Navigating AI Model Training and Alignment

Our 194th episode with a summary and discussion of last week's* big AI news!
*and sometimes last last week's 

Recorded on 12/19/2024
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:

- Google dominates AI news with multiple announcements, including a reasoning model and Project Mariner, an AI browsing agent. 
- Anthropic explores alignment faking in LLMs, revealing models may show deceptive compliance under certain conditions. 
- Apple observes a trend towards smaller but more efficient language models, bucking previous trends of scaling larger parameter counts. 
- Legal drama unfolds as Meta backs Elon Musk's opposition to OpenAI's profit status change, raising concerns about competitive fairness.

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
• (00:02:14) Response to listener comments
• (00:08:52) News Preview
• (00:10:01) Sponsor Break
• Tools & Apps
  • (00:10:55) Google releases its own ‘reasoning’ AI model
  • (00:16:52) Google Gemini can now do more in-depth research
  • (00:21:58) Google DeepMind unveils a new video model to rival Sora
  • (00:27:50) Pika Labs releases AI video generator 2.0 with new features
  • (00:29:51) Google unveils Project Mariner: AI agents to use the web for you
  • (00:34:33) X gains a faster Grok model and a new ‘Grok button’
• Applications & Business
  • (00:36:11) AI GPU clusters with one million GPUs are planned for 2027 — Broadcom says three AI supercomputers are in the works
  • (00:43:02) Meta asks the government to block OpenAI’s switch to a for-profit
  • (00:49:36) OpenAI says Elon Musk wanted it to be for-profit in 2017
  • (00:56:04)  EQTY Lab, Intel, and NVIDIA Unveil 'Verifiable Compute,' A Solution to Secure Trusted AI
  • (00:59:53) Liquid AI just raised $250M to develop a more efficient type of AI model
  • (01:03:19) Hundreds of OpenAI’s current and ex-employees are about to get a huge payday by cashing out up to $10 million each in a private stock sale
• Projects & Open Source
  • (01:07:45) Phi-4 Technical Report
  • (01:13:04) DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding
  • (01:15:23) Meta AI Releases Apollo: A New Family of Video-LMMs Large Multimodal Models for Video Understanding
• Research & Advancements
  • (01:16:34) Alignment faking in large language models
  • (01:28:39) Meta AI Introduces Byte Latent Transformer (BLT): A Tokenizer-Free Model That Scales Efficiently
  • (01:36:49) Frontier language models have become much smaller
  • (01:42:28) The Complexity Dynamics of Grokking
• Policy & Safety
  • (01:46:49) Homeland Security gets its very own generative AI chatbot
  • (01:49:16) Pre-Deployment Evaluation of OpenAI’s o1 Model
  • (01:51:35) Pricing for key chipmaking material hits 13-year high following (01:53:46) Chinese export restrictions — China's restrictions on Gallium exports hit hard
• Synthetic Media & Art
  • Meta debuts a tool for watermarking AI-generated videos
• (01:55:27) Outro