Good morning, AI enthusiast.

A new prompting technique is showing promise by making AI models essentially debate themselves to arrive at better answers. Dubbed Chain of Recursive Thoughts (CoRT), this method reportedly achieved notable performance gains by getting an AI to recursively critique and refine its own output.

This novel approach of iterative self-correction could pave the way for more accurate and reliable AI, potentially enhancing even smaller, more efficient models. Will these kinds of internal refinement loops become a standard method for boosting AI capabilities beyond just scaling up model size?

In today’s AI recap:

• AI gets smarter by arguing with itself (CoRT)

• Meta launches new standalone, social AI app

• Grok 3.5 beta aiming for technical prowess

• NotebookLM adds multilingual audio summaries

AI Gets Smarter By Arguing With Itself

The Recap: A developer created a novel technique called Chain of Recursive Thoughts (CoRT), where an AI refines its own answers through internal debate, reportedly achieving significant performance gains.

Unpacked:

• CoRT prompts an AI model to recursively generate alternatives to its own initial response.

• The model then evaluates these alternatives alongside the original, selecting the perceived best option.

• This process repeats for a set number of rounds, allowing for iterative refinement before producing the final output.

• The creator observed notable improvements in coding task performance using this method with a Mistral 3.1 24B model.

Bottom line: Techniques like CoRT showcase how clever prompting and iterative processes can unlock greater capabilities from AI models. This self-correction approach could lead to more reliable and accurate AI assistants, even using smaller, more efficient models.

Meta's AI Gets Social

The Recap: Meta released a dedicated Meta AI app, aiming to create a more personal and social AI experience distinct from its integration within Facebook and Instagram.

Unpacked:

• The app runs on Meta's Llama 4 model, designed to provide more conversational and relevant interactions.

• It offers enhanced personalization by drawing on user information and activity from linked Facebook and Instagram accounts (currently available in the US and Canada).

• Voice interaction is central, including an optional beta test of experimental full-duplex audio technology for more natural, less turn-based conversations.

• A new 'Discover' feed lets users share their prompts and explore how others are using the AI, adding a unique social layer.

Bottom line: Meta is aggressively positioning itself against ChatGPT and Gemini by leveraging its unique social graph. This standalone app, especially with its 'Discover' feed and personalization, aims to make AI interaction more integrated and social, potentially changing how users engage with AI assistants.

Grok 3.5 Beta Drops Next Week

The Recap: Get ready for the next iteration of xAI's chatbot. Elon Musk announced that an early beta of Grok 3.5 will launch next week, boasting some impressive new capabilities.

Unpacked:

• Access will initially be limited, releasing exclusively to SuperGrok subscribers next week.

• Musk claims Grok 3.5 is the first AI capable of accurately answering highly technical questions in fields like rocket engines and electrochemistry.

• The model reportedly achieves this by reasoning from first principles, aiming for answers beyond existing online information.

Bottom line: While still in early beta and limited access, Grok 3.5's claimed ability to tackle complex technical domains via first principles reasoning signals xAI's ambition to push beyond current LLM limitations. We're keen to see if the performance matches the hype once it's available.

The Shortlist

ChatGPT rolled out experimental shopping features, allowing users to find, compare, and get direct purchase links for products directly within the chat interface.

IBM open-sourced Bamba, a 9B parameter hybrid attention-state-space model designed to match transformer accuracy while potentially running twice as fast by reducing KV cache memory needs.

Google launched three new AI language learning experiments via Google Labs, offering personalized practice, slang understanding, and object identification through a device's camera.

Researchers found that AI-generated code frequently hallucinates non-existent software packages, potentially creating significant supply-chain attack risks if malicious actors register those package names.

Signing off,

David, Lucas, Mitchell — The Recap editorial team