HNHacker News
TopNewBestAskShowJobs

swyx

24,625 karma · joined December 9, 2015

Writer/Angel in devtools and smol AI enjoyer.

ai engineer pod/news: https://latent.space/

conference: https://ai.engineer

devrel/devtools: https://dx.tips/

personal blog: https://swyx.io/ideas

book: https://learninpublic.org/

but you can also email username [at] ai dot engineer

---

meet.hn/city/37.7792588,-122.4193286/San-Francisco

Socials:

- x.com/swyx - youtube.com/@swyxtv

submissionscomments
swyx··on Ultramarine 1 Flash
> 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

swyx··on FLUX 3 Image Klein
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)

swyx··on Desire to Move Permanently to U.S. at New Low
i guess indicates the amount of "load" the destination country would have if they said yes to everyone. >1 is certainly too much
swyx··on FLUX 3 Image Klein
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.
swyx··on FLUX 3 Image Klein
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
swyx··on Desire to Move Permanently to U.S. at New Low
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%

swyx··on Tesla Roadster
you fail to understand the psychology of billionaires
swyx··on Tesla Roadster
> an expertly crafted extra-terrestrial vehicle.

hmmmm. spacex-tesla merger signalling?

swyx··on Grace Hopper Celebration Cancelled
> 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.

swyx··on Grace Hopper Celebration Cancelled
were there any hints that this might happen at all? so abrupt
swyx··on Character.ai has initiated voluntary Chapter 11 bankruptcy
how exactly did character.ai lose $2b in 2 years?
swyx··on Unikernels were hard. key word: were
i'm sorry what? eli5?
swyx··on Cloudflare acquires Deno
but even in that scenario, the node competitors absolutely spurred new energy and innovation in node, as node themselves have acknowledged
swyx··on Cloudflare acquires Deno
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
swyx··on Margaret Hamilton has died
[flagged]
swyx··on No new Ghostty updates since March
superlogical was announced in july https://mitchellh.com/writing/superlogical
swyx··on Homa: The end of TCP for AI clusters [video]
(talk curator here) we also transcribed with a light summary https://ai.engineer/talks/eZ8WWZzoaR0-homa-end-tcp-ai-cluste...
swyx··on Homa: The end of TCP for AI clusters [video]
curator here - was very honored that John came inbound to do this with us. AMA on the behind the scenes. note: we transcribed and summarized the talk here as we now do for all talks - https://ai.engineer/talks/eZ8WWZzoaR0-homa-end-tcp-ai-cluste...
swyx··on Cloudflare K2: serverless event streams
https://swyx.io/cloudflare-go :)
swyx··on Jev: System One Models for Prod, Not God – With Diogo Almeida, CEO, TypeSafe AI
literally no one red it lol
swyx··on Show HN: JevBench, a reproducible benchmark for typed decision models
jev ceo on why he eschewed benchmarking: https://www.latent.space/i/216783460/privacy-benchmarking-an...
swyx··on Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
> Now its collecting data to make a domain specific BERT and do what Jev does.

this misses the point of jev somewhat - the point is that this is a foundational, general purpose classifier model - see some good sources https://x.com/mparakhin/status/2101683565520199887?s=12

swyx··on Jev: System One Models for Prod, Not God – With Diogo Almeida, CEO, TypeSafe AI
hi HN, i know there's a lot of content about Jev, but this is as far as I know the first in depth technical interview with Diogo, creator/co-founder and CEO. hoping this is a valuable/valid contribution to the HN discussion!

AMA on any threads we left unanswered.

swyx··on Claude Code now reads AGENTS.md if there is no Claude.md
except they have similar pretrain/rlhf data which is the thing u really want to tune for
swyx··on Claude Code now reads AGENTS.md if there is no Claude.md
fwiw, i am with thariq https://x.com/trq212/status/2092302273099796842 in that prompts should be tuned for models and in fact blindly applying agents.md is probably an antipattern unless you want all models to basically converge to some common ill defined of instruction following - good local minima, bad global minima for model diversity and exploration of intelligence.

aka, sometimes it really is too early to force a standard

swyx··on How GLM built its own inference infrastructure
well, yknow, apart from his vested interest in having 1.4 billion more people to sell GPUs to
swyx··on Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents
kudos making the core open source. if you didnt do it, i wouldve. good luck
swyx··on Why I didn’t sign the Fields medallists’ letter
why does a fields medallist not have enough money to have a blog that isnt festooned with ads? good god wordpress.
swyx··on Hister: A private search engine for the pages you visit and the files you keep
is this like a pihole? is there a design difference you are going for here?
swyx··on We got admin access to Baseten's production GitHub
> Baseten handled this well. The timeline was:

> July 13, 11:10 PM: I reported the live basetenbot token, the public Harbor project, and the repository permissions.

> July 14, morning: Baseten made the Harbor project private. I flagged that the token itself still worked.

> July 14, 4:34 PM: Anton from Baseten Security confirmed the issue as critical and said they had made the Harbor project private and rotated the token. He also asked us to securely delete the images we'd pulled.

> July 14, 5:05 PM: We confirmed deletion and sent over two lower-severity findings from the same scan.

> July 17: Baseten closed out the remaining findings.

> September: We let Baseten know we planned to disclose the finding publicly and sent them a draft of this post.

They also sent us some T-shirts and sweatshirts as a thank-you for finding this critical bug.

well done all around. i think my only open question is what default security boundaries should all vibecoded internal agents follow as a learning we can take from this

← PreviousPage 2 of 34Next →