stared

On another note - who needs WordPress in the age of Astro and LLMs?

I mean, it's not a taunt, but a serious question - do people keep WordPress because it used to be the easiest solution to set up years ago, or are there still clear use cases where one can argue it's the best solution?

See the "caveats" section.

It was also our initial assumption, but to our surprise there were no signs of models "knowing" the solutions (we investigated all trajectories). Compare and contrast with SWE-Bench Verified, for which models suddenly generate solutions, or at least, have substantial hints, https://openai.com/index/why-we-no-longer-evaluate-swe-bench....

It is on the way! Sadly, this Kimi K3 API rate limits are devastating, at least on the OpenRouter. So far it solved all The Intro levels (as expected) - I am much more curious how it works with The Lake.

I am almost certain it is star or spark.

But again, if someone is looking for assholes, they will find them.

In a similar vein, some time ago I got curious why the Grafana logo looks like the Zerg emblem see https://www.reddit.com/r/grafana/comments/1o79zxy/grafana_lo....

It was our original assumption. Yet, we went through trajectories and agents did not recall solution. It is with a sharp contrast with task for which agents magically generate solution, e.g. https://openai.com/index/why-we-no-longer-evaluate-swe-bench....

In a few instances (we covered it in Caveats) Gemini 3.5 Flash "knew" which level it was, but misremembered, and went with a wrong solution.

In the spirit of ARC-AGI-3-like challenges, we just tested if frontier AI models are able to solve a lovely puzzle game, Baba Is You: https://quesma.com/blog/baba-is-bench/

A year ago, Sonnet 4 barely solved the first level. Now, both Fable 5 and GPT-5.6 Sol beat the first two stages. GPT 5.2 is slow, but efficient, while Gemini 3.1 Pro and 3.5 Flash struggle.

Open-source repo with the benchmark (requires a copy of the lovely game): https://github.com/stared/baba-is-harbor

I am curious how it works for people with ADHD.

I mean, I would love to read more books (and I have a very long backlog), but very rarely I find them stimulating enough to sustain attention. And if they do... well, then time flies and other commitments are missed.

Sure, everyone's ADHD is different and I know quite a lot of people, for whom book are go-to. For me there is a narrow bar between not enough stimulated to start, and too stimulated so in my own words of a thousand thoughts.

Since when Zuckerberg is an AI expert?

I am puzzled (and irritated) why there is „rotate left” without „rotate right”. Does any of you know why?

There was "It’s Not Enough to Be Right – You Also Have to Be Kind" https://news.ycombinator.com/item?id=21490714

But I think the core part is WHY we want to be right? To prove something to others, or to ourselves? To feel better? As a compulsion? As a gambler's fallacy? Many motivations are less lofty that we dare to admit.

I wasted way to much time arguing online. It was mostly wasted time, and wasted emotions. I mean, I also had many eye-opening and enlightening discussions, but these rarely were fights.

Nice!

If you want to play a hyperbolic minesweeper, Hyperrogue features that https://hyperrogue.fandom.com/wiki/Minefield

Yes, it gets really hot really fast.

As much as I was tempted to use it on longer projects, I had some reservations about whether it would put too much strain on my MacBook.

All experiments with Qwen 3.6 required no more than 48GB Apple Silicon. I believe you can go even further with more aggressive quantizations - one can go down even further.

In any cases, from the economic point of view, running models on laptops make little sense. Even at the pure cost of energy consumption, it might be hard to beat pricing at tokens generated at scale.

At the same time, it is a breaktrough, that will change the game. Previously such vibe coding on consumer device was not hard or costly - it was impossible.

I didn't know it has a name. But I have been using similar thing, https://github.com/harbor-framework/terminal-bench-science/p...

See also Gribouille: A Grammar of Graphics for Typst, discussed here a week ago https://news.ycombinator.com/item?id=48541062.

Violin plots have an interesting reputation (https://xkcd.com/1967/, https://www.reddit.com/r/labrats/comments/91ex4u/is_it_just_..., https://jabde.com/2022/12/22/banned-violin-plots/).

For showing distributions, I much prefer strip plots (https://seaborn.pydata.org/generated/seaborn.stripplot.html), perhaps with opacity, or swarm plots (https://seaborn.pydata.org/generated/seaborn.swarmplot.html) - no averaging with an unknown kernel, no hiding distributions behind a box plot, and the data is directly visible. We also directly see whether it is based on 5, 100, or many more points.

When using histograms, binning is usually more straightforward than kernels. And in any case, the mirror reflection of a histogram is not needed.

A year ago, added R to the pipeline (with multiple complications) just to use ggplot2 - even though Python was the main tech.

https://quesma.com/blog/sandboxing-ai-generated-code-why-we-...

Good, that ggplot2 can run inside in WASM, vide https://github.com/QuesmaOrg/webr-ggplot-playground

For malware detection, many models are biased for or against detecting a threat (likely a thing that can be adjusted with a prompt).

I suggest tasks cannot be guessed (find, not tell). And 2d charts, both for ROC and pricing, vide https://quesma.com/benchmarks/binaryaudit/

A plain dictionary makes this harder or virtually impossible (you most often will silently fail)

I didn't know about these Typst restrictions. Silent fails are the worst, so if a constructor is necessary to prevent these, good it is there.

Thanks for explaining! (And for developing gribouille in the first place!)

My take: does something add value OR is there because were are just used to?

Gribouille is not ggplot2, or other. Syntax is different. Superficial keyword similarity is (usually) a false friend. Reusing a keyword might be useful, but keeping an unnecessary construction is (in my view), a cargo cult.

Typst itself breaks with a lot of LaTeX stuff, and it is good that it does not pretend it is LaTeX-with-Rust, but has a fresh look.

Tauri is getting traction in the meantime.

A non-native UI has some issues, but also one clear advantage - it is easier to make a cross-system app with the same looks.

Interesting! If I get it right, the API is in the spirit of Observable Plot (https://observablehq.com/plot/), less ggplot2.

In any case, I'm curious whether aes is necessary, or whether it would suffice to drop this function entirely and just use keys in the mapping (similarly for labs). Or, more broadly, whether using patterns from other implementations of the Grammar of Graphics is a conscious decision, or some sort of legacy baggage.

It revolves around the sentiment of "go deeper" - but I think it is a double-edged sword. Sure, entropy, tensors and gradients are important - and yes, they are pretty much requirements.

But from what I see, it is the opposite - a lot (if not virtually all) progress in the last decade of deep learning was not because of a fundamental idea, but incremental, experimentally-verified practice. Even though I think there is good intuition for why ReLU is better than sigmoid (tl;dr: last layer is log(sigmoid) ~ ReLU, putting anything different inside kills the gradient), the original paper by Hinton himself was more or less "because it trains 3x faster".

Re-thinking fundamentals might help, but most "let's change the fundamentals" is rarely how it works. Even the most seminal papers, i.e. AlexNet and "Attention Is All You Need", are refinements of existing ideas, and show how they help.

Machine learning is an experimental science. Many mathematically cool ideas do not work. Many engineering ones do.

I've tweeted before that one of the most important traits in a researcher is healthy paranoia. Be paranoid!

I have seen so many PhDs burned out to cinders; I don't think it is any more a good piece of advice than "depression is good for philosophers". Sure, be a relentless explorer.

In short, holding on to ideas for too long can actually be counterproductive. Stay open-minded and refuse to let ego cloud your judgement.

Which I think is true.

https://xcancel.com/arena/status/2066957802741043641#m

Military race isn't a good think either, but you don't want to be on the losing side.

If there were European MicroSoft or Google, there would be a preference.

Let’s autonomous Russian drones, and Europe is at mercy of two other empires, who capitalize on this opportunity.

Sure.

At the same time, it made in many cases EU dependent on the US. A lot of governments are basically dependent on MS Office or Google Cloud.

With AI, it is even more strategic.