with unsolved problems as much as others, like completeness. In the case of programming, it's context and understanding. It's great for small chunks of code but people think you can vibe code entire interactive applications with no programming knowledge, but LLMs simply don't understand, so they can't keep a cohesive idea of what the end goal is in mind. The larger the codebase it needs to work on the more likely it is to make catastrophic errors, create massive security flaws, or just generate nonfunctional code.

Programming LLMs will become awesome when we create more narrowly targeted LLMs rather than these "train on everything" models.

I personally use it for a lot of SwiftUI work. I parallelize it across at least 3 projects at once. I use only the largest models on highest thinking modes. I give instruction on implementation and provide reference implementations.

I also use it for adding features/feature polish that address user pain points but that I can't prioritize for my own manual work just yet. There are a variety of user requests which sometimes LLMs are able to get done very suddenly when I give it a shot quickly. For these tasks, it's ok to abandon & defer them if the LLM spins its wheels.

It's the same as saying that writing good commit messages is a boon for extroverts and neurotypicals. It's a computer. You're giving it instructions, and the only difference to traditional coding is that the input is English text.

"we will just be offline more"

I think it will be quite some time into the future, before AI can impersonate humans in real life. Neither hardware, nor software is there, maybe something to fool humans for a first glance maybe, but nothing that would be convincing for a real interaction.

If your solution to this problem is the web of trust, to be blunt, you don't have a solution. I am techie whose social circle is mostly other techies, and I know precisely zero people who have ever used PGP keys or any other WoT-based system, despite 30 years of evangelism. It's just not a thing anybody wants.

"Web of Trust" has been the proposed answer for, what, 30 years now? But everyone is too lazy to implement and abide by it.

well, he was arguing that it's not worse than 99% of the human slop that gets posted, so where do you draw the line?

* well crafted, human only? * Well crafted, whether human or AI? * Poorly crafted, human * well crafted, AI only * Poorly crafted, AI only * Just junk?

etc.

I think people will intuitively get a feel for when content is only AI generated. If people spend time writing a prompt that doesn't make it so wordy, and has personality, and it OK, then fine.

Also, big opportunity going to be out there for AI detected content, whether in forums, coming in inmail inboxes, on your corp file share, etc...

Smaller communities too.

I grew up in... slightly rural america in the 80s-90s, we had probably a couple of dozen local BBSes the community was small enough that after a bit I just knew who everyone was OR could find out very easily.

When the internet came along in the early 90s and I started mudding and hanging out in newsgroups I liked them small where I could get to know most of the userbase, or at least most of the posing userbase. Once mega 'somewhat-anonymous' (i.e. posts tied to a username, not like 4chan madness) communities like slashdot, huge forums, etc started popping up and now with even more mega-communities like twitter and reddit. We lost something, you can now throw bombs without consequence.

I now spend most of my online time in a custom built forum with ~200 people in it that we started building in an invite only way. It's 'internally public' information who invited who. It's much easier to have a civil conversation there, though we still do get the occasional flame-out. Having a stable identity even if it's not tied to a government name is valuable for a thriving and healthy community.

Honestly, having seen how it can be used against you, retroactively, I would never ever engage in a discussion under my real name.

(The fact that someone could correlate posts[0] based on writing style, as previously demonstrated on HN and used to doxx some people, makes things even more convoluted - you should think twice what you write and where.)

[0] https://news.ycombinator.com/item?id=33755016

id.me?

Not government owned, but even irs.gov uses it

I do this every time I find myself typing something I could get written up over or even fired for.

1. I'm usually too emotional to write out why I feel that way instead of saying what I feel.

2. I really don't like the person (or their idea) but I don't want to get fired over it.

Claude is really great at this: "Other person said X, I think it is stupid and they're a moron for suggesting this. Explain to them why this is a terrible idea or tell me I'm being an idiot."

Sometimes it tells me I'm being an idiot, sometimes it gives me nearly copy-pasta text that I can use and agree with.

> People will be more than willing to say, "Claude, impersonate me and act on my behalf".

I'm now imagining a future where actual people's identities are blacklisted just like some IP addresses are dead to email, and a market develops for people to sell their identity to spammers.

I mean, that's fine I guess as long as its respectable and respects the forum.

"Claude write a summary of the word doc I wrote about x and post it as a reply comment," is fine. I dont see why it wouldnt be. Its a good faith effort to post.

"Claude, post every 10 seconds to reddit to spam people to believe my politics is correct," isn't but that's not the case. Its not a good faith effort.

The moderation rules for 'human slop' will apply to AI too. Try spamming a well moderated reddit and see how far you get, human or AI.

That's fine, because once someone is banned, the impersonations are also banned.