1
Code for IF by @deepfloydai is up, amazing work by the team.
Weights in a couple of days along with blog posts, inference & other tools, key research Qs & more.
Release will be like SD, research => open release after feedback as new model type.
github.com/deep-floyd/IF
2
Lots of people don't get open source AI is inevitable & required.
Regulated & private data is never going to be sent to black boxes with unknown data.
Will be legally required to use open models to turn that regulated data into knowledge.
Folk also want to own their weights.
3
Stable Diffusion XL beta now available in the API as well as DreamStudio.ai
We are refining this rapidly with human feedback so expect significant improvements over the coming period as we head to open source release, do give feedback!
stability.ai/blog/stable-di…
4
GPT4 was a feat of engineering by a great team
Here is the (all highly skilled) team breakdown:
1. Pretraining: 83
2. Long context: 12
3. Vision: 57
4. RL & alignment: 128
5. Evaluation & analysis: 147
6. Deployment: 86
343 unique folk in total (some did more than one area)
5
The other aspect of this is we can also now adjust facial expressions using AI in media we create and pretty much live as well.
Think about that.. twitter.com/rowancheung/st…
6
The 32k token window version of GPT-4 will be a bit of a shock to many in programming.
It's like 20k words of instruction..
You can chain lots of them together too to check their outputs at.. not much cost.
Big danger to SaaS and more.
Will only improve from here. twitter.com/mckaywrigley/s…
7
Stable diffusion XL #SDXL is in testing with our partners & dreamstudio.ai, getting data to refine it before open source release.
It is a 2.3 billion parameter variant (original 900m) with a range of improvements that will feed into Stable Diffusion 3
🚀 ✨
8
Folk saying open source AI is unsafe are idiots.
Better/essential for:
1. Interpretability 👀
2. Auditability 📝
3. Private data 🕵️
4. Regulated data ⚖️
5. Academia, reproducibility 📚
6. Optimisation ⏩
7. Security, less bugs 🔐
8. Alignment 😊
9. Liberty, no panopticon🗽
9
Not your weights not your brain 🧠 twitter.com/DrJimFan/statu…
10
On AI ethics & governance
Most powerful AI created by an org that has rejected its founding principles of transparency, openness, and collaboration
They say it could:
i) be an existential threat to humanity
ii) destroy our current society
Controlled by one unelected person 👍🏾 twitter.com/futurism/statu…
11
Open offer to anyone @OpenAI who actually wants to work on Open AI:
We will match your salary, benefits etc but you can work on any open source AI projects you like, ours or others.
Collaborate, be open and prioritise good outcomes over self interest:
openai.com/blog/introduci… twitter.com/jjvincent/stat…
12
Software was eating the World.
Generative AI will eat Software.
13
One reason we need free range, organic AI for all.
It’s a question of Liberty.
Also what happens when you get deplatformed from centralised AI.. twitter.com/Meaningness/st…
15
Google and Microsoft are going all in with generative AI as core to their future.
There is no “we are still early” here, trillion dollar companies are shifting their whole strategy and focus.
I can’t ever recall a technology and strategy shift as fast and meaningful as this.
16
Putting aside $$s for a moment is there anyone who would rather have the next generation using closed, centralised, corporate-aligned AI versus open, customised and personally-aligned AI?
It’s a choice we can make right now at the start of this crazy takeoff 🛫 🚀 🌟
17
Stories about how chatGPT will kill Google are a bit silly.
Google have the best full stack LLM team and infra with custom chips (PaLM, LaMBDA, Chinchilla, MUM, TPUs etc)
Nobody can bet them on innovation, cost or go to market.
Institutional inertia is only limiting factor.
18
The AI art debate misses that future models will likely be zero/few shot learners.
Model base will have just a few dozen million images, licensed
Show it a portfolio of images, write prompt and instant generation in that style
See MJ v4 image promoting as an example/dreambooth
19
So, open source #ChatGPT? 🤔
I have some thoughts 💭
But first what does everyone else think 🧐 👇
20
Access to next generation AI is a human right.
Otherwise we will have augmented humans and everyone else, divide will grow.
Restricting generative AI access is like restricting internet, phone or camera access.
It must be open.
It must be a public good.
Support free speech. twitter.com/shl/status/159…
21
With distilled stable diffusion, chatGPT (which is awesome, great job by @OpenAI team 👏) and other stuff hitting soon we are almost at real time human-level generation of media and content.
How will this impact folk reading this tweet?
I’m surprised how few grok what’s coming
22
Distilled #StableDiffusion2
> 20x speed up, convergence in 1-4 steps
We already reduced time to gen 50 steps from 5.6s to 0.9s working with @nvidia
Paper drops shortly, will link, model soon
Will be presented @neurips by @chenlin_meng & @robrombach
Interesting eh 🙃 twitter.com/EMostaque/stat…
23
Some raw output from #StableDiffusion2
Like actually try it folks, it's only going to get better from here, each of these was generated in a couple of seconds.
We will release some optimised settings soon. 🚀✨
24
Some user notes on V2! 🧵
V2 prompts different and will take a while for folk to get used to.
V2 is trained on two models, a generator model and a image-to-text model (CLIP).
V1 used the CLIP-L14 model open sourced by @OpenAI which was great, but nobody knew what was in it 1/ twitter.com/StabilityAI/st…
25
Nothing major, just take brain recordings and reconstruct what folk are seeing at the time from them.
Exponentials eh. twitter.com/_akhaliq/statu…
Not your weights not your brain.
@stabilityai