Scene_Cast2

I wonder if there's research into fitting gaussian splats that are dependent on focus distance? Basically as a way of modeling bokeh - you'd feed the raw, unstacked shots and get a sharp-everywhere model back.

@dang sorry for the meta-comment, but why is yfontana's comment dead? I found it pretty insightful.

The juiciest one is the Albin Counter-gambit. If you follow the "ideal line" where white blunders and takes the bishop bait, there's a neat knight underpromotion to win a queen.

From my own play, I typically see knight f3 from white on move 4, which still results in interesting games.

Are they competitive per Watt and per mm2 of die?

Another interesting take is how GPUs handle branching (latest NVidia GPUs in particular - see the register interconnect and how they swap threads without saving registers).

I'll copy-paste a comment I saw on Nikon Rumors (not sure whether it's copypasta or not)

---

I drove my electric BMW the other day, blasting a simulated V8 noise from speakers. It was a cold grey murky day but no rain. I stopped by the gas station to fuel my stomach by a bag of chips and the Snickers bar, because I went without eating a breakfast that morning. I saw a lonely dog by the roadside. It looked sad. I took my digital retro-styled camera with film simulation function out of my retro Billingham bag and took a photo. A little speaker in the camera has simulated the film advance noise just like in the past. Doggo looked at me with its sad eyes and went away. I took a glimpse of a photo of a dog and pressed "film grain +2" in the menu. Lovely shot. I'll post it to the Insta, probably. Then I entered the store, bought my bag of chips and the Snickers bar and saw a vinyl record corner. Man, I love vinyl. Those digital files pressed onto tangible, tactile surface. An AI-generated woman looked at me from the record artwork. Fonts were crooked. The price was $8.99 with a discount. I knew it's a pop record right away. Though, I'd love to blast an IDM track from speakers in my electric BMW alongside with simulated V8 noise, a pop record with vocoder vocals and autotune is also good. I took a record to place the vinyl on the bookshelf in my room. I know I'll be listening to the music via Spotify anyway. Man, I love vinyl. Just like film photography, it reminds me I'm alive. I'm real.

I have a OnePlus 8 Pro with an IR camera. It's pretty nifty - nature photography looks cool, seeing through stovetops is neat (and seeing when they heat up), and VR things are also often playing around with IR (plastic transparent to IR, IR LEDs, etc).

I ended up having to flash Lineage, as there was some outrage that in a highly limited set of circumstances, thin see-through T-shirts became slightly more see-through and OnePlus disabled that camera in their later firmware updates.

I use History Trends Unlimited. Works decently well for my use cases.

I find that Chrome has a fairly crippled history by default (worse than any other browser I've ever used). It's so bad that I ended up installing a history extension. Works much better.

Oh that 9mm is my absolute favorite lens (a close second being a 105mm). It's surprisingly versatile, too - I've done street photography, architecture (probably where it excels the most), landscape, and even macro.

I love the look of flame fractals, but I never figured out how to tweak things to get them to look more like how I want. I could never predict what a particular change would do to the look, or which change is needed to get a particular look.

I tried Apophysis (or one of the many spinoffs), but it never quite clicked.

Nope. Every single system like that will misbehave if given a bad set of weights, or even a random set of weights. I'd go as far as saying that even with "good" weights, it's likely to have some sigma of misbehavior.

Just like with their last release, they only released the architecture and not the weights. It may be useful for analyzing the system if you're a competitor (but from my last dive into it, it seemed like a strict subset of fancier, industry-leading rec systems), or perhaps getting into rec / retrieval systems as a newcomer.

However, this gives roughly zero insight into how Twitter's feed behaves.

Oh interesting. I was considering getting a Chinese phone soon (a vivo x200 Ultra), this really changes things. Is there somewhere I can read more about this?

I like Lynx launcher, and would switch to Kvaesito if Lynx wasn't available.

Why can't you put ads on RSS? Either in the story itself (by the site or aggregator) or as a "promoted item" in the feed (by the aggregator). If anything, the Google Discover (or whatever it's called) is not too different, just that you don't control which exact news sources you're subscribed to.

I can imagine an alternate timeline where Google Reader turned into a sort of Twitter (or FB or IG) feed.

I'd love for type checking in ML code. For example, here's an einsum from my code: "bhsc,cst->bhst". I'd love for a static analyzer to be able to tell if I accidentally swapped two of my dimensions and the einsum won't run, or what the output dimensions of an operation are.

If you think that this is sad, take a look at Canada and Cohere. I think it's not that Europe is lacking something, but rather the US and China are the only ones that are able to pull something off.

There's also the Box Tab X C, the color variant. Has color (and thus a backlight and the DPI tradeoff), and the pen is not Wacom.

There's also the Fujitsu Quaderno gen 3 as well - half the weight but can't do epub or mobi.

The paper is hard to read. There is no concrete worked-through example, the prose is over the top, and the equations don't really help. I can't make head or tail of this paper.

A bit unrelated, but I have the same soldering helper as in the post. It's called Omnifixo (I found out about it from a YouTube video (Norm from Tested)). I was initially skeptical, but it's made soldering a whole lot easier - highly recommended.

I've done some mouse PCB mods myself (swapping dead switches mainly). My biggest annoyance is resoldering through-hole components - unfortunately aftermarket PCBs for mice are quite rare, and my favorite mouse isn't all that popular in the "mouse community".

My issue with upgrading to the 5090 for workstation ML use is that it both has higher TDP than the 4090 and it can only be limited to 70% power (not 50% like the 4090).

How would this matrix get trained with PyTorch? I currently have a toy Transformer network - I ended up marking the matrix as sparse and using SparseAdam - gives a bit of a performance boost, but at the same time I can't use torch.compile() on the fetch from this matrix.

It's the latter. The scale is meant for real-time monitoring of rapidly varying force (the primary application is about monitoring the force derivative and repeatable max force logging). It uses an aluminum load cell if you're familiar with those, there's a tad of a multi-kHz resonance that is typically overshadowed by the object properties.

I've built a scale with a kHz sampling rate and gram precision at +/-100kg range.

One thing I found out is that getting calibrated accuracy beyond 0.1% is hard and expensive despite having all that precision.

That's pretty sad. I find their color film resolves better than Ilford's B&W stuff (I know it's apples to oranges, but if I'm shooting B&W, I kind of want the detail...). Their film is also generally easier to find than Provia or Velvia, for medium format at least.

I find it interesting that the architectures of modern open weight LLMs are so similar, and that most innovation seems to be happening on the training (data, RL) front.

This is contrary to what I've seen in a large ML shop, where architectural tuning was king.

I found a fairly large improvement in my toy transformer model where I added a "global" token akin to the CLS token in ViT.

Another approach I've seen is the "Diff transformer" from MS Research (https://github.com/microsoft/unilm/tree/master/Diff-Transfor...).

Not the OP, but I have a similar dilemma. I'm currently sitting on a SOTA ML model for a particular niche. I'm trying to figure whether I should try selling it to the incumbents (in some shape or form), or if I should publish a paper on the techniques, and/or if I should OSS it.

I wonder how much performance is left on the table due to it not being zero-copy.