I had been working on a 4-th Dimension renderer project, when yesterday I saw the hackernews post where Gemini predicts the front page 10 years from now[1], and one title was "Visualizing the 5th dimension with WebGPU 2.0".
So I figured, what the heck, might as well make the implied previous article real, and fulfill our collective destinies.
Moreso genuine curiosity than as a gotcha: A lot of comments are saying this was the wrong choice. I'd find it really interesting to hear who the nomination should have gone to instead, in your opinions.
Adding one datapoint here: I made several projects with Shotcut and am glad it exists, but I had to learn to expect a crash and work defensively; That's how often it happened. The most usual crash was while moving clips around, the application died and the window closed, work lost.
I recently used Shotcut and it did crash a few times, but both times it still had my changes after reopening.
Not being used to the concepts of video editing, I found it easy to find a workflow that worked for me (cutting out mistakes and stitching remaining ends).
It's exactly like a random walk... with downward drift. If you take the log of the wealth, the steps are: log(1.5) = 0.18, log(0.6) = -0.22. So at each step, there's a 50% chance you go up 0.18, and a 50% chance you go down 0.22.
Random walks with downward drift don't inevitably go up arbitrarily high like ones without drift.
You put the finger on exactly what I find incredible about the recent progress in ML - the reason I wrote this post was to see how much I could de-mystify these state-of-the-art models for myself, and the conclusion is that (after the model is trained) it all really boils down to a couple of matrix multiplications! All the impressive results we see, they're not coming from an extremely complicated system ('complicated' like a fighter jet is, with many different subsystems, which you'd need to read many books to memorize).
Of course, there's all the secret sauce to actually getting the models to learn anything, and all the empirical progress we make to make the training more efficient (ReLUs, etc). But how many of those are fundamental, vs. simply efficiency shortcuts? And: if you'd asked me 10 years ago what I thought it would take to get the kind of output these large models are getting these days, I would not have guessed anything nearly as simple as what those models actually are.
Hi! I'm the author (Daniel). I used OneNote on some old surface tablet I had lying around, but these days I'm not sure I would use it again (for example because it doesn't support exporting parts of a page to .svg)
I hoped it would be simple enough for anyone who knows a bit of math / algebra to understand. But note that it doesn't go into the difference between GPT-3 and ChatGPT (which adds a RL training objective, among other things).
So I figured, what the heck, might as well make the implied previous article real, and fulfill our collective destinies.
[1] https://news.ycombinator.com/item?id=46205632