AI Breakfast
Posts
Could OpenAI's 'Sora' Be Released Today?

Could OpenAI's 'Sora' Be Released Today?

December 09, 2024

Good morning. It’s Monday, December 9th.

Did you know: You can watch a livestream of Day 3 of OpenAI’s 12 days of Christmas today at 10:00am PST?

In today’s email:

Sora Drop Coming Soon?
Aurora Image Generator Pulled
Meta Llama 3.3
OpenAI’s 12 Days of Christmas Continues
5 New AI Tools
Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

^{In partnership with MASTERWORKS}

Over the last seven elections, this asset class has outpaced the S&P 500

Instead of trying to predict which party will win, and where to invest afterwards, why not invest in an ‘election-proof’ alternative asset? The sector is currently in a softer cycle, but over the last seven elections (1995-2023) blue-chip contemporary art has outpaced the S&P 500 by 64% even despite the recent dip, regardless of the victors, and we have conviction it will rebound to these levels long-term.

Now, thanks to Masterworks’ art investing platform, you can easily diversify into this asset class without needing millions or art expertise, alongside 65,000+ other art investors. From their 23 exits so far, Masterworks investors have realized representative annualized net returns like +17.6%, +17.8%, and +21.5% (among assets held longer than one year), even despite a recent dip in the art market.*

It's easy to get started at Masterworks, and our readers can SKIP THE WAITLIST HERE.

Past performance not indicative of future returns. Investing Involves Risk. See Important Disclosures at masterworks.com/cd.

^{Thank you for supporting our sponsors!}

Today’s trending AI news stories

OpenAI Signals Imminent Release of Sora Video Generator

OpenAI is on the cusp of unveiling an enhanced version of its Sora video generator, integrating sophisticated capabilities like text-to-video, text-and-image-to-video, and text-and-video-to-video generation, accommodating clips up to one minute in duration.

Sora v2 release is impending:
* 1-minute video outputs
* text-to-video
* text+image-to-video
* text+video-to-video
OpenAI's Chad Nelson showed this at the C21Media Keynote in London. And he said we will see it very very soon, as @sama has foreshadowed.
— Ruud van der Linden (@RuudNL)
3:57 PM • Dec 7, 2024

OpenAI’s Chad Nelson revealed these details during the C21Media event in London, corroborated by recent API leaks indicating a faster, more efficient design. The rollout, potentially aligned with OpenAI’s winter event, could also introduce GPT-4.5 and advanced image-generation updates for GPT-4o. If realized, this suite of enhancements would further consolidate OpenAI’s standing in generative AI. Read more.

𝕏 adds, then quickly removes, Grok’s new ‘Aurora’ image generator

Elon Musk's 𝕏 briefly debuted Aurora, a photorealistic image generator integrated into the Grok assistant, only to remove it shortly after. Listed as “Grok 2 + Aurora (beta),” it was replaced by “Grok 2 + Flux (beta)” within days. Musk admitted the tool’s beta status, promising swift upgrades. While Aurora’s ability to conjure images of public figures and copyrighted characters caught eyes, flaws like anatomical distortions and provocative outputs sparked scrutiny over content moderation and technical maturity.

Grok Aurora’s celebrity likenesses are scary good.
— fofr (@fofrAI)
5:38 PM • Dec 7, 2024

Aurora’s rollout aligns with Grok’s pivot to a freemium model, offering 10 prompts and three image analyses every two hours, expanding access beyond its initial paywall. Developed by xAI, which recently secured $6 billion in funding, these shifts suggest a calculated push to broaden AI adoption while grappling with thorny challenges of compliance and refinement. Read more.

Meta Launches Open Source Llama 3.3, Shrinking Powerful Model Into Smaller Size

Meta’s Llama 3.3 is a leaner, meaner version of its predecessor, packing the punch of a 405-billion-parameter model into a far more efficient 70 billion parameters. This upgrade slashes GPU memory needs—up to 24 times less load for the same performance—making it a cost-effective choice for developers.

Run Llama 3.1 1B offline at 60+ tok/s. Entirely on your phone.
Siri can’t save you offline, but a local model just might.
— Aaron Ng (@localghost)
8:23 PM • Dec 8, 2024

At just $0.01 per million tokens for generation, it’s a competitive alternative to GPT-4 and Claude. Llama 3.3 shines in multilingual reasoning with a 91.1% accuracy on MGSM, outdoing Amazon’s Nova Pro in key NLP benchmarks. Trained on 15 trillion tokens and fine-tuned on 25 million synthetic examples, it features a 128k token context window and Grouped Query Attention for enhanced scalability.

Meta also made strides in sustainability, ensuring a net-zero emission footprint despite the hefty computational demands. Llama 3.3 is a powerful, budget-friendly tool with an environmental conscience. Read more.

OpenAI's 'Ship-Mas' Continues with Reinforcement Fine-Tuning for Expert Models

OpenAI's new Reinforcement Fine-Tuning (RFT) method gives its o1 models the ability to tackle complex tasks with minimal training data. Unlike traditional fine-tuning, RFT allows models to explore potential solutions before evaluation, reinforcing effective reasoning and penalizing mistakes.

In collaboration with Thomson Reuters, RFT demonstrated its prowess, with the o1 Mini model outperforming standard versions in legal tasks. RFT also proved valuable in bioinformatics, hitting 45% accuracy in gene identification for rare genetic diseases.

OpenAI is now inviting select organizations to participate in an early access research program, with wider availability slated for 2025, cementing RFT's potential in expert systems and niche AI applications. Read more.