AI Roundup 095: QwQ

November 29, 2024.

Nov 29, 2024

∙ Paid

QwQ

This week Alibaba released QwQ-32B-Preview, a new open-source AI model that supposedly rivals OpenAI's o1 in reasoning capabilities.

Between the lines:

The model demonstrates impressive capabilities in math and logic tasks, even outperforming OpenAI's o1-preview on certain benchmarks.
In some ways, the model is more open than o1 - the model has an Apache license, its "reasoning" tokens are visible, and its weights have been published. Yet there are still limitations, including clear signs of Chinese regulatory compliance.
Reasoning models like QwQ-32B-Preview reflect a broader shift as major labs explore new architectures and techniques, such as test-time compute, to advance AI capabilities.

Elsewhere in frontier models:

Amazon is reportedly developing a new multimodal LLM, codenamed Olympus, capable of processing images, video, and text.
Nvidia demoed Fugatto, an AI model for generating music and audio that can also modify voices, trained on open-source data, as the company weighs whether to release it.
And Anthropic introduced the Model Context Protocol, an open-source standard for connecting AI assistants to data sources, plus custom writing styles for Claude.

Amazonthropic

Amazon doubles down on its Anthropic investment with an additional $4B, but with a significant condition: Anthropic will use both AWS and Amazon's custom chips to train its models.

Why it matters:

This deal represents more than just funding - Amazon is making a strategic move to secure its position in AI infrastructure, given Anthropic’s massive compute needs.
While Anthropic reportedly prefers Nvidia chips, the company's projected $2.7B burn rate for 2024 likely made Amazon's terms too attractive to refuse.
Yet, despite both Google and Microsoft under fire for potential AI antitrust concerns, it remains to be seen whether Amazon will face similar scrutiny, as this now makes $8B invested into Anthropic.

Elsewhere in the FAANG free-for-all:

Microsoft beta tests its Recall AI feature for Copilot+ PCs, including Click to Do, a feature similar to Google's Circle to Search.
Sources say xAI plans to debut a consumer app as soon as December, as the company is set to close a new $5B funding round.
And Threads is testing AI-generated summaries of user discussions in the app's Trending Now section in the US.

Sora subject

On Tuesday, artists temporarily leaked access to OpenAI's Sora video generation tool, protesting the company's treatment of early testers.

Between the lines:

A small group of artists publicly took credit for the leak with an open letter criticizing OpenAI's practice of using creative professionals as "PR puppets" with minimal compensation.
The leaked access allowed public users to generate 10-second videos using Sora (via HuggingFace) but was stopped after OpenAI shut down all Sora API endpoints.
While OpenAI maintains that participation was voluntary, this is only the latest incident in mounting tensions between AI companies and creative professionals.

Elsewhere in AI anxiety:

Writers and publishers criticize Spines, a startup that charges authors to use AI for editing and distributing books.
An analysis finds over 54% of longer English-language posts on LinkedIn are likely AI-generated.
US coding boot camp graduates face a tough job market due to AI coding tools and mass layoffs, with developer job listings down 56% since 2019.
And utilities, regulators, and researchers in six countries say the power demand surge caused by AI and data centers is being met in the near-term by fossil fuels.