#196 - Nvidia Digits, Cosmos, PRIME, ICLR, InfAlign

Our 196th episode with a summary and discussion of last week's* big AI news! *and sometimes last last week's  Recorded on 01/10/2024 Join our brand new Discord here! https://discord.gg/nTyezGSKwP Hosted by Andrey Kurenkov and Jeremie Harris. Feel free to email us your questions and feedback at contact@lastweekinai.com and/or hello@gladstone.ai Read out our text newsletter and comment on the podcast at https://lastweekin.ai/. Sponsors: The Generator - An interdisciplinary AI lab empowering innovators from all fields to bring visionary ideas to life by harnessing the capabilities of artificial intelligence. In this episode: - Nvidia announced a $3,000 personal AI supercomputer called Digits, featuring the GB10 Grace Blackwell Superchip, aiming to lower the barrier for developers working on large models.  - The U.S. Department of Justice finalizes a rule restricting the transmission of specific data types to countries of concern, including China and Russia, under executive order 14117. - Meta allegedly trained Llama on pirated content from LibGen, with internal concerns about the legality confirmed through court filings.  - Microsoft paused construction on a section of a large data center project in Wisconsin to reassess based on new technological changes. If you would like to become a sponsor for the newsletter, podcast, or both, please fill out this form. Timestamps + Links: (00:00:00) Intro / Banter (00:04:52) Sponsor Break Tools & Apps (00:05:55) Nvidia announces $3,000 personal AI supercomputer called Digits (00:10:23) Meta removes AI character accounts after users criticize them as ‘creepy and unnecessary’ Applications & Business (00:16:16) NVIDIA Is Reportedly Focused Towards “Custom Chip” Manufacturing, Recruiting Top Taiwanese Talent (00:21:54) AI start-up Anthropic closes in on $60bn valuation (00:25:38) Why OpenAI is Taking So Long to Launch Agents (00:30:08) TSMC Set to Expand CoWoS Capacity to Record 75,000 Wafers in 2025, Doubling 2024 Output (00:33:10) Microsoft 'pauses construction' on part of data center site in Mount Pleasant, Wisconsin (00:37:23) Google folds more AI teams into DeepMind to ‘accelerate the research to developer pipeline’ Projects & Open Source (00:41:59) Cosmos World Foundation Model Platform for Physical AI (00:48:21) Microsoft releases Phi-4 language model on Hugging Face Research & Advancements (00:50:16) PRIME: Online Reinforcement Learning with Process Rewards (00:58:29) ICLR: In-Context Learning of Representations (01:07:38) Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs (01:11:44) METAGENE-1: Metagenomic Foundation Model for Pandemic Monitoring (01:15:45) TransPixar: Advancing Text-to-Video Generation with Transparency (01:18:03) The amount of compute used to train frontier models has been growing at a breakneck pace of over 4x per year since 2018, resulting in an overall scale-up of more than 10,000x! But what factors are enabling this rapid growth? Policy & Safety (01:23:45) InfAlign: Inference-aware language model alignment (01:28:44) Mark Zuckerberg gave Meta’s Llama team the OK to train on copyrighted works, filing claims (01:33:19) Anthropic gives court authority to intervene if chatbot spits out song lyrics (01:35:57) US government says companies are no longer allowed to send bulk data to these nations (01:39:10) Trump announces $20B plan to build new data centers in the US   See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.