- Starter AI
- Posts
- OpenAI & FT team up, Sanctuary’s Phoenix, Introducing: The Instruction Hierarchy
OpenAI & FT team up, Sanctuary’s Phoenix, Introducing: The Instruction Hierarchy
The quest for data lives on.
Sthefania
April 29, 2024
Hello, Starters!
When we're away from the weekend and transition into Monday, it often feels like we've been missing for months. AI develops so rapidly that you might need to catch up again if you lose sight of it for even a moment. We're here to help!
Here’s what you’ll find today:
OpenAI and The Financial Times are teaming up
Sanctuary AI presents Phoenix
The Instruction Hierarchy: Steps to safety
Moderna’s AI-based approach
Vidu, a Sora-like text-to-video model
And more.
As the quest for data continues, OpenAI looks for more trustworthy providers. The newest one is the UK's Financial Times, with which the company has reached a licensing agreement.
Data is one of the fuels of AI, and as models continue to evolve, companies are redirecting a significant portion of their efforts toward seeking valuable information for training. In the past, this shift led to numerous copyright infringements. That's why OpenAI is focusing on doing things right, with most of its licensing deals involving media outlets. Additionally, data from journalism might help decrease biases and inaccuracies.
🤖 Sanctuary AI presents Phoenix (3 min)
Sanctuary AI has unveiled the 7th generation of Phoenix, its general-purpose robot, powered by Carbon, the company's groundbreaking AI control system. Phoenix can tackle a wide range of tasks, making it more than just "general" purpose; it might almost mimic human brain subsystems.
Upgrades in both hardware and software make Phoenix even faster, with enhanced sensory capabilities, while keeping manufacturing costs reduced. Sanctuary claims that Phoenix's similarities to humans make it capable of supporting workers with complex tasks.
A recent paper by OpenAI introduces the Instruction Hierarchy as a viable way of reducing the vulnerabilities of LLMs to prompt injections and jailbreaks. The proposition involves using the instruction hierarchy to give priority to prompts from trusted sources, while teaching LLMs to ignore instructions from unreliable sources and only allowing them if they align with the primary sources.
This results in models like GPT-3.5 enhancing their robustness, making them less prone to attacks while maintaining their functionality.
💉Moderna and OpenAI have been partners for a while. However, the company is now harnessing ChatGPT Enterprise to enhance most of its business processes aiming to use AI as a valuable asset to accelerate drug development, especially mRNA medicines.
🎞️With the arrival of Sora, the video generation field is slowly witnessing a change in the pace of innovation. Shengshu Technology, a Chinese AI company, has recently introduced Vidu—a model capable of creating 16-second HD videos at 1080p resolution with just a click.
⚡️Quick links
What did you think of today's newsletter? |