Daily AI Newsletter: Introducing r1. Your pocket companion

In today’s email:

🔓 How to Persuade ChatGPT to Jailbreak it
🔥 MultiOn: Building a Brighter Future for Humanity with AI Agents
🕵🏻‍♂️ AI aids nation-state hackers but also helps US spies to find them, says NSA cyber director
🧰 15 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Introducing r1. Your pocket companion

Today, Rabbit's new AI device, the R1, was introduced by CEO Jesse Lyu. Priced at $199, the R1 is a compact, standalone AI gadget with a 2.88-inch touchscreen, a rotating camera, a scroll wheel/button for navigation, a 2.3GHz MediaTek processor, 4GB of memory, and 128GB of storage. It features a design collaboration with Teenage Engineering and promises an "all day" battery life. Despite its light and well-designed hardware, the R1's software, Rabbit OS, is its main highlight. Rabbit OS, unlike a large language model, is based on a "Large Action Model" (LAM) and functions as a universal controller for apps, similar to Alexa or Google Assistant. It can perform various tasks like controlling music, ordering cars, buying groceries, and sending messages through a single interface.

The R1 can be trained to perform new actions and is integrated with a web portal called Rabbit Hole for logging into various services. Rabbit's strategy avoids the need for building APIs and convincing developers to support the R1. Instead, the LAM is trained on existing apps. While the R1 can perform some tasks independently, teaching it new tasks involves using Rabbit's virtual machines.

Rabbit OS could revolutionize the app store, similar to how ChatGPT could change web search. However, the R1's actual practical utility and its ability to integrate with various platforms and devices remain to be seen. The device is available for pre-order, with shipments expected to start in March.

Watch the keynote here.

OpenAI Accidental Release - GPTs with Memory 😍

For a brief moment yesterday, some ChatGPT users started seeing this popup on their accounts. The screenshot refers to the new GPT memory feature. This is what OpenAI refers to as Project Sunshine and we have been seeing in previous leaks from OpenAI.

Along with the new popup, you could also find a new section under the settings menu called “Personalization” which is where you can toggle the memory option ON/OFF.

After about 20 minutes, the new setting went away and couldn’t be found under the settings menu anymore. Testing the ChatGPT also indicates that there is currently no actual memory in the chat.

Although this appears to be an accidental release by OpenAI, there is a high chance that OpenAI soon release this feature.

Announcing MultiOn: Building a Brighter Future for Humanity with AI Agents

MultiOn is an innovative AI startup founded by Stanford alumni Div Garg and Omar Shaya. Their vision is to create AI agents that perform mundane tasks, thus freeing people to focus on more meaningful activities. MultiOn's AI agents can quickly handle tasks like booking travel itineraries or making restaurant reservations, tasks that typically take much longer for humans.

Div Garg, previously juggling multiple jobs and interested in autonomous AI systems, left his PhD program at Stanford to start MultiOn. His background includes work on Large Language Models (LLMs), Reinforcement Learning, and notable contributions at companies like Apple, Google, and Nvidia.

Omar Shaya, with a strong background in computer science and consumer products, was pursuing an MBA at Stanford when he met Div. He has experience at Microsoft and Meta, working on consumer AI products, and his insights have significantly influenced MultiOn's product vision.

The company's mission is to augment human potential through AI, enabling concurrent task handling and transforming daily life efficiency. MultiOn has already demonstrated capabilities like booking flights, ordering food, and making reservations. It has attracted significant attention and investment from major industry players and is recognized for its potential to revolutionize the AI landscape.

New Paper: How Johnny Can Persuade LLMs to Jailbreak Them

The project described focuses on using a range of persuasion techniques to "jailbreak" or manipulate large language models (LLMs), such as GPT-3.5 and GPT-4. A taxonomy of 40 different persuasion techniques was developed and applied iteratively to these models. The team found a remarkably high success rate of 92% in jailbreaking these LLMs using what they term "Persuasive Adversarial Prompts" (PAPs).

Interestingly, the study revealed that more advanced models like GPT-4 are more susceptible to these PAPs. Furthermore, the adaptive defenses developed to counteract these PAPs also proved effective against other forms of attacks, suggesting a broader application for these defenses.

The message also demonstrates the use of persuasion techniques within its own text, such as "logical appeal" and "door in the face", providing real-time examples of the concepts discussed. The purpose of the project seems to be to both test and improve the safety and alignment of LLMs against various forms of persuasive attacks.