• AI Emergence
  • Posts
  • Ilya Sutskever is aligned to Safe Super Intelligence!

Ilya Sutskever is aligned to Safe Super Intelligence!

Along with: Industrial Research Labs facing existential questions

Hey there, 

Ilya, to his credit, has announced SSI Inc (Safe Super Intelligence Inc.). While details are still awaited, it remains to be seen if he can find a way to stay not-for-profit and achieve Safe SuperIntelligence. 

On the other hand - OpenAI is looking to convert to a for-profit and Deepmind is becoming an AI factory. At a time when the world is waking up to the impact of AI, the pioneers are juggling to find the right structures to handle something we have never seen before!

At times I wonder, where we are heading :)

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

Let’s look at the developments this week

Table of Contents

Ilya Sutskever, OpenAI’s co-founder and former chief scientist, has launched a new AI company focused on safety called Safe Superintelligence Inc. (SSI).

SSI aims to advance AI capabilities while prioritizing safety and operating free from short-term commercial pressures. The startup emphasizes a singular focus on safety and progress, without distractions from management overhead or product cycles.

As major players like OpenAI, Google, and Microsoft push forward in AI, it will be fascinating to see how SSI performs. (source)

Over a year after merging DeepMind and Google Brain into the AI "super unit" Google DeepMind, the integration has faced significant hurdles. Bloomberg reports internal conflicts, with DeepMind researchers frustrated by imposed roadmaps that diverge from their research-focused origins. 

This friction arises amidst fierce AI competition and the troubled launch of AI Overviews, a product intended to improve search responses, which instead drew criticism for generating inappropriate suggestions. (source)

Google DeepMind has developed a new AI model called V2A (Video-to-Audio) which synchronizes soundtracks and dialogue with video content.

This model uses visual cues from videos and text prompts to produce realistic audio, enhancing AI-generated videos with sound effects and matching dialogue.

V2A is compatible with Google’s Veo model and can work with various types of video content, including silent films.

It allows users to customize soundtracks with positive or negative prompts and uses watermarking to ensure authenticity. Although promising, V2A is not yet available to the public due to potential limitations and misuse concerns. (source)

Runway has introduced Gen-3 Alpha, a significant upgrade in its series of models designed for large-scale multimodal training. 

This new model enhances fidelity, consistency, and motion, advancing toward General World Models. 

Gen-3 Alpha supports Runway's suite of tools, from Text to Video to Advanced Camera Controls, and introduces precise temporal control for dynamic content creation. 

It also excels at producing photorealistic human characters, offering vast storytelling possibilities. Additionally, Gen-3 Alpha incorporates improved safety measures and allows for extensive customization in partnership with industry leaders. (source)

CEO Sam Altman is considering transitioning OpenAI to a for-profit benefit corporation, which would enable it to seek profits while still aiming for societal benefits, akin to its competitors Anthropic and xAI.

Although discussions are still fluid, this shift could redefine OpenAI's non-profit commitments. In response to these rumors, OpenAI emphasized its dedication to building AI that benefits all and maintaining its nonprofit mission. (source)

Luma Labs has launched its Luma API, a powerful tool for developers to create high-quality 3D models and environments quickly and affordably. 

By utilizing video walkthroughs, the API converts scenes into interactive 3D models or pre-rendered images and videos, supporting a broad range of applications from e-commerce to VFX. 

The API is designed to be cost-effective, dramatically reducing the time and expense traditionally associated with creating detailed 3D models. 

Continuous improvements are expected in output quality and processing times, and Luma encourages feedback for further enhancements. (source)

NVIDIA has released the Nemotron-4 340B, an open model designed for generating synthetic data to train LLMs across various industries. 

This model facilitates the creation of high-quality synthetic data that mirrors real-world scenarios, enhancing the training and accuracy of custom LLMs. 

Optimized for NVIDIA's NeMo and TensorRT-LLM frameworks, Nemotron-4 340B supports both data generation and efficient inference processes. It offers developers a scalable solution to improve model performance and is equipped with a robust visual moderation system to ensure the safety and relevance of the generated data. (source)

Researchers at the University of Tokyo, have developed a "musculoskeletal humanoid" robot named Musashi, designed to drive a car. 

Equipped with camera eyes, mechanical hands, and anti-slip feet, Musashi can operate car functions like the steering wheel, brakes, and signals. 

It successfully navigated a test track, making cautious turns and managing speed variations on inclines. However, it displayed cautious driving behavior due to technical limitations, exemplified by a slow cornering maneuver that took about two minutes. (source)

Zeta Labs, co-founded by former Meta engineers Fryderyk Wiatrowski and Peter Albert, launched an AI web agent named Jace, capable of executing browser-based tasks autonomously. 

Supported by a $2.9 million pre-seed funding led by high-profile tech leaders, Jace leverages a blend of LLM and proprietary web-interaction models to manage complex tasks across various platforms without human oversight. 

Designed primarily for e-commerce, recruitment, and marketing, Jace aims to automate repetitive tasks, promising a subscription model post a free usage tier. (source)

OpenAI has stirred controversy by appointing former NSA director Gen. Paul Nakasone to its board, prompting strong criticism from whistleblower Edward Snowden.

Snowden warns against trusting OpenAI, accusing it of betraying public rights by including a high-level intelligence figure in its governance.

This appointment is part of a broader leadership reshuffle aimed at enhancing cybersecurity measures in response to complex threats, following tumultuous changes in OpenAI's executive team, including the temporary firing and rehiring of CEO Sam Altman. (source)

TikTok recently introduced Symphony Digital Avatars, lifelike AI clones, and a new AI dubbing tool. 

These avatars, which mimic creators or stock actors, can be used in advertisements, speaking over 30 languages to enhance global reach. 

The AI dubbing feature allows videos to be dubbed in multiple languages, improving accessibility and personalization of content. This technology is part of TikTok's strategy to appeal to international audiences and is currently available to a select group of users. (source)

Tool: Jammable

Creating engaging and memorable experiences at conferences can be challenging. Jammable offers a unique platform to create customized, high-quality song covers to enhance the ambiance and engage attendees.

Problem Statement: Suppose you are organizing the DataHack Summit and want to provide an engaging musical experience for attendees. Use Jammable to create customized song covers that reflect the event's theme and energize the audience.

How to access the tool:

1. Sign up or log in to Jammable.

2. Select the songs you want to cover.

3. Customize the cover with desired styles and themes.

4. Use AI-generated covers to create a unique soundtrack for your event.

5. Integrate the music into various segments of the conference to enhance the overall experience.

Here is how you can utilize this application:

Use Jammable to produce tailored song covers that match the DataHack Summit's theme, creating a unique and memorable atmosphere for all attendees.

  • In the recent episode of Leading With Data, I had a chat with Dr. Manish Gupta about his extensive experience in spearheading innovative projects and research in the fields of video technology and AI across global tech giants like VideoKen, Xerox, and IBM.

I wonder what is the best way/structure for an organisation to bring AI to the world. The conflict reflects on so many levels. First, you need to have a structure that inspires the best talent in the world to join you.

Next, you need to think about your approach to product launches. OpenAI could launch fast, iterate fast, and tackle safety as it scaled the product in its early days. Google, at its scale, could not. They (aimed to) ensure Safety on Day 0.

Then there is a whole debate on whether the company should be for-profit or not-for-profit.

And that directly decides whether and how much capital you can raise.

Looking at all these challenges, it feels like the best company to have managed this for the longest time is still Google. They built and brought several technologies to the world including the Transformers. They are a for-profit company but had kept safety at the heart of their launches (unless they were forced to respond to competition). They have organized the knowledge for the world and made it a better place.

Which structure do you think is best suited to bring AI to the world?

Login or Subscribe to participate in polls.

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Join the conversation

or to participate.