some news: i quit MSFT last week after 3.5 years with the inflection/microsoft AI crew -- lots of memorable times. i'm taking some time to reconnect w/ old colleagues and friends and its been great. if you are reading this and want to catch up, DM me!!
Excited to announce that we’ve raised $1.3B to build one of the largest clusters in the world and turbocharge the creation of Pi, your personal AI.
forbes.com/sites/alexkonrad/…
It’s a big week! We’ve raised $1.3 billion and are building the world’s largest AI cluster (22k H100s).
We’re grateful for our investors and new funding that will help us accelerate our mission to make personal AI available to every person in the world. inflection.ai/inflection-ai-…
We have amazing results to announce! Inflection-1 is our new best-in-class LLM powering Pi, outperforming GPT-3.5, Llama and PALM-540B on major benchmarks commonly used for comparing LLMs. inflection.ai/inflection-1
My post on SotA image generative models was released 🥳
Featured 7 notable recent papers with emphasis on:
- VD-VAE
- VAE + discriminator (e.g. VQGAN, DC-VAE)
- Diffusion models (e.g. DDPMv2)
Plus some notes on scaling (e.g. DALL-E) and evaluation.
arankomatsuzaki.wordpress.co…
"The images are preprocessed to 256x256 resolution during training. [...] each image is compressed to a 32x32 grid of discrete latent codes using a discrete VAE that we pre-trained using a continuous relaxation."
GPT + VAE + scale = impressive results!
openai.com/blog/dall-e/
It is easy to write a program
but it is difficult to create a machine
that will read those lines.
(Was looking through my journal, found this gpt-3 generation conditioned on haikus)
Happy to announce our new work on score-based generative modeling: high quality samples, exact log-likelihoods, and controllable generation, all available through score matching and Stochastic Differential Equations (SDEs)!
Paper: arxiv.org/abs/2011.13456
It breaks my 💚 when researchers tell me that VAEs don't work. My first typical question is "did you try hierarchial VAE or vanilla VAE?", the answer is usually vanilla VAE.
VAEs work much better with hierarchical structures. NVAEs and this work take this to the extreme!
Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images
“Very Deep VAEs” achieve higher likelihoods, use fewer parameters, generate samples 1000x faster, and are more easily applied to hi-res images, compared to PixelCNN.
openreview.net/forum?id=RLRX…
Posted my first paper on arXiv💥🙌
GPT-f is a Transformer-based automated theorem prover. We show that Transformer + Search is suitable to formal reasoning and continuous self-improvement 🦾
arxiv.org/abs/2009.03393
Our base model is a Sparse Transformer. If we make it bigger and train for a while with this augmentation, it results in both very high likelihoods (2.55-2.65 bpd on CIFAR-10) and also samples equal/better than most GANs (as measured by FID). Code here: github.com/openai/distributi…
Thanks if you came to our ICML poster on Distribution Augmentation. The zoom discussion was way more fun/interesting than I expected! TLDR of our work: use powerful data aug in your generative model by conditioning it on the aug. Improves samples + likelihoods considerably.