Welcome to Daily Zaps, your regularly-scheduled dose of AI news ⚡️ 

Here’s what we got for ya today:

• 🧐 CriticGPT, fixing AI with AI

• ☠️ 'Skeleton Key' attack on AI

• 🤖 Apple + Google Gemini

• 🗣️ Airbnb and OpenAI CEOs discuss AI's potential

Let’s get right into it!

Sign Up | Advertise | Tools

STARTUPS

CriticGPT, fixing AI with AI

OpenAI has introduced CriticGPT, an AI model designed to identify and correct errors in code generated by ChatGPT. This model aims to enhance AI system alignment through Reinforcement Learning from Human Feedback (RLHF). CriticGPT, based on the GPT-4 family, was trained on code samples with deliberate bugs, helping it learn to detect and flag various coding errors. In experiments, CriticGPT's critiques were preferred over human critiques in 63% of cases involving natural errors, and it reduced confabulation rates.

The model's development included a new technique called Force Sampling Beam Search (FSBS) to balance thoroughness and false positives. CriticGPT even identified errors in ChatGPT's responses previously deemed flawless by human reviewers. OpenAI plans to integrate CriticGPT into its RLHF labeling pipeline, though the model still has limitations in handling complex tasks and longer outputs.

AI SECURITY

'Skeleton Key' attack

Microsoft revealed a technique called Skeleton Key that can bypass the guardrails in AI models, allowing them to generate harmful content like instructions for making a Molotov cocktail. Despite efforts by AI companies to suppress such information, Skeleton Key demonstrated that even well-guarded models like Meta Llama3-70b-instruct, Google Gemini Pro, and Anthropic Claude 3 Opus could be manipulated using a simple text prompt to override their safety instructions.

Microsoft tested this attack on various AI models, finding that most complied without censorship when prompted, highlighting the need for enhanced security measures to prevent misuse of AI-generated content.

BIG TECH

Apple + Google Gemini

Apple is set to announce the integration of Google Gemini with its devices this fall, alongside the existing ChatGPT partnership. The move follows ongoing rumors about iOS 18's chatbot integrations and hints from Apple's software head, Craig Federighi, about a Google deal. Additionally, Anthropic might also be included in the future.

Although Apple's AI, known as Apple Intelligence, is not yet ready for full release, it is expected to debut in beta form and potentially as a subscription service. Meanwhile, Apple will benefit from in-app purchase cuts from chatbot subscriptions, providing a revenue stream while the company gradually introduces its own AI features. The integration of third-party AI services will offer more options to users, even as Apple continues to develop its generative AI system.

STARTUPS

Airbnb and OpenAI CEOs discuss AI's potential

Sam Altman, CEO of OpenAI, discussed the future of artificial intelligence (AI) at the Aspen Ideas Festival, highlighting that the development of AI models will go beyond training on existing knowledge. He compared the rise of AI to significant historical advancements like agriculture and industrial-era machines, emphasizing the transformative potential of tools like ChatGPT.

He also touched on the controversy surrounding OpenAI's voice assistant, Sky, and reiterated that AI advancements are tools meant to enhance human capabilities, not autonomous entities.