AI News April 7th 2024

Researchers from Princeton NLP developed SWE-agent, an open-source project that turns GPT-4 into an AI software engineer. The agent achieved accuracy similar to Devin and resolved 12.29% of issues autonomously.

https://twitter.com/_carlosejimenez/status/1775157915789463935

OpenAI made a big upgrade to DALL-E 3, now allowing users to edit images directly in ChatGPT. Users can edit images directly in the chat across the web, iOS, and Android apps by selecting an area of the image and prompting changes.

https://twitter.com/OpenAI/status/1775569161759985737

Apple researchers introduced ReALM. It's a new AI system that can understand on-screen tasks, conversational context, and background processes. The model 'substantially outperformed' GPT-4 in some areas.

Microsoft and OpenAI are planning project "Stargate" — a $100 billion data center project. The project could reportedly cost over $100 billion, which is 100 times more expensive than today's largest data centers.

Stanford researchers just introduced Octopus v2, a new framework for on-device AI agents. With just 100 training samples, the model achieved 98% accuracy in selecting the right function, surpassing GPT-4.

https://twitter.com/nexa4ai/status/1775993692509323582

Stability AI just released Stable Audio 2.0, an updated AI audio gen model that can create songs up to 3 minutes long from a prompt. The new model even introduces audio-to-audio, allowing users to upload and edit their own audio samples using prompts.

https://twitter.com/StabilityAI/status/1775501906321793266

Higgsfield AI revealed a new foundational model that can create personalized character videos for social media. Alongside the announcement, they dropped a new iOS app that allows users to take a selfie and get a personalized character.

https://twitter.com/higgsfield_ai/status/1775523757136007354

Researchers from Google DeepMind introduced Mixture-of-Depths, a method for dynamically allocating compute in transformer-based models. It allocates compute across sequences, reducing computational costs without sacrificing performance.


Researchers introduced CameraCtrl, a method that lets users control the camera movement in AI-generated videos. It allows for camera control of different video styles, allowing users to generate AI video by specifying the text and camera movement.

https://twitter.com/_akhaliq/status/1775345383943659966

Anthropic researchers discovered a new "jailbreaking" technique called "many-shot jailbreaking". It evades the safety guardrails of LLMs by exploiting expanded context windows.

https://twitter.com/AnthropicAI/status/1775211248239464837

Hope you enjoyed this weeks AI news breakdown. Let me know if you want to know more about any of these topics in a video!