现已上线！2025 年 Telegram 研究 — 年度关键洞察 
HN Best Comments
前往频道在 Telegram
Comments from https://news.ycombinator.com/bestcomments Source code: https://github.com/border-radius/hn-best-comments
显示更多4 193
订阅者
-324 小时
-57 天
-1030 天
数据加载中...
吸引订阅者
七月 '26
七月 '26
+5
在1个频道中
六月 '26
+60
在6个频道中
Get PRO
五月 '26
+64
在5个频道中
Get PRO
四月 '26
+137
在9个频道中
Get PRO
三月 '26
+61
在9个频道中
Get PRO
二月 '26
+93
在4个频道中
Get PRO
一月 '26
+134
在8个频道中
Get PRO
十二月 '25
+61
在6个频道中
Get PRO
十一月 '25
+126
在9个频道中
Get PRO
十月 '25
+227
在6个频道中
Get PRO
九月 '25
+429
在6个频道中
Get PRO
八月 '25
+1 000
在6个频道中
Get PRO
七月 '25
+249
在5个频道中
Get PRO
六月 '25
+45
在2个频道中
Get PRO
五月 '25
+48
在1个频道中
Get PRO
四月 '25
+25
在0个频道中
Get PRO
三月 '25
+23
在1个频道中
Get PRO
二月 '25
+604
在2个频道中
Get PRO
一月 '25
+303
在2个频道中
Get PRO
十二月 '24
+25
在5个频道中
Get PRO
十一月 '24
+24
在2个频道中
Get PRO
十月 '24
+711
在3个频道中
Get PRO
九月 '24
+59
在4个频道中
Get PRO
八月 '24
+60
在3个频道中
Get PRO
七月 '24
+53
在1个频道中
Get PRO
六月 '24
+47
在3个频道中
Get PRO
五月 '24
+65
在7个频道中
Get PRO
四月 '24
+88
在2个频道中
Get PRO
三月 '24
+133
在4个频道中
Get PRO
二月 '24
+110
在4个频道中
Get PRO
一月 '24
+153
在8个频道中
Get PRO
十二月 '23
+84
在2个频道中
Get PRO
十一月 '23
+130
在7个频道中
Get PRO
十月 '23
+122
在2个频道中
Get PRO
九月 '23
+58
在0个频道中
Get PRO
八月 '23
+107
在0个频道中
Get PRO
七月 '23
+67
在0个频道中
Get PRO
六月 '23
+78
在0个频道中
Get PRO
五月 '23
+334
在0个频道中
Get PRO
四月 '23
+326
在0个频道中
Get PRO
三月 '23
+51
在0个频道中
Get PRO
二月 '23
+36
在0个频道中
Get PRO
一月 '23
+294
在0个频道中
Get PRO
十二月 '22
+1 360
在0个频道中
Get PRO
十一月 '22
+71
在0个频道中
Get PRO
十月 '22
+36
在0个频道中
Get PRO
九月 '22
+55
在0个频道中
Get PRO
八月 '22
+43
在0个频道中
Get PRO
七月 '22
+89
在0个频道中
Get PRO
六月 '22
+203
在0个频道中
Get PRO
五月 '22
+42
在0个频道中
Get PRO
四月 '22
+23
在0个频道中
Get PRO
三月 '22
+23
在0个频道中
Get PRO
二月 '22
+38
在0个频道中
Get PRO
一月 '22
+47
在0个频道中
Get PRO
十二月 '21
+36
在0个频道中
Get PRO
十一月 '21
+75
在0个频道中
Get PRO
十月 '21
+44
在0个频道中
Get PRO
九月 '21
+14
在0个频道中
Get PRO
八月 '21
+46
在0个频道中
Get PRO
七月 '21
+71
在0个频道中
Get PRO
六月 '21
+172
在0个频道中
Get PRO
五月 '21
+287
在0个频道中
Get PRO
四月 '21
+15
在0个频道中
Get PRO
三月 '21
+16
在0个频道中
Get PRO
二月 '21
+28
在0个频道中
Get PRO
一月 '21
+34
在0个频道中
Get PRO
十二月 '20
+639
在0个频道中
| 日期 | 订阅者增长 | 提及 | 频道 | |
| 04 七月 | 0 | |||
| 03 七月 | +1 | |||
| 02 七月 | +1 | |||
| 01 七月 | +3 |
频道帖子
Re: Immich 3.0
I teach a free software development course to my undergrad students. It's really exciting to stumble upon one of the work they did for my class in the wild (it's the first listed bug fix — which is the last of the three pull requests this student got merged in Immich during my course). I feel so proud! :)
p4bl0, 19 hours ago
| 2 | Re: Costco is the anti-Amazon
> Even if you think it is preferable at an individual level, there are good reasons to question the social value of the logistical complexity that it necessitates. Home delivery of single-packaged items entails an entirely different cost structure than freight trucks driving to consumer-facing warehouses delivering entire pallets of goods to be driven home by customers themselves.
Ok, so 100 people can all drive to the store, or one delivery truck can drive to everyone's house. (Ignoring the packaging waste for a second,) I suspect delivery of single items cuts back significantly on trips to the store.
gwbas1c, 9 hours ago | 23 |
| 3 | Re: 60% Fable cost cut by converting code to images and having the model OCR it
In Gemini at least, if you look at how they process PDFs, they do an OCR and then feed the text + image to the model, without charging you for the text tokens (I believe).
So my guess is that Claude’s backend is doing the same — so this hack is probably more of a loophole in token accounting that might get closed if Claude is doing what Gemini does
aabhay, 7 hours ago | 172 |
| 4 | Re: Memorizing session transcripts isn't useful
Blog posts like this just blow me away.
> I believed this so strongly that my company built an entire product around this concept. I used to tell folks that "session transcripts were the new oil," that they were more valuable than the code itself.
> […]
> We don't really write code by hand anymore.
Honestly, isn't this just influencer spam? What possible value is there in reading about people who used to have products, but no longer write their own code, complaining about the inscrutable prediction machine they have handed that job and their livelihoods to?
Like, if you have complaints about the thing, perhaps you should address them to your supplier directly. None of your readers can help, and nobody's magic folk solution to your problem is better than yours.
And there are so many of these sorts of posts. Are we not entirely cooked?
(I think I have concluded that if people writing about AI aren't writing about interesting things they have achieved with small, local LLMs — which for clarity I am fully interested in reading - then I'm done reading. This whole blogging-about-cloud-AI genre is just weird and irresponsible now)
dofm, 4 hours ago | 325 |
| 5 | Re: Jamesob's guide to running SOTA LLMs locally
I play with local LLMs a lot. I've spent more on hardware than I should. I'm friends with a local group of people who have spent a lot more than I have.
The warning I would have for everyone is to temper your expectations and read the fine print carefully. The big build in article starts off with a $40K budget and then includes 4 GPUs that are $12K each. For those doing the math, this build is going to cost more like 50-55K.
Local setups also often rely on quantization and techniques like REAP to fit the models on their hardware. You will read a lot of claims that 4-bit quantization is lossless, but those claims come from KL divergence measurements on a small corpus. Use one of these 4-bit models on long context coding tasks and the quality will be noticeably less. Even for non-coding tasks like dataset analysis, I can measure a substantial quality difference between 4-bit models, 8-bit quants, and even some times the full 16-bit source.
This article is also encouraging the use of a REAP model, which means someone has cut out some of the weights to make it smaller. The idea is to remove weights that are less useful for certain tasks, but again this is going to reduce the overall quality of the output.
The trap is that people say "I'm running GLM-5.2 locally!" and it sounds amazing when you look at the GLM-5.2 benchmarks. However they're not actually running GLM-5.2, they're running a model derived from GLM-5.2 that discards most of the bits and drops some of the experts. It does not perform the same as what you see in the benchmarks. In my experience, the divergence between a quantized/REAP model and the parent model is unnoticeable when you try it on very small tasks or chat, but becomes painful when you start trying to use it on long-horizon tasks where little errors start compounding.
Then you get into the slippery slope of thinking you're $50K deep into this project, but what you really need is just one or two more of those $12K GPUs to use the next level of quantization that might improve the quality a little more and make your investment worthwhile...
Aurornis, 5 hours ago | 285 |
| 6 | Re: Why Switzerland has 25 gbit internet and America doesn't
This article came up before. It's heavy on the clickbait, if you couldn't guess from the title.
Some important points that they leave out:
- 25G internet isn't available everywhere in Switzerland. It's just the fastest tier available in some locations.
- The United States is 85 times larger than Switzerland. The entire country of Switzerland is the size of a small US state. Covering the US with broadband is much harder than Switzerland.
- 25G internet is also available in some locations in the United States.
- As another commenter discovered, the average speed test results of US and Swiss internet connections are pretty similar. The average Swiss person isn't connected to the internet faster than the average United States person.
Aurornis, 14 hours ago | 323 |
| 7 | Re: Half-Baked Product
One important part of the story is in the very beginning: The founder’s motivation. To become wealthy.
You see this in the startup world a lot. Founders with 5+ failed startups in different sectors, because said founder picked the fields mainly by doing some market analysis. Not domain expertise.
There’s then a big mismatch between what the founder thinks is possible, and what the domain expert thinks is possible.
The defense is of course that some people can do that - Musk did it, so why not?
Another defense is that blindingly naïve optimism is sometimes needed to move the needle, as the concept “that can’t be done” simply doesn’t exist to some people.
I’ve sat through some pitches like that, where it is very obvious that the founder/CEO has limited knowledge and expertise in what they’re pitching, where the product is limited, but their enthusiasm is off the charts.
EDIT: The very latest happened only a couple of weeks ago. A startup had reached out to my employer as they’re developing a platform in our domain. Higher ups liked what they’d seen, enough to arrange a real meeting.
Startup is only 3 months old, and the moment I opened the platform I recognized a vibecoded (likely using clause) platform identical to almost all other launched on a daily basis.
So I probed a bit about data sources, serious questions regarding security, etc. but the guy was pretty fluent in consultant (turns out he had worked as a management consultant before launching), and the CTO was just nodding along.
In the end they wanted our data, and promised the moon on features - but as mentioned, I’m sure the whole product was entirely vibecoded.
TrackerFF, 6 hours ago | 357 |
| 8 | Re: CarPlay Is Additive
Author says "I literally will not buy a car that does not support CarPlay."
From July 2022: https://www.cnbc.com/2022/07/22/apple-carplay-could-be-a-tro...
Apple engineering manager Emily Schubert said 98% of new cars in the U.S. come
with CarPlay installed. She delivered a shocking stat: 79% of U.S. buyers would
only buy a car if it supported CarPlay.
“It’s a must-have feature when shopping for a new vehicle,” Schubert said
during a presentation of the new features.
valgaze, 18 hours ago | 364 |
| 9 | Re: Alibaba to ban Claude Code in workplace over alleged backdoor risks, source says
[flagged]
p0w3n3d, 7 hours ago | 433 |
| 10 | Re: Zuckerberg 'Admits' Meta's Layoffs Were Ineffective
I wonder when (if ever) the companies realize that demoralizing your workforce (and destroying that sector of the job market) doesn't have only advantages.
I know plenty of people that reacted with the desired fear, putting in long hours to avoid layoffs, willingness to accept lower pay because the job market sucks, etc. - but I think there are also plenty of the the mythical 10x engineers that just checked out, stopped being 10x engineers, and are just collecting their paychecks and waiting for the layoff now. And I'm not sure you can "get them back", ever.
At least some companies reacted to this with more top-down management, stricter metrics etc. which kills motivation further and leads to metric optimization. Tell a good, smart, motivated engineer that you want more AI usage, and he's going to maybe start using some AI where it makes sense, but mostly ignore the metric while trying to do useful work. Demotivate the same engineer and make clear that his paycheck depends on metrics, and he'll give you what you're asking for, except https://github.com/dtnewman/burn-baby-burn is probably not what you _wanted_...
tgsovlerkhgsel, 4 hours ago | 438 |
| 11 | Re: Markets are competitive if and only if P = NP
The actual paper's title is
"Markets are competitive if and only if P != NP"
Seems that HN's auto-headline rewriting in this case has made a critical error :)
>Artificial intelligence, by expanding firms' computational capabilities, is pushing markets from the competitive regime toward the collusive regime, explaining the empirical emergence of algorithmic collusion without explicit coordination.
I have to dig more into the paper but I don't see how this follows, except in the most straightforward way. Basically, if everyone uses the same methods to derive price, of course there will be "collusion", or in other words, everyone will have the same price. But this doesn't seem like a result of compute per se, but simply better communication networks and information flows. You could have gotten the same result in medieval England by having everyone post their selling prices on the town square board.
Again, I haven't dug into the paper yet, but it seems like what really matters for firms is "compute"/$ (if the "compute" is an LLM or an assistant that has to go walk the 10 minutes down to the square makes little difference)
Edit: Isn't another implication of this, that increased compute -> collusion imply that increased compute -> communism becomes feasible?
I think this goes to my point above though, the primary problem preventing fully automated luxury communism isn't compute per se, but actually observing the information flows to make it possible. Capitalism famously solves this information problem through the pricing mechanism. So in effect, he's arguing that extra compute makes information gathering more efficient, and at the limit you get perfect information. Which, yeah, I guess so. Assuming everything can be perfectly measured, even theoretically.
xxpor, 2 hours ago | 415 |
| 12 | Re: ZCode – Harness for GLM-5.2
I'm somewhat surprised that this is not open source (from what I can tell). Compare to Mimo Code https://github.com/XiaomiMiMo/MiMo-Code (which is a CLI, while this is a desktop app).
seizethecheese, 2 days ago | 458 |
| 13 | Re: Why jet engines aren't made in China
The more parsimonious explanation is that commercial jet engine production is downstream of commercial airbody production and China's currently limited by COMAC's scaling woes. All the money and talent in the world can't replicate real users generating real data that you can use to improve.
I'd argue the opposite that jet engines have a market structure that's uniquely terrible for traditional free market societies. There's a few industries where structurally, companies can only exit the market but it's almost impossible for a new company to enter. Airframes, jet engines, CPU manufacturing, lithography etc.
What this dynamic doesn't make any company immune from though is corporate rot. You've seen the rot take down Boeing and Intel from the inside as a slow moving car wreck. There's no reason the rot can't take down ASML, TSMC or Airbus as well. The free market fundamentally doesn't have a good response to this problem, excess capital is taken out of these companies during good times and then they run to governments seeking bailouts during bad times but governments don't know how to mandate good corporate governance.
I think a lot of the jet engine manufacturers are seeing this same corporate rot process, the number of high profile scandals across the industry and reports of insiders on how the number crunchers are taking over the business are strangely reminiscent of what we heard out of Boeing and Intel.
shalmanese, 1 day ago | 463 |
| 14 | Re: Android Developer Verification: Threat masquerading as protection
It doesn't solve the current issue, but in case we don't manage to push back on this, some people might not know that there are various actual linux OSes for mobile:
- SailfishOS: still linux based and seems fairly community inclusive, but the UI part of the stack is closed source. Is the only one officially allowed to run android apps, via emulation. Has existed for a very long time, it's lightweight and I think the most stable/bug-free in this list.
- Ubuntu Touch: fully open source and community driven, it uses snap packages for security, you might be able to run android apps. Last time I run it also seemed fairly stable/bug-free.
- PureOS: fully open source and privacy focused. I think it's the only one that, released with the Librem 5, can avoid using proprietary blobs for interfacing with the hardware. Seems less stable than SailfishOS and Ubuntu Touch. You would need to buy a fairly expensive-but-old phone(librem 5) to run it.
- PostmarketOS: fully open source, focused on being lightweight and revive old phones, has a huge amount of phones it has been tested on, is based on Alpine.
- Mobian: mobile version of Debian, it's fairly new on this list.
There are many more linux mobile OSes, but as far as I know these are the main ones. There might also be some inaccuracies on this post, I tested some of these a long time ago, and I never actually run the last 2.
sambuccid, 1 day ago | 470 |
| 15 | Re: Reality has a surprising amount of detail (2017)
The first time I built a freestanding bookshelf, I put a lot of effort into making the feet level and the back straight and at a right angle to the feet. Once I put it up against the wall I'd built it for, I realized I'd solved completely different problem than the problem I really had. I needed crooked bookshelf, since the wall was totally tilted.
In the end I screwed some wall shelves in and called it good enough.
WastedCucumber, 17 hours ago | 437 |
| 16 | Re: CarPlay Is Additive
For pure interaction, CarPlay as a generic solution is very hard to beat infotainment systems that are deeply integrated with the vehicle itself. Its advantages are mainly these:
- Consistency. You get into a new car, all you need to figure out is how to open CarPlay, no need to learning a completely different and often complicated infotainment system.
- "It's on your phone". You can decide what playlist to play or where to navigate before you even get into the car.
- Stays up to date over the long term. Just look at cars from five years ago. Most built-in infotainment systems are still stuck in that era, no matter how smart they once looked. CarPlay uses your phone as the main computing platform, and the car's infotainment system only needs to act as a thin client for I/O, it keeps updating with iOS.
linzhangrun, 7 hours ago | 464 |
| 17 | Re: Half-Baked Product
Oh, I'm glad I don't work in the oven business. We're just starting a stealth startup that's revolutionizing dishwashers, and the prototypes are amazing. They use less water, less detergent, and this weekend we're hoping to solve the last remaining issue: occasionally, they break glasses.
sixtram, 3 hours ago | 523 |
| 18 | Re: Since Linux 6.9, LUKS suspend stopped wiping disk-encryption keys from memory
Yes, if you simply suspend your laptop on most stock Linux distributions, then everything including the master key is still kept in memory. But Debian pioneered the (optional) cryptsetup-suspend addon. This issues a luksSuspend command which is supposed to wipe the key from memory, and on resume asks you to resupply your passphrase.
Up to kernel 6.8, this worked as described; starting with kernel 6.9, it silently didn't.
IngoBlechschmid, 19 hours ago | 565 |
| 19 | Re: CarPlay Is Additive
Surprised no one's mentioned it so far, but CarPlay / Android Auto aren't just features, they're consistency. Across makes, models, years. I know what I'm getting when I connect my phone - and if anyone uses my car with their own device, they get their own dashboard, as well. One interesting use case I saw was a couple where one used a left-to-right interface and the other a right-to-left UI. CarPlay makes this easy, because the interface is linked to your own personal device.
faitswulff, 7 hours ago | 563 |
| 20 | Re: Exapunks (2018)
For those who don't know -- while Zachtronics is no longer making games, Zach Barth is still active now under the company Coincidence Games. They just game out with a spacecraft engineering puzzle game:
https://store.steampowered.com/app/2536720/UVS_Nirmana/?cura....
jawarner, 11 hours ago | 612 |
