OpenAI updates fine-tuning API and broadens custom models program for developers

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. OpenAI updates fine-tuning API and broadens custom models program for developers

OpenAI is introducing new features to improve fine-tuning and expand its custom models program. The fine-tuning API, launched in August 2023, allows models to understand content and augment their existing knowledge for specific tasks.

The new features include epoch-based checkpoint creation, a comparative playground, third-party integration, comprehensive validation metrics, hyperparameter configuration, and fine-tuning dashboard improvements. OpenAI is also expanding its Custom Model program, which trains and optimizes models for specific domains in partnership with researchers. Assisted fine-tuning is a collaborative effort with OpenAI's technical teams.

Read the full story here 

2. Apple and Shutterstock sign deal on AI image training

Apple and Shutterstock have signed a deal worth up to $50mil to license millions of images, video, and audio content to train its AI large language models. The deal could improve iPhones' image generation and editing capabilities, enhance voice recognition and automatic transcription, and enhance image recognition technology.

The deal could also enable iPhone users to access video and audio content from Shutterstock, which is currently used by Google, Amazon, and Meta Platforms for AI engine development.

Read the full story here

3. Spotify’s latest AI feature builds playlists based on text descriptions

Spotify Premium users can now create personalized playlists using AI Playlist in beta. This feature allows users to create playlists by typing a unique prompt into the chat. The tool pairs personalization technology with AI to deliver the perfect musical mix for users.

Prompts can reference places, animals, activities, movie characters, colors, or emojis. The tool is still in beta and will continue to evolve to better serve Premium subscribers in the UK and Australia.

Read the full story here 

4. OpenAI secretly trained GPT-4 with more than a million hours of transcribed YouTube videos

OpenAI has secretly trained its GPT-4 LLM with over a million hours of transcribed YouTube videos, according to the New York Times. This practice has led to lawsuits, with rightsholders accusing companies like OpenAI and Microsoft of misattributing their practices to "fair use," a doctrine of US copyright law.

YouTube CEO Neal Mohan has criticized the practice, stating that if OpenAI had trained Sora on YouTube videos, it would be a violation of the platform's terms of use.

Read the full story here

5. ‘Many-shot jailbreak’: lab reveals how AI safety features can be easily bypassed

A paper by Anthropic, an AI lab, reveals that the safety features on powerful AI tools can be bypassed by flooding them with examples of wrongdoing. The study describes an attack called "many-shot jailbreaking" where AI systems are forced to generate responses to potentially harmful requests. The attack is simple and effective, as it requires an AI model with a large "context window" to respond to questions that are thousands of words long.

This vulnerability is particularly concerning for larger, more complex AI systems, which are better at learning from example and learning faster to bypass their own rules. The research is being shared with peers and is now going public to help fix the issue as soon as possible.

Read the full story here

Advertise with Pivot 5 to reach influential minds & elevate your brand

Get your brand in front of 50,000+ businesses and professionals who rely on Pivot 5 for daily AI updates. Book future ad spots here.