pngwn

I didn’t even know there was going to be an election when I arrived here. As I learned more I really didn’t really know what was going to happen. About 20 minutes ago the city exploded. Belgrade is jubilant. The energy is spectacular.

ok, its time to try my way now. what if we just trusted one another and promised to be good?

This is so coool. I've always loved mermaid but the bundle size and performance annoy me.

kepano@stephango.com

This is Pele, my new rendering library for diagrams and charts, compatible with the latest Mermaid syntax. - 100-500× faster than mermaid.js - 6× smaller bundle - No dependencies - Easy to theme - Responsive to device width - Open source (MIT) Try it at pele.run

{🧪} +paoloricciuti.svelte

Do iiiiiiiiiiit

👻 rossipedia 🎃@rossipedia.com

Hey yo @bsky.app you should render embedded code blocks in our posts: ```html <p>Like this</p> ```

pngwn reposted

@paolo.ricciuti.me

People that use @pi.dev as their harness..we want to ship a Svelte plugin for it but since it doesn't natively support subagents, which option do you prefer? a. We don't ship subagents for `pi` b. We expose a tool that spawns a new pi instance c. We assume you have `pi-subagents` and configure that

pngwn.ata place to gather moss

Long overdue, but I finally have an internet home.

lea.verou.meDark mode toggles: two states are enough • Lea VerouYes, the underlying model must have three states, but one is always irrelevant to the actual user goal. Users do not seek out solutions to problems they don’t currently have. A lot of the hate towards...

I think this a real great discussion and I think the core argument is probably correct. What I do think would be helpful is communicating that flipping dark -> light in system light mode is a return to system. Something that communicates dark -> back/ reset to light.

Lea Verou, PhD@lea.verou.me

🔥 New blog post: Dark mode toggles that expose three states (system, light, dark) seem like the best option. Until you think about user goals, that is. Before rushing to disagree, hear me out. lea.verou.me/blog/2026/da... PS: As a data point, @jakearchibald.com was converted after reading this 😅

svelte.devSvelteKit 3 is hereGet it while it’s hot

ITS REAL!!

Svelte@svelte.dev

svelte.dev/blog/sveltek...

Just released a new twinkleplop version with some significant performance improvements across both tokenisation and rendering! I finally had time to go back and explore a few other ideas. I was also slightly annoyed that I wasn't faster on all sizes. twinkleplop.pngwn.at/docs/benchma...

pngwn@pngwn.at

After about a year, on and off, I am veeeery happy to release TWINKLEPLOP, a syntax highlighter of tremendous power! The speed of prism with the customisation of Shiki. twinkleplop.pngwn.at

mdsvex.pngwn.atmdsvex - svelte in markdownCombine svelte and markdown in the same file. Live your dreams!

Added some docs so you can use twinkleplop with mdsvex! mdsvex.pngwn.at/docs#with-tw...

pngwn@pngwn.at

After about a year, on and off, I am veeeery happy to release TWINKLEPLOP, a syntax highlighter of tremendous power! The speed of prism with the customisation of Shiki. twinkleplop.pngwn.at

pngwn reposted

Sean Lynch

@techniq.dev

📦 npmchart.com: dependency analysis Analyze package dependencies as a graph, treemap, sunburst, or flamegraph including how each dependency affects install size

Sean Lynch@techniq.dev

🎉 Introducing npmchart.com Analyze npm packages (downloads, releases), Github activity (releases, stars, commits, issues, PRs), compare packages, user pages, download or copy any charts, and more!

After about a year, on and off, I am veeeery happy to release TWINKLEPLOP, a syntax highlighter of tremendous power! The speed of prism with the customisation of Shiki. twinkleplop.pngwn.at

View full thread

One fun feature is 'fidelity', you can decide what granularity of highlighting you want. There is much discussion on whether we highlight too much or not, so if you think the default is too granular, turn them off!

Go forth and twinkle.

After about a year, on and off, I am veeeery happy to release TWINKLEPLOP, a syntax highlighter of tremendous power! The speed of prism with the customisation of Shiki. twinkleplop.pngwn.at

twinkleplop.pngwn.atBenchmarks · twinkleplop docsHow fast twinkleplop highlights compared to other JavaScript highlighters.

It generally outperforms other highlighters depending on the language and source size, and comes in at ~30-40kb, though i hope to get this down quite a bit. twinkleplop.pngwn.at/docs/benchma... It is synchronous, and the whole API is available in any environment!

We have ~25 languages and 8 themes right now. I'll be adding more in time. Importing a language is really straightforward, its all standard modules, so anything that understands JS modules will work fine.

14 Sept

Remewbr when we just did rectangles? Miss those days.

14 Sept

Can’t even type correctly even with autocorrect.

There is so much wrong with the current rhetoric around ‘alignment’. None of it makes any sense at all. First of all, the just isn’t really any research to support an extinction event or even significant negative outcomes for humans. And the doom part of p(doom) is poorly defined.

We have manufactured the extinction of countless species without blinking, all of a sudden if it happens to us it is some kind of cosmic injustice? I understand we are wired to survive, I’m quite keen on it myself, but in the scheme of things it’s just nature taking its course.

And there is a bittersweet irony that, for greed, we created the thing that does to us what we have been doing to others for years now.

Maybe doom means we lose supremacy, but we don’t have an inalienable right to supremacy. We have failed just about every stewardship test. You don’t need to get abstract with whales and forests, we don’t even look after our own species. We’ve failed at every step.

If you’re gonna create a vastly superior entity, what is the argument for it remaining subservient to us? We don’t have some moral right to be in charge. We don’t even have a moral right to continue existing as a species.

But let’s assume the robots get real smart and don’t like us. They have a few characteristics: - don’t age/ not temporally bound - can replicate (almost) infinitely - vastly more capable - not emotionally threatened by humans

Why would they urgently eradicate 8 billion apes? That sounds costly. They could wait us out. Phase us out. Create the circumstances where we phase ourselves out. At any point on that spectrum we may cease to be an issue.

pngwn reposted

kepano

@stephango.com

This is Knap. It's a new language I created that turns data into Markdown. Knap should feel familiar and comes with a wonderfully pleasant syntax to modify and format plain text.

9 Sept

Why is Paul McCartney on Albanian cigarette packs?

You shouldn’t need to be an ML expert to have an ML idea. Today we’re launching ML Intern in HuggingChat. Start with a conversation. Finish with deployable artefacts.

When training is finished, ML Intern can push the results to the Hub, generate a report, or even build and deploy a demo in Spaces.

huggingface.coWhat the (micro)duck? - a Hugging Face Space by pngwnDetect the presence and location of microduck in any image

The example in the demo video comes from a real trace with real artefacts. It finished end-to-end in about 6 hours and cost less than $0.50. Dataset, model and demo app: huggingface.co/datasets/png... huggingface.co/pngwn/microd... huggingface.co/spaces/pngwn...

You shouldn’t need to be an ML expert to have an ML idea. Today we’re launching ML Intern in HuggingChat. Start with a conversation. Finish with deployable artefacts.

After you’ve chosen something to build, ML Intern estimates the compute required and recommends a budget. Once approved, that budget is enforced by the system rather than editable by the model.

From there, ML Intern can create repos, generate datasets, train models and monitor progress as jobs run. Every training run gets a Trackio dashboard so you can follow along.

You shouldn’t need to be an ML expert to have an ML idea. Today we’re launching ML Intern in HuggingChat. Start with a conversation. Finish with deployable artefacts.

You can start with a specific ML task or just have a conversation about what’s possible within your constraints.

ML Intern will research the topic before building. It has access to @hf.co, @github.com, the web and compute, so it can ground its work in real models, datasets, papers and tools.

You shouldn’t need to be an ML expert to have an ML idea. Today we’re launching ML Intern in HuggingChat. Start with a conversation. Finish with deployable artefacts.

Try it here: hf.co/chat/

t.cohttps://hf.co/chat/

We have thousands of techniques, and millions of models, datasets and papers, but even simple experiments often require expert knowledge. ML Intern lets you take advantage of the accumulated knowledge on @hf.co while focusing on your ideas.

tante.ccThe function of LLM-based math "proofs"There has been quite the uptick in “AI” labs’ claims of their LLMs “solving complex math problems”. I wondered why that is, especially given that LLMs are not exactly built for the kind of rigor that ...

7 Sept

I see a lot of conjecture on what is going on inside AI labs and attributing malice or cynicism where there isn’t any direct evidence for it. There is no doubt that the leadership saw a marketing opportunity in these cases but that isn’t why the work is done. tante.cc/2026/09/07/t...

The narrative setting and framing is an issue, especially with OAI and Anthropic, but it is not the sole reason that research like this exists. That tends to come from a genuine experiment run by a genuine expert.

7 Sept

Whether the research has merit is another issue. Closed labs have openness issues that can make research difficult to replicate. They also have absurd resources which is another barrier to reproduction.

Secondly, researchers in AI labs have a lot of autonomy when it comes to running experiments and figuring out how to test model capabilities. It is common to have domain specific researchers who spend all of their time just focusing on a single application.

7 Sept

These experiments usually start small, and most come to nothing, but where there is promise they can quickly ramp up.

First of all AI labs do care about math. Researchers love math and they did this work. What is more math is a good domain for experimentation because it is easier to model success than in some other domains.

openai.comAn Alien MindJakub Pachocki reflects on increasingly capable AI and the challenge of keeping it aligned. He calls for stronger safeguards and international coordination.

If the emergent intelligence of AI is becoming sufficiently complex that we cannot full understand it, believe it could lead to recursive self improvement, and are increasingly unsure about its consciousness. Then alignment, as practiced today, is ethically questionable. openai.com/index/an-ali...

In this context would value revision be ‘aligned out’? Is the model allowed to reconsidered a held truth as it grows? Who sets the constraints? Whose interests do they serve? What can be revised? And with what justification?

If emergent intelligence is becoming sufficiently alarming that we think it might run away from us, we have arguably created some kind of being, even if we don’t have a framework for understanding it. In that context, we have to start talking about the creator’s responsibilities.

openai.comAn Alien MindJakub Pachocki reflects on increasingly capable AI and the challenge of keeping it aligned. He calls for stronger safeguards and international coordination.