RT by @saradurovic: 🧠 We'll have James Betker (@neonbjb) from @OpenAI talking about his DALL-E 3 work in my Discord server tomorrow! Do-NOT-miss-this-one :) Discord link: https://discord.gg/peBrCpheKE James used to work at Google as a software engineer in the security space before he started working at OpenAI and is now the first author on the DALL-E paper/tech report! His journey into ML is fascinating, he's an autodidact, same as me, so that attracted me to his story instantaneously. He built 2 servers in his garage with 8 RTX 3090s and a DIY cooling/power systems (those beasts are hungry, even my smaller rig with 2 RTX 3090 is drawing significant current). He then used one server to pay for the bills of the other one (by renting on http://vast.ai) :D and used that second one to train this amazing project called tortoise TTS which is the thing that eventually what caught an attention from people from OpenAI & DeepMind and ... the rest is history :) Check out tortoise TTS here: https://github.com/neonbjb/tortoise-tts And join us! If you have a recommendation on whom to invite next let me know! Comment down below. :) I'm always on the lookout for extraordinary people.

R to @gordic_aleksa: argh almost forgot: October 24th, 6 pm CET (9 am PST) :)) Again, server link: https://discord.gg/peBrCpheKE

R to @gordic_aleksa: GANs strike back. Oh wait… they never left

what power elite don't want you to know... before the generative AI boom the AI was super super discriminative think about that for a second do you really want to live in such a world? discriminative AI is still there they say

R to @vkhosla: VinodAI I'd buy that domain name

tired: generative AI wired: discriminative AI

R to @gordic_aleksa: Too bad one can't see the videos if one was not enrolled while the course was active Welp, gotta wait till October the 31st I guess

R to @PrakamyaMishra: Great!

R to @PrakamyaMishra: Tips and tricks for training transformer models (position of layernorm+use of weight decay) is covered in Tutorial #3 from @BorealisAI. Quantization and pruning is also covered in the lectures and papers by @songhan_mit. @ykilcher has also covered ALiBi

R to @PrakamyaMishra: More about Positional Encoding and Transformer variants for longer sequences (Attention Mechanism variants + decomposition variants): Tutorial #16: Transformers II: Extensions by @BorealisAI https://www.borealisai.com/research-blogs/tutorial-16-transformers-ii-extensions/

R to @PrakamyaMishra: More about Scaling LM models (Magetron LM): Video by @gordic_aleksa https://youtu.be/hc0u4avAkuM?si=E1FxlMiVHtxL8VCS

RT by @jag_singh95: 🤖🔥 AI news highlights from last week! * Flash-Decoding! A continuation of the flash attention work that makes LLM inference much more efficient. Exploits the fact that during inference we have a single query token which causes suboptimal GPU packing. Speeds up the attn layer further ~60x (server-grade GPUs presumably similar benefits for high-end commercial)! https://pytorch.org/blog/flash-decoding/ * New open-source multimodal (vision-language) model in town: Fuyu-8B by team at @AdeptAILabs. Simpler architecture, lacks an encoder (e.g. the type that DeepMind's Flamingo used), uses plain-old transformer decoder. https://www.adept.ai/blog/fuyu-8b demo: https://huggingface.co/spaces/adept/fuyu-8b-demo Also check out CogVLM from Chinese Tsinghua university (also open-source and even more powerful but larger): https://github.com/THUDM/CogVLM/blob/main/assets/cogvlm-paper.pdf * New non-open-source 😅 vision-language model from Google: PaLI-3. SOTA across various VLM benchmarks while only 5B params (no comparison with above 2 as they're concurrent work): https://arxiv.org/abs/2310.09199 * @laion_ai just released CLARA - a contrastive audio-text model that allows for better emotion parsing. The accent is on emotions - super important given the growing importance of AI assistants: https://laion.ai/blog/clara-release/ Also check out the new strategic games dataset they've released to improve the reasoning capabilities in LLMs: https://laion.ai/blog/strategic-game-dataset/ * DALL-E 3 is now available to ChatGPT Plus/Enterprise users. I used it to generate the image below :) Also - we'll have James Betker from OpenAI, one of the authors of DALL-E 3, in my Discord server on Tuesday! Join us here: https://discord.gg/peBrCpheKE https://openai.com/blog/dall-e-3-is-now-available-in-chatgpt-plus-and-enterprise * Amazon strengthens its collaboration with Agility Robotics and starts deploying bipedal humanoid robots in their facilities. Super impressive capabilities, give it 10 more years and I can't start to imagine where we're heading towards. Abundance. https://www.youtube.com/watch?v=NggB00SRhBg * Eureka, an open-ended agent that designs reward functions for robot dexterity at super-human level. LLMs so far were mostly good at generating coarse grained plains & actions but they fall short for fine-grained motor controls. This work basically enables reward through code generation. Super cool ideas, check out the paper: https://arxiv.org/abs/2310.12931 So much going on, the world is accelerating and the everyone is obsessed with AI. 2020 felt like an AI winter in comparison even though that's when this whole new cycle started with the release of GPT-3 - and things were moving fast even back then, hah. Keep learning :) keep building.