Played with this a bit and from what I gathered it's purely a re-arch of pytorch models to work as .tflite models, at least that's what I was using it for. It worked well with a custom finbert model with negligible size reduction. It converted a quantized version but outputs were not close. From what I remember of the docs it was created for standard pytorch models, like "torchvision.models", so maybe with those you'd have better luck. Granted, this was all ~12 months ago, sounds like I might have dodged a pack of Raptors?
> Then we called in our Tech Lead / Manager, who had a reputation of being a human JavaScript compiler. We explained how we got here, that Math.abs() is returning negative values, and whether she could find anything that we were doing wrong. After persuading her that we weren’t somehow horribly mistaken, she sat down and looked at the code. Her CPU spun up to 100%, and she was muttering in Russian about parse trees or something while staring at the code and typing into the debug console. Finally she leaned back and declared that Math.abs() was definitely returning negative values for negative inputs.
It seems most of the things Netflix produces is optimized by the algo for attention. When I feel it directing me gives me the ick. Looking at you Squid Game.
It's part of the same phenomenon we see in social media. The first waves of social media and YouTube were predicated on the idea that you either seek out content yourself or view a feed of content you'd already taken action to subscribe to/follow. Services like Twitter, Facebook and YouTube pivoted to go from "pull" where users select content or stay within their own networks, to a "push" model where the algorithm predicts and autoplays content, mostly from strangers, based on highly accurate predictions of virality and eyeball-retaining potential.
Things like Netflix realized it too and buried the "Continue Watching" at a randomized index in an endless carousel, added Autoplay and even starts autoplaying something different after you finish a series. And of course, newer things like TikTok have always been this way. All these things are, I'd argue, user-hostile in that they're optimizing toward, in the extreme case, complete addiction.
No need to suspect, they are advertising that openly and are proud of it. Famously, years ago they invented House of Cards TV show by looking at the most popular search tags and picked the most popular ones to select a genre and theme of a new show. It was a story of many articles about Netlix back then.
No, that's not at all what happened. House of Cards was a highly regarded UK TV series from BBC (made in the early 1990s). Like many UK TV series, it was ripe for an American adaptation. Netflix won the bidding war for that adaptation.
Making up "famous" examples doesn't make your case stronger, but the opposite.
> Famously, years ago they invented House of Cards TV show by looking at the most popular search tags and picked the most popular ones to select a genre and theme of a new show
This does not appear to be true based on any articles I can find. I do believe they heavily follow the trends from their analytics in what the shows they buy and what they cancel, though.
Homogeneous systems kill innovation, with that in mind, I guess it’s a good thing DeepSeek disregards licenses? Seems like sledding down an icy slope, slippery. and they suck.
DeepSeek is a card trick. They came up with a clever way to do multi-headed attention, the rest is fluff. Janus-Pro-7B is a joke. It would have mattered a year ago but also just a poor imitation of what's already on the market. Especially when they've obfuscated that they're using a discrete encoder to downsample image generation.
This feels more like misery porn disguised as a crusade to 'protect' the vulnerable. I’m not suggesting we make it easier for people to get scammed, but intentionally engaging seems like a recipe for a harsher clap back. Education > Escalation.
This sounded fun and I wanted to try it, so I built a web version. It's open-source built with React Native. The code is here https://github.com/HatmanStack/savorswipe or it's hosted here https://www.savorswipe.fun You can easily add recipes to the swipe list by snapping a photo of ingredients and directions then adding the photo to the app.
Is it just me or is this stream janky? I had a 30+ min lag for it to start and checked it on my phone with the same wait. I guessed it was happening for everybody but didn't see anything else posted.
