We have released *dot* (aka Deepfake Offensive Toolkit), an extensible library for injecting real-time deepfakes in virtual cameras. At no training cost.
🚀 First time at the the top of Hacker News news.ycombinator.com/news
👍 Congrats to team!
Follow
Click to Follow GiorgioPatrini
Giorgio Patrini
@GiorgioPatrini
Giorgio Patrini’s Tweets
original tweet
Quote Tweet
Paypal Mafia was yesterday. Look at the impressive offsprings of the OpenAI Mafia.
Show this thread
Why so many people exited OpenAI for their own gig? How is that financially sensible given the company trajectory? A downside of its initial non-profit structure (no employee equity)?
e.g. Anthropic is a "public benefit" but for-profit company
Image credits
2
4
13
Show this thread
Credits: 3d_printer_stuff
1
1
Show this thread
We have now clients who are like:
can't you do it faster, with chatGPT?
2
1
We have now clients who are like:
can't you do it faster, with chatGPT?
2
1
Credits: 3d_printer_stuff
1
1
Show this thread
FYI, people are already building adversarial attacks to GPT-detectors in the real world
8
14
I'm happy to announce Hugging Face Competitions! With 🤗 Competitions, you can create public/private competitions with full control of datasets, metrics and evaluation! Check out our first competition: AI or Not huggingface.co/spaces/competi 1/N
19
121
471
Show this thread
FYI, people are already building adversarial attacks to GPT-detectors in the real world
8
14
chatGPT is providing false information about the GPT3 API pricing. I didn't want to open the docs but I will have to now 😅
1
2
Looks like we're really hitting that AI exponential curve when you learn *the same day* about (web UI for AI apps in python) & (vector search API) 🤯
2
11
ok this got my attention 🫣
LLM to *replace* backend entirely (no implementation!)
api logic and memory storage are inferred on the fly
github.com/TheAppleTucker
3
3
ok this got my attention 🫣
LLM to *replace* backend entirely (no implementation!)
api logic and memory storage are inferred on the fly
github.com/TheAppleTucker
3
3
chatGPT is providing false information about the GPT3 API pricing. I didn't want to open the docs but I will have to now 😅
1
2
1
Looks like we're really hitting that AI exponential curve when you learn *the same day* about (web UI for AI apps in python) & (vector search API) 🤯
2
11
A huge amount of practical experience has been distilled into this great guide to tuning deep learning models. twitter.com/GeorgeEDahl/st
This Tweet is unavailable.
8
106
604
Quote Tweet
Large Transformers are powerful but expensive to train & use. The extremely high inference cost is a big bottleneck for adopting them for solving real-world tasks at scale. Check out my new post on some ideas on inference optimization for Transformers: lilianweng.github.io/posts/2023-01-5
🔥 New (1h56m) video lecture: "Let's build GPT: from scratch, in code, spelled out."
youtube.com/watch?v=kCc8Fm
We build and train a Transformer following the "Attention Is All You Need" paper in the language modeling setting and end up with the core of nanoGPT.
542
3,629
21K
Show this thread
I can't understand how does not have a "copy to clipboard" icon on their code blocks?
1
4
It looks like it isn't trivial to teach chatGPT to interpret images encoded as ASCII art. Here is a GAN generated face that I fed to it
1
3
8
Show this thread
I initially thought to do the same for images in byte strings, but the cap on chars number in chatGPT makes it difficult on the current UI.
Show this thread
I was inspired by
Quote Tweet
Turns out they might already
Can LLMs be few-shot *image* classifiers?
While prepping notes for cs324.stanford.edu, I revisited the idea... by turning images into ASCII
Check out the notebook here:
colab.research.google.com/drive/1I4RCsJ6
Can LLMs be few-shot *image* classifiers?
While prepping notes for cs324.stanford.edu, I revisited the idea... by turning images into ASCII
Check out the notebook here:
colab.research.google.com/drive/1I4RCsJ6Show this thread
read image description
ALT
1
Show this thread
I don't have a good hunch on what goes wrong here. Others have found out that working with ASCII art is particularly challenging for chatGPT
twitter.com/mmmbchang/stat
Quote Tweet
Can ChatGPT generate ascii art? Let's find out!
Some generations seem reasonable, but some are quite puzzling (further down the thread):
Show this thread
1
Show this thread
Given the many suprising emergency properties of GPT-like models, I am puzzled that ASCII images seem to be that hard to understand.
A 5yo can understand ASCII art, while python is likely too hard for most people at that age.
1
Show this thread
But then, I realise that these may just be random guesses all along, without much in-context learning
1
1
Show this thread
At this point, we may assume that this is a play:
- gpt doesn't know how to solve the problem, and RLHF has taught it well to be conservative under uncertainty
- when pushed, it will try to provide an answer anyway, and sometimes the answer is OK
1
Show this thread
I ask more specific questions, perhaps easier than elaborating a full description. For humans, it seems easy to guess gender, or an approximate age, for example.
chatGPT is being conservative again, and avoid an answer, at first. you can see a pattern
1
Show this thread
When asked to describe the image, it's just pure guess
1
Show this thread
I ask more specific questions, perhaps easier than elaborating a full description. For humans, it seems easy to guess gender, or an approximate age, for example.
chatGPT is being conservative again, and avoid an answer, at first. you can see a pattern
Show this thread
But it is? Let's try with another image. This one is larger. I won't show the original RGB, so you won't have advantage over chatGPT if you wish to play along
1
Show this thread
If corrected, it would seems to produce something meaningful, finally.
1
Show this thread
Indeed, there is no "eyes on the top left corner"
1
Show this thread
It's fascinating to see that with a little hint chatGPT can make sense of the image then. But upon closer inspection, I start to wonder if it's just hallucinating something in order to please me
1
1
Show this thread
Initially I can sense it is reluctant to give an answer, but if I push harder we will get a bad answer.
1
Show this thread