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.
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.
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.
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.
Now: benchmarking AI at: https://quesma.com/blog/
Previously: co-founder & CTO at https://quantumflytrap.com/