| Github | https://github.com/Libric0 |
The Hogwarts legacy image is the worst of them all concerning AI glow. Just for fun, I asked #chatGPT to do the same thing #DLSS5 does here. It also changed the faces significantly, and even destroyed the composition a bit (probably because the image is generated from scratch), but the #AI glow is still far less pronounced.
Almost like companies and researchers already learned that 'photoreal' does not mean 'pronounced lights coming out of nowhere'
No reason to buy a 5090 here. #Nvidia
Previous #DLSS iterations were about upsampling or improving lighting, while keeping the artistic vision. I admired the tech
#DLSS5 looks like someone took the image, put it and the prompt 'photorealistic, trending on artstation' through an image generator (Like #StableDiffusion or #chatgpt) and called it a day.
Its not 'better lighting', it changes faces as if the actors were replaced. It adds makeup. It even creates the uncanny #AI glow. No photorealism here. It wont improve #gaming
Today I learned: Using variables in Python lambdas is weird
Apparently they store the variable names and not the values. Like a reference, even for primitives. But if you supply the value as a default argument, it works as intended.
Hey people with #ADHD
Do you also hate re-planning the same tasks again and again, and forgetting things in the process? Me too! That's why I programmed an #app for that 
Routines lets you create and edit routines and follow them through in a step-by step manner (to avoid being overwhelmed)
Google play will take some time, but you can sideload it (.apk) from https://github.com/libric0/routines/releases
Btw: It's completely open source!
If the current political decisions at Meta are something you do not wan't to support, do the small step and install Signal
It is securer than Whatsapp (No Backdoors), developed by a Nonprofit (No enshittification), and works just the same
Just install it, and keep it on your phone. Thats the first step to taking back our communication from big coorporations
Android: https://play.google.com/store/apps/details?id=org.thoughtcrime.securesms&hl=en
iPhone: https://apps.apple.com/us/app/signal-private-messenger/id874139669
@signalapp #queer #lgbt #lgbtqia #trans #transgender #politics #uspol
For anyone #queer looking to emigrate from the #US to #Europe. I found this tool to compare European Counties by their Queer-Friendliness. Hope it helps
https://rainbowmap.ilga-europe.org/
Of course, we are not perfect either, but our continent might be a safer place for you
Stay safe <3
Edit: Also applies for other countries obv
After trying out #o1preview I must say its only impressive for the first question 😐
Not only doesn't it have tool use anymore, its also very verbose and sometimes forgets to answer or think completely
This points to the possibility of them using one end-to-end LLM to do both
It thereby mixes up both tasks
- Long Chain of Thought -> Verbose
- Expecting thinking counts as answer
I'd expect separate expert models would do better 🤔
The second problem is tokenization. #GPT s don't read text, they get tokens (numbers) corresponding to word parts. For strawberry, this is 73700. It's like asking a child who can talk, but not read.
If the training data did not contain examples, it will most likely be wrong.
I found 2 ways to fix this:
1) Capitalize the Rs to split the word into more tokens
2) Give Multi-Modal GPTs an image of the question. Then it will "read" it and find all Rs 🧵2/2
There is this fact going around that #chatgpt et. al. cannot count the R's in strawberry. This is not completely true.
Both humans and GPTs "think" heuristically, that is: quickly, but error prone. We need to remind ourselves to think slowly. If we remind ChatGPT to think slowly, i.e. using programming to count Rs, it answers correctly. 🧵1/2
#AI #GPT #genAI #machinelearning #deeplearning #artificial_intelligence #llm
🥳 I've just updated my dynamic loss-plotting library Loss Watch📊
A small bug was fixed, but more importantly, train losses now don't have to be, updated.
If you want a quick and pretty way to follow your models training / validation progess in 📔Jupyter or your 📺Terminal, give it a try!
If you like this project, I'd be thankful for a boost 
#python #deeplearning #AI #GPT #ChatGPT #StableDiffusion #library #useful #development #science #neural_networks #ki #development
