> But I'm beginning to wonder whether that will even matter in a few months.
At that point, even Bun itself doesn't matter. All intermediate tools don't matter if LLMs can reliably write something large.
The problem is that LLM is not quite there yet. The rewrite was only possible because they mostly stick to 1-to-1 translation resulting in non-idiomatic Rust code. So, what from there? I don't think they can really build up a sane codebase from that state. They only shot themselves with a bigger gun.
I've only posted one question to SO, and it was enough to dislike the whole platform. LLM didn't kill it. SO killed itself with the broken community structure.
It's ×(multiplication), not +(plus). Interface is not simply overlaid on Information; it actively changes how information is perceived and used, thus User Experience (UX).
Users have limited screen space, attention time, context retention, etc. So, apps must be wise about what information matters at any given moment, and make the best out of those limited resources.
Developers have been crazy about this: A/B testing, CVR, retention, churn, LTV, ARPU, DAU/MAU, North Star Metrics. Deploy and analyze, develop and optimize, rinse and repeat, and apps end up as revenue-generating machines.
A side effect of this optimization loop is that, apps become a designed thinking process for users. Apps decide what to show, what to hide, what to emphasize, and what comes next. They actively shape how people see and think, all to lure them into spending money.
So, "this could have been a webpage" misses the point of what apps are for, and, by extension, who apps are for.
Still, I see a bit of potential here. Document is a natural user interface -- almost all apps, including even SPAs, have document-like or document-driven views. Perhaps we've been too obsessed with computer-program-like UI. Documents can always be dynamic and interactive without being overdesigned. Perhaps this is what folks wanted to point out.
Yup, and calling out a human-written article as AI-generated would be a serious insult. AI-flagging would incur bigger damage to the community than just having AI-generated contents around.
Sure, that is a case, but I mainly wanted to quote INTERPOL (the international police):
> Criminals behind pirate sites can be part of organized crime groups. They can use the proceeds to fund other illegal activities, such as illegal online gambling, online sexual exploitation, drug trafficking, arms smuggling, and money laundering.
Another notable example would be Operation KRATOS 2 by EUROPOL[1]. The case was roughly the same — crime organization pirating video content. IPTV piracy is certainly a lucrative business, because videos are relatively more difficult to handle for normal people.
1. Piggybacking established brand names (Postgres + Rust)
2. … without practicality nor advancement (e.g. this solves no extra problems)
3. … without trust (i.e. LLM-driven rewrite, with no capabilities to thoroughly review it)
I think people get easily upset when the title has high-signal names like Postgres, and the title touts it somehow, yet it’s obviously impractical for obvious reasons (short-/long-term practicality, social trust & network effect, etc)
A dirty secret is that piracy is being abused by criminal organizations[1]. When people unknowingly access such sites to see contents for free, it generate ad revenue for those organizations, which can fund other crimes.
Corporations acting as if naive is a bit of problem in reality. For one thing, CF is probably the largest entity serving pirated content internationally while hiding the identities of actual perpetrators for privacy.
Same here: CF is basically giving malicious actors an ability to ship contents/data publicly while laundering the legal responsibility of those actors.
I’m quite sure there is a certain amount of correlation unfortunately, mainly because there are micro patterns (e.g. IO, allocator) that can’t be modularized into functions. Lots of manual copy-pasta.
Just a side note: prompts often get a disproportionate amount of attention. That is, when you copy-paste an error message into the prompt, the LLM will focus on pleasing you immediately by fixing the error message, rather than understanding and fixing the underlying issue.
A better workflow would be to let LLMs directly access the same verification tools you use. This allows LLMs to observe failures during the loop and incorporate the info more organically, without giving failures too much attention priority.
The above is based on my own experience. LLMs perform better in a positive context (e.g. constructive thinking, building outward, what to do) than in a negative one (e.g. restrictive thinking, carving context inward, what NOT to do). LLMs themselves are designed to be defensive & negative, but they get easily confused under lots of prohibitive rules. LLMs are good at expansive exploration, but suck at verification and pin-pointing what you want. (I'm not sure whether it's related, but this mantra is also true for image generation using Stable Diffusion)
> The fix is to manage the consumption and the sources. …
> Containing news consumption to defined windows of time …
> Choosing depth over volume
Golden.
TBH, we must concentrate on what matters to us. When people cross that boundary, they not only hurt themselves, but end up hurting someone close by for issues from far far away.
> Then, run the frame analysis pipeline, which will divide the video into separate video scenes (1s each, or 1fps)
> (…)
> Frames analyzed 57,537
Aha, it makes total sense. This number sounds much more reasonable than “669 GB”, since the actual total size of processed frames would be like 10-30 GB.
(Not downplaying anything. Doing-at-home always requires some math on practicality)
> Total compute time 67h 40m 42s
I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
In fact, Opus does the same. It finishes the job, and redo it from scratch before presenting the result to the user. This happens even for simpler writing tasks especially when I instruct it to create a text file.
Yeah, you got a very good point there. Yes, it may seem very weird, but that’s because China was not playing the game.
The game that I suggested earlier assumes that (1) those in power seek to maximize their gains, (2) and such behavior is NOT aligned with social gains. So, basically, a never-ending arms race between the authority and the people. (check Leviathan by Thomas Hobbes if you’re interested)
However, China (and also South Korea) got a weird alignment in interests between the authoritarian government and the people — both of them somehow sought after national economic growth. Since their interests are already aligned, they didn’t play the game that I suggested above.
I think such alignment cannot be reproduced through games nor social interaction b/w powers — rather, it is more of a humane part of the history.
I think the caste system is what’s hampering social and economic progress in India (or at least partially).
It’s a sort of game theory stuff here. A structured class hierarchy makes it inherently more difficult for individuals to challenge the authority in power, even under the democratic government. The system imposes an additional risk of social backlash/punishment/retaliation for anyone who attempts to disrupt the established order, thus people have more reasons to stay back. This kind of risk is largely absent in more egalitarian/classless societies.
I think seeing it as an “align”-ment problem puts too much blame on the local side. Also, autonomy has nothing to do with the problem of misalignment.
In authoritarian systems like China, mis-alignment with authority can carry serious political and social risks, so people are easily pushed toward dishonesty. What happened under the Mao’s rule is simply this; local officials were too afraid of criticizing the very father of the revolution, which could be interpreted as attacking the legitimacy of the revolution itself. It was a side effect of over-concentration, and gaining more control over local would have not made any differences.
Deng was successful only because he was exactly aware of this problem. In his speeches on the government reform (the Open-Door policy), he explicitly pointed out over-concentration as a major issue. He not only eased the concentration of power, but also redesigned the incentive structure, so that officials can adopt objective measures and even try their own experiments.
I’ve been trying this on some social and political science topics, and, tbh, Opus is never going to replace ChatGPT in terms of sharpness of claims. Claims generated by Opus alone are almost always dull and inadequately nuanced. It’s also heavily influenced by popular opinions that include common misunderstanding, preconception, and vague language. The context gets contaminated too easily by external texts, so I have to carefully control and craft the input to Opus.
So my pipeline is mostly stuck with: (1) brainstorm with Gemini (2) plan with GPT (3) augment the plan with Gemini (4) execute the plan with Opus.
Perhaps, given all the materials filtered and prepared, Opus is good for verification.
Last time i checked their issue tracker (in 2025), the main source of problem was the engine, not their Zig code. A lot of core dump was happening inside and around JSC.