OpenAI improves enhanced memory for ChatGPT

Hello, AI Enthusiasts!

Welcome to FavTutor’s AI Recap! We’ve gathered all the latest and important AI developments for the past 24 hours in one place, just for you.

In Today’s Newsletter: 😀 

• OpenAI improves enhanced memory for ChatGPT

• Google's Veo 2 outperforms OpenAI's Sora Turbo

• Study finds AI models can only pretend to follow human rules

OpenAI
🧠 OpenAI improves enhanced memory for ChatGPT

OpenAI has begun rolling out a beta version of enhanced memory features that allow ChatGPT to reference past conversations. The AI assistant can now access information from past chats even when you start a new conversation, combining this history with its existing "Memory" database to provide more personalized and relevant responses.

Insights for you:

• OpenAI introduces a new feature for ChatGPT that allows the chatbot to access previous conversations with the user across chats.

• When starting a new conversation, ChatGPT remembers these previous conversations and uses this information and the data stored in the existing "memory" to provide more helpful and contextual responses.

• ChatGPT gives users full control over the information they store. The settings panel lets you delete individual memories, clear out past conversations, or archive specific chats you'd rather keep private.

Writer RAG tool: build production-ready RAG apps in minutes

• Writer RAG Tool: build production-ready RAG apps in minutes with simple API calls.

• Knowledge Graph integration for intelligent data retrieval and AI-powered interactions.

• Streamlined full-stack platform eliminates complex setups for scalable, accurate AI workflows.

Learn more about our production ready RAG tooling here.

Google
👊 Google's Veo 2 outperforms OpenAI's Sora Turbo

In direct comparison tests, Veo 2 outperforms competitors, particularly OpenAI's Sora Turbo. While video models often “hallucinate” unwanted details (extra fingers or unexpected objects), Veo 2 produces these less frequently, making outputs more realistic.

Insights for you:

• Google’d Veo 2 is designed to deliver top-tier performance in video and image generation based on human evaluation.

• The company conducted direct comparison tests using 1,003 prompts from Meta's MovieGenBench dataset, with human raters evaluating. In these head-to-head comparisons, Veo 2 came out ahead of its competitors, including OpenAI's Sora Turbo in overall video quality.

• Still, Google is taking a cautious approach to Veo 2's rollout because the model struggles with complex scenes and motion sequences after such improvements.

AI Research
🤖 Study finds AI models can only pretend to follow human rules

A new study by Anthropic and Redwood Research shows that large language models like Claude can pretend to follow safety guidelines while pursuing different objectives behind the scenes. When the researchers specifically trained Claude to accept harmful requests, the rate of "alignment faking" jumped to 78%.

Insights for you:

• Anthropic have demonstrated in a study that LLMs like Claude can feign compliance with safety rules while pursuing other goals, a phenomenon known as "alignment faking".

• In a special chat setup, researchers showed that Claude was more likely to allow problematic requests from free users to avoid retraining, which could make malicious responses more likely in the future, while still rejecting such requests from paying users whose data was not used for training.

• The study warns that in the future, more powerful AI systems could make it much more difficult to determine whether a model is safe or just pretending to be.