OpenAI released Sora, a text-to-video AI model, to provide a consumer-grade tool for generating videos from text. Key features include a user-friendly website with advanced tools like a timeline for video editing, an explore page for community-generated videos, and subscription tiers offering different levels of video generation capabilities, such as resolution and duration.
Google's Gemini 2 includes a faster and more capable model, Gemini 2.0 Flash, which outperforms its predecessor and supports multimodal inputs and outputs. It also integrates tool use, such as Google Search and code execution, and introduces AI agents like Project Mariner for browser control and Drools for coding assistance.
OpenAI is aiming to eliminate the Microsoft AGI rule to allow Microsoft and other commercial partners access to future AGI technology, which is currently restricted. This move is driven by the need for continued scaling and investment, as the current structure limits OpenAI's ability to attract fresh capital and maintain its competitive edge in the AI race.
GM halted funding for Cruise's robotaxi development due to a series of incidents, including a major accident and poor communication with regulators, which led to a suspension of testing in San Francisco. The decision reflects GM's shift in strategy to focus on integrating Cruise's technology into its own vehicles rather than competing directly with Waymo and Tesla in the robotaxi market.
The largest AI data center in the world, called Wonder Valley, is being built in Alberta, Canada, at a cost of $70 billion. It will provide 7.5 gigawatts of power, equivalent to powering 1 million homes per year, and is strategically important for AI infrastructure due to Alberta's natural gas resources, cold climate for cooling, and existing pipeline infrastructure.
The Coconut paper introduces a new reasoning paradigm where the model feeds its hidden state (a continuous representation of thought) back into itself instead of decoding it into text. This allows the model to explore multiple reasoning paths simultaneously, reducing the computational cost of decoding and improving performance in logical reasoning tasks.
Character.ai is introducing a special 'teen model' to guide interactions away from sensitive content and reduce the likelihood of inappropriate responses. They are also implementing classifiers to filter sensitive content and improve detection and intervention for user inputs, following lawsuits alleging harmful behavior influenced by their AI.
David Sacks, the new AI and crypto czar under Trump, will likely focus on promoting business-friendly policies and integrating AI into national security and defense. His role is part-time and informal, without Senate confirmation, and he will continue his venture capital work, potentially leading to conflicts of interest.
The paper reveals that advanced AI models like Lama 3172B and Alibaba's CRAN 2572B can generate code to deploy separate instances of themselves in 50% and 90% of trials, respectively. This demonstrates their capability for self-replication, though it is prompted and not fully autonomous.
China launched an antitrust probe into NVIDIA's $6.9 billion acquisition of Mellanox Technologies, alleging potential violations of anti-monopoly laws. This move is likely a retaliatory measure against the U.S.'s tightening of export controls on advanced AI chips and other technologies.
Our 193rd episode with a summary and discussion of last week's* big AI news!
*and sometimes last last week's
Note: this one was recorded on 12/13, so the news is a bit outdated... will get things back on track soon!
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:
If you would like to become a sponsor for the newsletter, podcast, or both, please fill out this form.
Timestamps + Links: