AI Roundup 086: Llama 3.2

Sep 27, 2024

∙ Paid

Meta's Connect event came and went this week, with a number of new hardware and AI announcements.

Everything unveiled this week:

Llama 3.2, an update to the Llama 3.1 family that now includes vision-based models.
Orion, a new AR-glasses prototype that combines generative AI and real-time visual displays.
Updates for the Ray-Ban Smart Glasses that include real-time AI video processing, live translation, QR code scanning, and an Audible integration.
New Imagine features are expanding across Facebook, Messenger, and Instagram, enabling users to generate AI photos and image captions.
And a look at how Meta's AI products are being used, with WhatsApp and international usage dominating Instagram.

Elsewhere in the FAANG free-for-all:

Microsoft previews Correction, a tool to factually incorrect AI-generated text.
A profile of Amazon executive Rohit Prasad and his team developing AI products for Alexa and other businesses.
Some key staff working on Microsoft's distilled AI models (such as Phi) have reportedly left the AI unit led by Mustafa Suleyman.
And the US, along with Meta, Anthropic, Google, IBM, Microsoft, Nvidia, and OpenAI, launched the Partnership for Global Inclusivity on AI to advance "sustainable" AI development.

Meanwhile, Google released some incremental updates to its models, Gemini 1.5 Pro and Gemini 1.5 Flash.

Why it matters:

Gemini doesn't get nearly as much press as ChatGPT, but it's quietly becoming a solid alternative. The fact that it has a 2 million token context window enables some unique applications.
With these updates comes yet another price drop, making Gemini 1.5 Pro by far the cheapest GPT-4 class model.
Hopefully, these upgrades can help Gemini capture more developer mindshare. Google's “enterprise-ready” SDKs can be a hurdle for programmers experimenting with AI (not to mention its questionable naming scheme).

Elsewhere in Google:

Google updated its AI note-taking assistant NotebookLM to let users get summaries of YouTube videos and audio files, and create sharable AI audio discussions.
Snap announced it will use Google's Gemini to help add new features to Snapchat's chatbot, including translating menus.
Warner Bros. Discovery partnered with Google to add its "caption AI" tool built on Vertex AI to Max for automatically generating captions.
Google updated Google Photos' mobile video editor for Android and iOS, adding editing tools and AI presets to make trimming and tweaking clips easier.
And Google plans to include the standalone Gemini app as standard in its Workspace Business, Enterprise, and Frontline plans from Q4, replacing the Gemini add-on.

After weeks of rumors, Reuters has confirmed OpenAI's plans to restructure its core business into a for-profit benefit corporation.

Between the lines:

OpenAI's non-profit status used to be seen as a way to prevent AGI from falling into the "wrong hands" - now, it's as an obstacle to raising the necessary funds to build AGI in the first place.
The shift would also give CEO Sam Altman equity for the first time - he has famously said that he owns no equity in OpenAI.
However, the company still seems to be leaking execs: CTO Mira Murati abruptly announced her departure, as did the company's Chief Research Officer.

Elsewhere in OpenAI:

Sources reveal that OpenAI is working on improving Sora, its AI video generator, to create clips more quickly.
OpenAI agrees to allow inspection of its training data in a copyright lawsuit filed by authors.
The company adds five new voices as it rolls out Advanced Voice Mode to ChatGPT Plus and Teams users.
Thrive Capital has already provided over $1 billion to OpenAI as part of a $6.5 billion funding round.
Jony Ive and Sam Altman are reportedly working to raise up to $1 billion for an AI device project by the end of 2024.
And OpenAI launches The OpenAI Academy to educate policymakers and the public about AI.