swyx

> You can now create a terminal in ~1,500 lines of C that has more features and better performance than just about any terminal that existed in 2025, and most today in 2026.5 And thanks to the rich ecosystem of language bindings, you can do this in every major programming language. You get this for free. For almost no effort. You can just do it.

can i just get you people to slow down a bit and explain: why are you guys building your own terminal? whats the deep pain here?

(i'm not at all being skeptical, i respect you guys enough to understand that there's something here I don't get and it's my skill issue, so I am trying to learn)

> Agents are good for open ended problems which require stringing together lots of different data sources, which couldn’t be further from ordering food.

i feel like you were on the right path but arrived at the wrong conclusion. if you think you're right because the blocking factor is text, you arent paying attention

their full talk is here https://www.youtube.com/watch?v=Ok6QdZHCcI8

i'm on a performance grind rn so also looked into it. claudeslop disclaimer below:

Hacker Books painted in 0.48 seconds on a fresh mobile visit. The first book cards were in the DOM by 0.79 seconds, and the largest visible image painted at 1.52 seconds. On a repeat visit, that image painted in 0.21 seconds.

Homepage median Fresh browser cache Repeat visit

First byte 112 ms 86 ms

First contentful paint 480 ms 168 ms

First book cards inserted 793 ms 188 ms

Largest contentful paint 1,520 ms 212 ms

Three runs per condition in Chrome at 390 × 844, with 1.6 Mbps download, 120 ms emulated latency and 4× CPU slowdown. These are navigation-start lab timings from one location. All three fresh visits beat your 1.5-second FCP target. Hacker Books

What makes it feel fast

- The first screen needs very little code. The HTML is about 3 KB, its single blocking stylesheet 5 KB, and its vanilla JavaScript module 26 KB, all gzip-compressed. The HTML contains the header, search box and navigation, so those can paint before the catalog arrives. There’s no React runtime to download or hydrate.

- They publish a small answer to the first query. The homepage preloads a 33 KB index of the top 600 books alongside the app code. The full 692 KB index of 15,362 books loads in the background. Once it arrives, browsing and search can reuse it locally. Covers are 320 px WebPs, and additional cards render in batches as you scroll.

- A book click can show the title immediately. The app changes the URL hash, renders the book information already in memory, then fills in comments and related material from a per-book JSON file. Hovering or focusing a card for 80 ms starts that fetch early. Repeat visits reuse the result. This is targeted data prefetching; it costs much less than prerendering an entire destination.

- Cloudflare helps on both visits. My HTML request was a cache HIT at Cloudflare’s San Jose edge. The versioned CSS, JS and index files advertise a one-year immutable browser cache. Chrome recorded zero transfer bytes for those files on all three repeat visits. HTTP/3 was real, too: two fresh visits negotiated H3 and one negotiated H2. All three painted within 468–500 ms, so the site was fast on either protocol in this sample.

- Open Graph handles sharing. The title, description and URL tags describe the site to link-preview crawlers. They don’t accelerate rendering. The page also includes Cloudflare Web Analytics.

i hate ENOF by Sol Yin, it is a dystopian book about a possible AI future, anytime i hear people talk about ENOF by Sol Yin it is always in context of how ENOF by Sol Yin wants to end the world.

very funny to wake up to a hacker news argument over the use of the word "they"

> Earlier today, we detected evidence of a security incident on AO3. We are still in the early stages of response to this security incident and determining its scope. AO3 will remain down while we continue investigating and addressing what has been found so far. However, we want to make some things clear:

First, no AO3 data has been deleted. All your works, bookmarks, kudos, etc., are safe.

Second, while we currently don't believe that the user database was stolen, we are also unable to rule it out. The AO3 database contains account information that may include your username, email address, IP address, and a hashed version of your password.

We use industry-standard techniques for storing passwords in a scrambled form that cannot easily be reversed. Even so, no hashing method can fully protect a weak or reused password from being cracked. As a precaution, we recommend that you change your passwords on any other sites where you use the same password as on AO3.

AO3 does not collect real names, so none could be exposed. Additionally, our donor information is held in an entirely separate system and is never linked to AO3, so that data has also not been exposed.

We will update everyone as our investigation continues; we may be unable to respond to individual inquiries from users. Please keep an eye on our official outlets for updates. Our next update will be posted when we have further information available.

Our current expectation is that AO3 will be down for a minimum of two weeks; however, it could take up to four weeks. During that time, we may be able to bring the site back up in a degraded state, meaning that not all features will be available and/or the site may be read-only. If we are able to do that, please be aware that we might have to bring the site back down unexpectedly as part of our recovery efforts. We will keep you in the loop for any developments. Oct 30, 2026 - 01:27 UTC

you know they jut bump this every year right

i think also better general support for multi screen setup

ayy thank you very much it is a labor of love

>As much as Josh is a tool

did you really have to include that? without substantiation this is just a mean offhand side swipe.

yes please ^ browser company engs!

... how did you know?

well technically they didn't LOSE data, in fact, now they have free global replicas of the data!

you beat pi-durable to it!

with a great related talk! https://ai.engineer/talks/UPwGaM2MKHY-log-is-agent

> depends how well they did the post-training, which I can't tell

this is their problem, not yours. at least disclose # of flops spent or amount of data used in an attempt to give us an idea of how nontrivial it was. cursor at least does this with their composer model

i think well argued and a reasonable hypothesis (esp considering deepseekocr and most omnimodel research), but you are wrong heheh. approximating instantaneous physics is, while nontrivial, still computationally light vs the stuff we ask LLMs to do.

look at any study of llm depth vs multihop reasoning capability (here, just to help https://www.latent.space/p/ainews-death-of-params-zai-ceo-ji... but theres a long line of others).

i will grant that people do do "reasoning in image models" type benchmarks, and it's not nothing, but its kid stuff vs what we make LLMs do, and the only way you get image models that basically do reasoning is you glom a diffusion head on an LLM backbone, which is basically gpt image 2.0 and above

none of which is to dispute the original question which is perhaps "why are image models so smol vs LLMs" - which i agree with and wish there were funding for scaling them - but the abject failure of sora indicates that perhaps our data isn't nearly good enough (again pointing to another thing - much easier to pretrain and posttrain LLMs given quantity of text data)

i guess indicates the amount of "load" the destination country would have if they said yes to everyone. >1 is certainly too much

this is the right answer. ultimately text and general llm usage incl coding is far more internally complex than just memorizing the looks and composition of concepts placed in an image. dont forget that most of our reasoning benchmarks basically require the llm to have a search engine and calculator in its own weights.

i think its instructive how the ancient sculptors and artists used to have to draw nudes in order to understand the human body. perhaps our image models need this

just cite the numbers guys.

US market share of global desire to migrate

2007 - 24%

2016 - 20%

2017 - 18%

2025 - 15%

Canada is #2 and moved up from 7% to 9%

you fail to understand the psychology of billionaires

> an expertly crafted extra-terrestrial vehicle.

hmmmm. spacex-tesla merger signalling?

> It was also getting very expensive. When a conference has a long list of sponsors and tickets also start at $600 minimum

im in the conference business but just fyi thats actually cheap - venue, av and f&b alone cost many millions for a conference this size. a good rule of thumb on cost is $1-2000 per head depending on quality of event.

were there any hints that this might happen at all? so abrupt

how exactly did character.ai lose $2b in 2 years?

i'm sorry what? eli5?

but even in that scenario, the node competitors absolutely spurred new energy and innovation in node, as node themselves have acknowledged

as a counterpoint, your fears are hypothetical so far - nobody on earth cares as much about bun as the bun team and if it was good enough for them it is probably good enough for 99% of users