• Starter AI
  • Posts
  • Stability’s Stable Audio 2, Meet SWE-agent, Benign data vs. AI safety

Stability’s Stable Audio 2, Meet SWE-agent, Benign data vs. AI safety

Say goodbye to bugs!

Hello, Starters!

Streamline, optimise, improve—you might already be familiar with these words, and how could you not be? Especially if you're a developer searching for AI tools to help you code. Luckily, there are tons!

Here’s what you’ll find today:

  • Stability AI introduces Stable Audio 2

  • Princeton’s groundbreaking SWE-agent

  • Can benign data undermine AI safety?

  • Google’s AI features come at a price

  • Xbox might soon have an AI chatbot

  • And more.

Stability AI has recently launched Stable Audio 2, the latest iteration of its AI music generation tool with many features that make it stand out. Stable Audio 2 can compose songs up to three minutes long in high quality while also incorporating a new audio-to-audio feature that allows users to upload their files to craft fully produced samples or adjust them to a desired style.

This tool comes at a time when both Stability AI and AI music are in the spotlight, with the startup going through internal turmoil and prominent figures in the music industry releasing an open letter discussing the impact of AI on creativity.

Princeton's NLP team has presented SWE-agent, which has been designed to turn language models like GPT-4 into software engineering agents that address bugs and repository issues on GitHub in real time.

This AI agent showcases a 12.29% issue resolution, putting it as a close contender to Cognition's Devin, which is at 13.86%. However, SWE-agent is an open-source project and could potentially improve further thanks to user feedback.

A paper from Princeton University has revealed that fine-tuning models with benign data can inadvertently degrade safety. This means that, despite the efforts to train current LLMs with data that aligns with human values and assures the safe development of AI technologies, there are specific subsets of benign data that are prone to jailbreaking.

Recently, we've also covered Anthropic's approach to this phenomenon, which raises the question of what measures should be implemented to prevent bad actors from using these weaknesses in AI to their advantage.

💵According to sources, Google is looking for ways to integrate its AI-powered search into a paid service for its users. While regular search might remain free, those with a subscription will have access to results boosted by Gemini.

🎮Microsoft is developing an AI chatbot for Xbox as part of its effort to improve its platform and services. Reportedly, this would be an "embodied AI character" aimed at helping users with queries related to support and refund processing.

What did you think of today's newsletter?

Login or Subscribe to participate in polls.

Thank you for reading!