Isn’t this the basis of GAN (Generative Adversarial networks), which is how most GenAI image models work? The purpose of generator network is to generate data that is as close to the training set as possible. The purpose of discriminator network is to distinguish the original from generated data.
Is blue-team and red-team like a post-training generator and discriminator?
Apple M1 mac GPU is orders of magnitude slower than a single Nvidia T4 GPU with CUDA for certain machine learning tasks.
This was my experience while trying to run Stable diffusion Inpainting/Outpainting to expand images. This task took a mere 5-10 seconds on Nvidia T4, while it took almost 5-6 mins on M1 mac. Nvidia T4 was almost 30x or more faster. I’m not sure how much of this performance is attributable only to the GPU and how much to CUDA working only on Nvidia T4.
“The long-term solution to ATT, though, is to build probabilistic models that not only figure out who should be targeted, but also understanding which ads converted and which didn’t.”
What does this even mean? Is it really possible to build models where the conversion label itself is not deterministic due to ATT and has to be predicted by another probabilistic model?
My understanding of how FB advertising worked as the following:
1. You purchase ads on FB
2. You tell FB how much a conversion is worth to you (let's call that $W)
3. You link your site to FB (via a pixel etc) so that they can automatically measure conversions. They say this is for you so that you can see which ads are actually working.
4. In reality, FB now knows what a conversion is worth to you so they can keep upping the price per ad till it gets as close to $X as you can take before you stop advertising.
If ATT is in play now, FB doesn't know the conversion rate and then can't tell how much to "squeeze" you by raising the price.
That would therefore incentivize them to figure out other ways to figure out that rate otherwise how can they maximize advertising revenue if they can't squeeze you?
Honestly, I'm not sure how to say this politely, but your argument does not map to my understanding of how FB ads worked (I was there from 2013-18, predominantly working on ads).
Steps 1-3 are correct, step 4 is not.
What actually happened is that you'd say I want to spend $10 on FB ads. There's a system called pacing that reduces your bid most of the time so that you don't run out of budget before the end of the day.
Separate to this, FB estimated how likely a user was to click/convert on your ad, and multiplied your bid by this probability (times 1000). This number was used to rank your ad against other ads/content.
So, what would actually happen is that you'd see amazing performance on that $10 budget, and you'd be like FB rocks spend all the money. Unfortunately, because of the way the above worked, when you 10xed the budget it would all far apart.
If FB had actually wanted to maximise short term revenue they might have behaved like you said, but the reason they've made so much money over the years is that they were always much more focused on the longer-term revenue.
If you think about it as a diminishing returns curve, it becomes something that can be understood and explored. Of course, the sales people will try to get you to expand quickly and your most efficient spend will come first - after your inflection point, your incremental return will decrease.
However, fighting that pull is the optimization opportunities enabled only at scale. If I run 1 ad, I know how that performs. If I run 1000 ads, I know which ad performs best.
Think Python or what used to be called “How to think like a Computer scientist” is still one of the best beginner programming books IMHO. There is an edition which is upgraded to Python 3 which you can find here - https://greenteapress.com/wp/think-python-2e/
There are 3 kinds of apps you need to reduce phone addiction. Listing out examples in Android
1. Apps that Analyze usage - Your hour, Stay free etc. These apps also allow you to specify max number of duration you can use specific app/phone, but this seldom works. The analysis helps you in understanding your usage patterns
2. Apps that block specific apps at specific times in the day - like AppBlock. Based on your usage patterns and work hours, block the most used scrolling apps during your free time. Blocking 100% of the time does not work, allow 30-60min windows where scrolling apps are allowed.
3. Notification managers - like FilterBox - Use Android settings to block notifications for most scrolling apps. Even for essential apps like Gmail/Whatsapp, FilterBox gives you lot of control over when you want to allow/block notifications
Is Advent of code used mainly for learning new languages? I find it very different from Topcoder style questions. I tried 3 puzzles and part1 and part2 in each question makes this a good test of how to design code in order to make it extensible for future changes. Can someone who has done Advent of code in the past confirm if learning new languages and making your code extensible are the primary objective of AoC questions?
reply