Alex J. Champandard ❄️ Profile picture
Artificial Intelligence expert, Deep Learning #ML research, ex-R☆/Guerrilla #AI Developer, co-Founded #CreativeAI. #⚘
Jan 31 11 tweets 3 min read
That's a wrap folks, show's over! Diffusion models as *lossy* databases that can regenerate their training data:

"Diffusion models are less private than prior generative models [...] mitigating these vulnerabilities may require new advances in privacy-preserving training." Image
Jan 22 33 tweets 12 min read
I have successfully compiled and run GLM-130b on a local machine! It's now running in `int4` quantization mode and answering my queries.

I'll explain the installation below; if you have any questions, feel free to ask! 130B parameters on 4x 3090s is impressive. GPT-3 for reference is 175B parameters, but it's possible that it's over capacity for the data & compute it was trained on...

I feel like a #mlops hacker having got this to work! (Though it should be much easier than it was.)
Jan 22 6 tweets 2 min read
Anyone in Europe want to try this?

Write to GitHub, ask for contact details of their Data Protection Officer. If they refuse, explain it's mandated by GDPR to provide contact details. Ask DPO what's their policy on personally-identifying information under GDPR. Post response! Oh, it's hosted on Google. Do the same there!

If you go via Support, you'll probably have to ask three times for DPO contact because frontline Google Support is not GDPR aware, and will refuse a few times to see if you're serious.

Not sure if intentional or incompetence!
Jan 21 11 tweets 4 min read
I never realized just how fragile tokenization can be when you're crafting LLM prompts!

Say your model was trained to summarize with "\n\nTLDR:" and you decide to include an extra space after ": " so that the space is excluded from the generated output: it's different tokens. So the next sentence could be "This research ..." but the statistics would get messed up because of the extra space, as the tokenizer would have tokenized " This" to include the space before the capital.

I'm not sure this really is "engineering" it's more like prompt hacking...
Dec 5, 2022 6 tweets 1 min read
Alex's Rules of Automation

If you automate something, it disappears until something breaks! Alex's Rules of Automation

When automation eventually breaks, it's always something new (and annoying) and may even take longer to fix than doing it manually.
Dec 4, 2022 4 tweets 1 min read
An automated system took this photo. Cue the debate over whether it's copyrighted or not! ;-) "Someone carried the camera to the top of the hill and set up the trigger. Why don't they own the copyright?"
Nov 5, 2022 20 tweets 4 min read
It's been 36h since this thread, with many constructive discussions since!

One front I failed on goes something like this:
"I thought you were an AI coder. How come you want CoPilot to be withdrawn? Do you want to cancel large models?"

Let's unpack! 👇
First, there's no risk of CoPilot service being terminated and the technology abandoned. I don't want to see that and that's not the objective of their lawsuit either.

Second, I think medium- to large-models are absolutely worth pursuing technologically!
Nov 5, 2022 9 tweets 3 min read
The Chrome team is cutting support for the superior JPEG-XL codec in its browser — even before they enabled it!

The decision was made in secret under the direction of a single person who has conflicts of interests, and promotes the inferior AVIF alternative. AVIF is based on VP10 codec, like a successor to WEBP which is based on the VP8 codec. Google owns & controls VP10, so has interests in promoting it instead of superior alternatives.

This means there'll be ~50% more energy used, and thus carbon, for internet bandwidth. 🙄
Nov 4, 2022 4 tweets 2 min read
"The Right To Read Is The Right To Mine" was a campaign from ~2012-2015 to convince the public & legislators that machines should bypass copyright for data-mining.

IMHO we're at the next stage of this campaign, now for generative systems — should they act outside copyright? Articles like this one are at the tail end of the first pro-mining campaign and precursors to this new generative campaign?

It tries to establish that "reading by robots doesn’t count" and "infringement is for humans only".… (via @GradySimon)
Nov 4, 2022 4 tweets 1 min read
If you're working at a generative company, and worried about the lawsuit against GitHub for their generative model, please take some comfort in the fact that I think they made *many* missteps — with either a serious lack of due care, or the intent to break the law. For instance, Google announced they had a similar code model and they didn't release it. They used it internally & measured a 6% improvement on productivity while they understand the legal and ethical implications.

(Could also be that Google wanted to see others get sued first?)
Nov 3, 2022 17 tweets 5 min read
Reading through the GitHub CoPilot litigation submitted; although it was pulled off quickly — it's a solid piece of work!

My assessment is that the defendants, GitHub, Microsoft and OpenAI are in a very bad position... The documents show how Codex and CoPilot act like databases; they have three different examples of JS code that is recited verbatim — with mistakes — from licensed sources.

Including this debug code below isPrime(n):
Nov 3, 2022 4 tweets 2 min read
You know how hands & fingers are particularly difficult to generate?

Wouldn't it be funny if people having important conversations online (in the near future) used hand gestures in front of their faces, so both sides know it's not a #DeepFake.

Anchor: I'm sorry to ask Mr. President, but before this TV interview can proceed please make a creative gesture with your hands.

Pres: What?

Anchor: Well, in the last election multiple candidates were caught using DeepFakes to make them look & sound smarter than they are.
Nov 3, 2022 5 tweets 2 min read
In NVIDIA's new paper on #Diffusion Models, they show how more denoisers (for each stage) and more embeddings (text, image) helps with quality!

TL;DR: If you buy more GPUs, you get correct spelling too. #AI #ML With so many different labs rushing to research and deploy this kind of technology, this will quickly turn into a race for more efficiency as different providers compete on costs too.
Oct 17, 2022 4 tweets 2 min read
When large language models are explicitly trained to use Python and look-up Wikipedia, we'll be entering scary territory for #InfoSec#AI! OpenAI engineers probably did this a few months ago, now frantically trying to make sure their Python sandboxed environments are sufficiently safe...
Oct 17, 2022 5 tweets 2 min read
It's amazing how this great paper about prompt engineering from August ( is only really getting wide-spread attention now there are good open-source implementations:

Academic Impact: OSS or GTFO? Prompt-To-Prompt editing allows you to easily change your input text without needing to completely regenerating the image. This makes it much easier to control the diffusion!

Example from bloc97's GitHub, four seasons of the same scene:
Jul 7, 2021 4 tweets 2 min read
Is there a precedent for a DeepDream or Style Transfer image (famous or not) being taken down under copyright laws? I'm aware of the Blade Runner case, but it was a feed-forward network and I'd like to know if other "ML Art" algorithms have had similar fate.…
Jul 1, 2021 38 tweets 14 min read
Watch carefully as GitHub PR tries to (re)define copyright and set a precedent that the licensing terms of open source code don't apply in this case...

Left: Text from June 29th.
Right: Edited text on July 1st.

See FAQ section at ImageImage Just for the record: it's not considered fair use, @github.

It's highly controversial in the community.

Multiple datasets have been removed from the internet because of such legal issues, and courts have forced companies to delete models built on data acquired questionably.
Apr 14, 2021 6 tweets 2 min read
The EU's leaked draft on AI Regulation is basically trying to regulate the whole of software. Try making any software without statistics, search or optimization...

(Except the military and autonomous weapons, for some reason, that remains unregulated. 🤨) Full document here, via @j2bryson.…
Apr 12, 2021 4 tweets 2 min read
The most interesting aspect of this tech is that it separates voice "acting" from voice "style":

- Actors can perform the lines regardless how they sound.
- Designers can separately license the style they want.

Big-budget games will still use humans for both. (via @mtrc) I suspect the voice style will be easiest to generate and license, but the voice acting will remain the domain of professionals for a while longer.
Nov 24, 2020 19 tweets 6 min read
#DeepFake Tutorial.

In this thread I'm going to post some tips & tricks to identify Deep Fakes using these examples I found online.

📊Vote in this other thread first:
There are many ways to detect deep fakes. Here are three of them:
1) Impossible static poses.
2) Impossible movements.
3) Technology artefacts.

The first category seems to be easier to detect, but the second is more reliable. The third may go away soon!
Nov 24, 2020 19 tweets 5 min read
#DeepFake Alert!

I've been digging through various propaganda and conspiracy websites (so you don't have to) and finding a surprisingly large number of deep fake appearances of Mr. Bіden.

Here is a thread with videos+polls to test your skills at discerning what's real or not... VIDEO #1

Original clip at higher quality (mp4/vp9):