jjani

> ... And the 50 millions yearly tourists be damned.

It does seem like the plan is to slash that number too, and they're succeeding.

Would be interesting if they'd add another one on non-fiction creative writing. For example, turning a set of investigative notes and findings into a Pulitzer-prize winning article that wouldn't be out of place in a renowned, high-quality newspaper.

IME, for LLMs (just like humans) this skill doesn't necessarily correlate with fiction writing prowess.

This is probably harder to judge automatically (i.e. using LLMs) though, maybe that's why they haven't done it.

The "everything is securities fraud" meme is really unfortunate, not quite as bad as the "fiduciary duty means execs have to chase short-term profit" myth, but still harmful.

It's only because lying ("puffery") about everything has become the norm in corporate America that indeed, almost all listed companies commit securities fraud. If they'd go back to being honest businessmen, no more securities fraud. Just stop claiming things that aren't true. This is a very real option they could take. If they don't, then they're willingly and knowingly commiting securities fraud. But the meme makes it sound to people as if it's unavoidable, when it's anything but.

I have, dozens of times, and it's generally better than 3.7. Especially with more context it's less forgetful. o1-pro is absurdly expensive and slow, good luck using that with tools. Virtually all benchmarks, including less gamed ones such as Aider's, show the same. WebLM still has 3.7 ahead, with Sonnet always having been particularly strong at web development, but even on there 2.5 Pro is miles in front of any OpenAI model.

Gemini subscription? Surely if you're "using LLMs correctly" you'd have been using the APIs for everything anyway. Subscriptions are generally for non-techy consumers.

In any case, just straight up saying "it isn't good" is absurd, even if you personally prefer others.

> They cant produce new things, and so meta and google are struggling to compete when actual merit matters, and they cant just brute force the solutions.

You haven't been keeping up. Less than 2 weeks ago, Google released a model that has crushed the competition, clearly being SotA while currently effectively free for personal use.

Gemini 2.0 was already good, people just weren't paying attention. In fact 1.5 pro was already good, and ironically remains the #1 model at certain very specific tasks, despite being set for deprecation in September.

Google just suffered from their completely botched initial launch way back when (remember Bard?), rushed before the product was anywhere near ready, making them look lile a bunch of clowns compared to e.g. OpenAI. That left a lasting impression on those who don't devote significant time to keeping up with newer releases.

I didn't think that was confusing, your article pointed it out fine and clearly. However it doesn't address my point. "(Un)intentional" is solely a qualifier regarding the intentionality behind becoming a moderate, not the "moderate" part itself, which is what I'm arguing against. Independent thought doesn't necessarily lead to a moderate position. It leads to a position that almost certainly isn't perfectly aligned with a single -ism ideology. In fact, many of the individual beliefs held may well be considered the opposite, "extreme" in the sense of the Overton window (which is what "moderate" is really about).

It is unfortunately also abandoned. The creator got an offer from Cursor to integrate it last Sepmtember and abandoned it, seems likely it was part of the contract. By the looks of it he also didn't refund subscribers, not a good look.

I'm sure it does, as Gemini 2.5 Pro has been making every other model look pretty bad.

> “The Respected Comrade’s brilliant economic agenda is an all-encompassing plan to revive our economy by unleashing Chosun energy, implementing tariffs to level the playing field, and bringing trillions of won in historic investments to our country’s manufacturing sector,” minister Paek Sông-Ryông said.

Those people will next use their mother's ID, their grandparents' IDs, and so on. Then move on to stolen IDs. Which is not preventable unless you do full-on hardware attestation tied to real ID, which means China-level surveillance.

Source: Living in a country where gaming accounts are already tied to real identities :)

I can call something "purple" and then give a specific explanation I mean "computer" by it, but I just shouldn't be using purple like that in the first place. The word "moderate" is too entrenched, gives people an immediate instinctive, emotional reaction based on the established meaning of the word. This is not the word to repurpose here.

I can strongly sympathize. The image with the squares and circles hit home hard, from an early age, it's been pretty lonely. Depending on your environment it can be super hard to find others part of the 1%, so you really need to treasure them when you do find them.

One point of criticism:

The usage of the word "moderate". It seems PG's article is the one to blame here. The word "moderate" when used about politics means something to people in English. And given that meaning, saying that independent thought leads to one being "moderate", is straight up wrong. What the article is really talking about is that independent thought leads to a set of beliefs that is unlikely to be a very good fit for any particular ideology, and therefore, political party. That's true! But that's not "moderate". That's.. diverse, pragmatic, non-ideological. Those words aren't ideal either, but "moderate" is definitely not it.

The 99%/1% is also greatly overstated in a way. Firstly, it's definitely dependent on locale, culture, subculture, environment, as the writer already says themselves. More importantly, if you manage to somehow get people 1:1 in an environment where they feel safe, it turns out that many actually aren't that tribal/ideological after all, and they do actually have beliefs that span different mainstream tribes. But then that conversation finishes, and they go back to being a tribe member.

I'm pretty sure there's plenty of experiments that directly show the above. That when you give people policy choices that are non-obvious (e.g. they've never thought about), and then make them vote on them, they'll often vote against their tribe. But if you'd beforehand tell them which tribe voted which way, they'll always vote with the tribe.

At the risk of sounding very arrogant, I've found this incredibly obvious even when I was just 18 years old. Decades have passed, plenty of my beliefs have changed, but this one hasn't.

The chance that one "ideology", whether it's liberalism, conservatism, anarchism , fascism or any-ism is always the right answer to every single societal question, is 0. It's comparable to the idea of exactly 1 of the (tens of) thousands of religions being the true one, correct in everything, with all of the others being wrong.

And this extends to politics. Where I'm from, the political landscape is very different from the US, with at least 5+ different parties that support different policies in various ways. At the same time, it's similar - there isn't a single one that approaches things on a case-by-case basis, each of them being ideology-based.

> So I don’t think being a centrist implies one is not tribal, rather that the degree to which your beliefs on a variety of issues align with the “default” of a group implies how tribal you are.

Absolutely, "centrism" is an ideology in itself. This is also why the usage of the word "moderate" in the article and by PG is very unfortunate. That word too comes with a whole lot of baggage, and saying that independent thought leads to one being "moderate" in the way that most people think of that word, is straight up wrong. We need a different word, but I'm not great at coining those. "pragmatic" is the best one I can come up with. I can feel a "pragmatism is an ideology!" coming, but "the ideology of not looking at things from an ideological perspective" is entirely different from anything else. I'm sure the bright minds here can give better words.

> In other words, a politically thoughtful and independent person probably has a basket of opinions that don’t fit into neat left or right, liberal or conservative, etc. categories.

Very much so. And as the article points out, this is unfortunately a very lonely experience, so it's completely logical that most don't opt for this, instead choosing the warmth of a dogmatic community.

I am. I'm saying that what you see as its main selling point:

> It's easy to customize!

May be very overstated, with lots of users not caring or even knowing about that as a selling point. They just use it for all of the other selling points.

> The thing is literally distributed via copied files and updated via diffs instead of being a package. The entire cargo cult that lead up to that is 100% the idea they'll modify it. It's just either don't out of apathy (and should have just used a component library), or do and do so terribly (and should have just used a component library).

It definitely started out that way. Just like sports brands started out selling trainers for, you know, sports. And now 99.9% of the minutes-worn for them is during non-sports activities. But they're meant for sports! That's their selling point! Sure, most people couldn't care less though, and don't even buy them with the idea of using them for sports.

Back up the thread it was about the whole market, but then we started talking about AI developers (=the vacuum salesmen) and how much they're affected by the mindshare, which is also the part I quoted. The average consumer, absolutely, they all currently use GPT as the data shows, and are indeed led by Ghibli memes - though there's still plenty of time to change that by the huge players, á la Chrome eating all other browsers despite Microsoft's IE monopoly. AI developers though? Much less so.

> satisfy all the people lying to themselves that one day they'll actually modify that ui folder.

What I'm saying - they're not lying to themselves that they'll do that. They've never even considered it as being something to potentially do! They consider it as a package to use as-is. "A re-design later on? Who knows, by that time we'll have people who know their UI stuff, they'll figure it out. Whether that will be based on shadcn? Who cares, not important." I bet the premise that all these people are using shadcn with the idea of some day modifying it just isn't the most common reality.

Thanks. I think Zalo's only big in Vietnam, so we're at one example outside of China, where the rules have never been the same anyway. GGP's premise doesn't seem to hold up, though I can see the idea of banks getting exceptions being potentially more accurate.

Though if it's similar to Line I wouldn't even count Zalo either, as said, from what I've seen with Line it doesn't have an app-store-in-app, all functionalities are predetermined and preinstalled. It just does a lot of things, but that in itself doesn't go against any App Store rules.

> I feel like you underestimate the degree to which AI developers are influenced in just the same ways as everybody else.

I feel like you're overestimating that. I work every day on developing LLM-based products, use them a lot while developing those, and have been doing so since November 2022 (beta release of GPT 3.5). So I should be a prime representative of the demographic you're talking about.

When a new frontier model is released, I run a series of tests, play around with it, and decide whether or not to use it based on that. Anything else would be ridiculously stupid, leaving either cost savings, product quality improvements, or development velocity increases on the table. We're not talking about databases or cloud VMs where there's a big switching cost. It's changing a model string and provider, so frictionless that it's a no-brainer.

We actually have some data that's on my side. Take a look at OpenRouter's model usage statistics. Note how OpenAI mmodels make up a tiny share on there, despite having more general mindshare on social media than all other model providers combined. Their data also shows people are very quick to switch when improvements to price/performance are released.

It's a bit like how vacuum cleaner specialists never use Dysons despite them having the most mindshare and overwhelming marketing budget. Regardless of vacuum cleaner specialists also just being normal people. Even if this one doesn't hold in your region, there's countless of these examples.

> They don't have a design system, don't know what that even means, and their sinful hands should not be tweaking any UI libraries, at most being limited to modifying a rich set of theme tokens that force consistency

I feel like the majority of shadcn users doesn't tweak it whatsoever though. In fact - I bet you a large percentage has never even considered the ability to tweak it, just seeing it as a plug-and-play set of UI components. Think old Android and iOS apps, where almost everyone just used the default components.

> Apple has a lot of carve-outs and exceptions (see do-everything apps that contain mini app stores)

Which ones are you thinking of? Does Grab operate this way?

The China case is well-known, but that's really its own beast. KakaoTalk (Korea), while more of an 'everything' app than those in the West, is still a far cry from containing a mini app store. A user can't choose to add any new functionality by installing something - it's all included right from the get-go. My (limited) experience with Line (Japan, Taiwan, Thailand) is similar. So I'm curious if there's any non-Chinese apps you can name.

FWIW I'm not arguing against your fundamental premise, would just like to know which do-everything apps you mean.

> Every single medical provider today relies on Oracle.

All over the world? If not, then there's the replacements, although I'm sure they're not FOSS.

East Asia isn't on Whatsapp. In most of South-East Asia it's also not the #1. That makes for ~2 billion people.

Not trying to dismiss your anecdote, as both can very much coexist separately, and I actually think your analogy is spot on for LLMs! But it did make me think of something.

I once was in an environment where A got along with everyone, and B was hated by everyone else except for A. This wasn't because B saw qualities in A that no one else recognized; it was just that A was oblivous to/wasn't personally affected by all the valid reasons why everyone else disliked B. A to an extent thought of themselves as being able to see the good in B, but in reality they simply lacked the understanding of the effects of B's behavior on others.

> Even a 1 month project with one good senior engineer working on it will get 20+ different files and 5,000+ loc.

For what it's worth, I mostly work on projects in the 100-200 files range, at 20-40k LoC. When using proper tooling with appropriate models, it boosts my productivity by at least 2x (being conservative). I've experimented with this by going a few days without using them, then using them again.

Definitely far from the massive codebases many on here work on, small beans by HN standards. But also decidedly not just writing one-off scripts.

How much lines of context and understanding can we, human developers, keep in our heads, taken into account and refer to when implementing something?

Whatever the amount may be, it definitely fits into 300k tokens.

The narrative of this article is that Big Tech can and should be doing more, that they're the enablers.

But if you look carefully:

> The roots of “pig butchering” scams reach back to the offshore gambling industry that emerged from China in the early 2000s. Online casinos had become hugely popular in China, but the government cracked down, forcing the operators to relocate to Cambodia, the Philippines, Laos, and Myanmar.

China didn't wait for whatever tech was enabling them to start policing. They cracked down themselves, and clearly this is not happening in the lawless borderlands of Myanmar. That's the main enabler here, and the problem to solve. Big tech implementing proposed measures would at best be a stopgap measure that the scammers would find ways around.

> It is advertised as an autonomous software developer

By a few currently niche VC players, I guess. I don't see Anthropic, the overwhelming revenue leader in dollars spent on LLM-related tools for SWE, claiming that.

> The biggest thing I've found is that if you give any hint at all as to what you think the problem is, the LLM will immediately and enthusiastically agree, no matter how wildly incorrect your suggestion is.

I'm seeing this sentiment a lot in these comments, and frankly it shows that very few here have actually gone and tried the variety of models available. Which is totally fine, I'm sure they have better stuff to do, you don't have to keep up with this week's hottest release.

To be concrete - the symptom you're talking about is very typical of Claude (or earlier GPT models). o3-mini is much less likely to do this.

Secondly, prompting absolutely goes a huge way to avoiding that issue. Like you're saying - if you're not sure, don't give hints, keep it open-minded. Or validate the hint before starting, in a separate conversation.

> One, your projects are small enough that you can reasonably provide enough context for the language model to be useful. Two, you’re using the most common languages in the training data. Three, because of those factors, you’re willing to put much more work into learning how to use it effectively, since it can actually produce useful content for you.

Take a look at the 2024 StackOverflow survey.

70% of professional developer respondents had only done extensive work over the last year in one of:

JS 64.6% SQL 54.1% JTML/CSS 52.9% PY 46.9% TS 43.4% Bash/Shell 34.2% Java 30%

LLMs are of course very strong in all of these. 70% of developers only code in languages LLMs are very strong at.

If anything, for the developer population at large, this number is even higher than 70%. The survey respondents are overwhelmingly American (where the dev landscape is more diverse), and self-select to those who use niche stuff and want to let the world know.

Similar argument can be made for median codebase size, in terms of LOC written every year. A few days ago he also gave Gemini Pro 2.5 a whole codebase (at ~300k tokens) and it performed well. Even in huge codebases, if any kind of separation of concerns is involved, that's enough to give all context relevant to the part of the code you're working on. [1]

[1] https://simonwillison.net/2025/Mar/25/gemini/