News Earth Simulator, NIMs, RTX AI, Rubin - NVIDIA Computex June 2024

[Deleted member 2206]

NVIDIA is one of the biggest players when it comes to the AI revolution. The world was SHOCKED and STUNNED as they always are with these announcements during their March keynote earlier this year in which they announced...

Spoiler: March 2024 Announcements

• New incredibly powerful 'Blackwell' chip. Around 2-30 times as powerful as the previously used chips and consumes 25x less electricity and cost. These can be used to run trillion-parameter models and corporations that develop AI are already adopting these.
• Project GR00T; a foundation model for humanoid robots that allows them to mimic human movement through literally just watching them.
• Development of a new 6G cloud platform AI.
• Among other things...

3 months later in the heat of June we get another slew of very interesting announcements such as...

• Earth 2 - An Earth Simulator using AI to predict climate patterns with very high resolution.
• Nvidia Inference Microservices (NIMs) - Pretrained, specialized AI models with software interfaces able to run on CUDA.
• ACE Digital Humans - Digital humans running on the NIMs framework integrating several AI systems similar to GPT4o with the ability to input sight and sound and output AI-generated responses and speech; these digital humans will have AI avatars and agentic capabilities, replacing human workers (potentially).
• Further Chip Developments - Alongside more information about new 'Blackwell' chips releasing this year they promise 'Blackwell Ultra' chips for 2025 and 'Rubin' chips for 2026 with ambitious computational capabilities.

All in all it seems like even if we don't get AGI by 2030 I still believe that there will be cataclysmic effects being felt by everyone by 2030 requiring UBI. We are not even halfway through the decade and we already have AI with agentic capabilities, super-duper serious automation opportunities for business owners to salivate over, the emergence of autonomous robots that can learn by observing humans, rapid chip developments and earth simulators coming into play.

 

[クーロ]

Project GR00T; a foundation model for humanoid robots that allows them to mimic human movemen

They are being very vague on it
View attachment 73ba42c3-53c0-4b51-8cde-ee7d63a54e4a_1600x900.webp
But I guess it uses video data for imitation learning then adds onto it with reinforcement learning using NVIDIA's GPUs? They're using a lot of acronyms I don't get lmao. Looks interesting though considering these models usually take a ton of power to train

 

[クーロ]

Also some Chinese dudes in Hefei are making progress on AI wives

Chinese Researchers Create Jia Jia – a Super-Lifelike 'Robot Goddess'

中国科学技术大学

en.cs.ustc.edu.cn

Jia Jia is their latest robot and they're working on making it have realistic emotions. We'll have to see what companies say at the fair on 4/21 before they move on with it though

 

[MelaninQueen]

Nvidia engineers have big brains, but... 28GB VRAM only on 5090s!? Dude, I want to run a finetuned Llama 3 70b model without having to spend two months of Western salary on SLi setup.

 

[MelaninQueen]

Also some Chinese dudes in Hefei are making progress on AI wives

Chinese Researchers Create Jia Jia – a Super-Lifelike 'Robot Goddess'

中国科学技术大学

en.cs.ustc.edu.cn

Jia Jia is their latest robot and they're working on making it have realistic emotions. We'll have to see what companies say at the fair on 4/21 before they move on with it though

MY WIFE!

 

[Deleted member 2206]

Updates:

Spoiler: Omniverse/Gr00T symbiosis:

Spoiler: Nims/ACE Digital Humans:

Spoiler: Earth 2:

Spoiler: Chip Developments:

Ya we keep getting new updates every 2 weeks or so. 2 weeks ago it was Google and their big announcement and around that time GPT4o also. Now NVIDIA. I don't see how anyone can claim we are stagnating or in an AI winter. Feels like we're heading toward Cyberpunk 2030s.

 

[Deleted member 2206]

Nvidia engineers have big brains, but... 28GB VRAM only on 5090s!? Dude, I want to run a finetuned Llama 3 70b model without having to spend two months of Western salary on SLi setup.

I am too poor
The best I can run locally right now is 10.7B models, if I want something with actual intelligence I have to hook up an Anthropic/OpenAI API. Although 10.7B models are good enough for roleplay atleast.

 

[MelaninQueen]

I am too poor
The best I can run locally right now is 10.7B models, if I want something with actual intelligence I have to hook up an Anthropic/OpenAI API. Although 10.7B models are good enough for roleplay atleast.

I wish I could run WizardLM 70B with large context on my GPU... but I only have 24GB.