Tricking AI with Emotions: Jailbreaking Language Models by Humanizing ❤️‍🩹

PLUS: OpenAI Whisper 2x Faster, Chat with Mistral 7B on Your Phone

Shubham Saboo

Jan 10, 2024

Today’s top AI Highlights:

Speculative Decoding for 2x Faster Whisper Inference
AI Updates from AMD, Samsung, Volkswagen, and more from CES 2024
Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs
Chat with Mistral 7B on your phone without internet
More AI Tools for quick model deployment, building your own Slack AI, automated cap table management

& so much more!

Read time: 3 mins

Latest Developments 🌍

OpenAI’s Whisper with 2x Speed 🚝

OpenAI's Whisper TTS has demonstrated exceptional performance in speech transcription across multiple languages and challenging audio conditions. Despite its latest large-v3 model topping the OpenASR Leaderboard, Whisper faces a significant challenge in inference speed, taking over six minutes to transcribe a one-hour audio clip on a 16GB T4 GPU. Addressing this bottleneck, a new technique called Speculative Decoding has been introduced, increasing the transcription time by 2x without any loss in output quality.

Key Highlights:

This technique involves a two-step process where an assistant model first generates a sequence of candidate tokens, followed by their verification by the main Whisper model. This approach significantly accelerates the transcription process, effectively doubling the speed of the original Whisper model.
The Speculative Decoding technique maintains the same level of accuracy as the standalone Whisper model, achieving less than 30% word error rate (WER) on 42 out of 58 languages in the Common Voice 15 dataset. This balance between speed and accuracy is achieved by ensuring that the assistant model is at least 3x faster than the main model, while still capable of correctly predicting the majority of the tokens.
This update provides a seamless integration into existing Whisper pipelines, offering a straightforward upgrade for users without the need for major modifications. The technique is compatible with both English and multilingual variants of Whisper, making it a versatile tool for a wide range of transcription tasks.

Volkswagen Brings ChatGPT to its Cars 📺

CES 2024 witnessed a heap of updates from leading tech companies, and clearly, AI was at the forefront for all. Here are our favorite AI announcements from the event:

AMD has announced the Ryzen™ 8000G Series desktop processors with dedicated AI neural processing units (NPUs), a first for desktop PC processors. These NPUs are designed to accelerate AI software capabilities, optimizing AI workloads for enhanced productivity and efficiency.
Samsung’s new AI-powered SmartThings home platform includes an interactive "map view" of smart home devices that can be created using photos, and lidar-enabled devices, and displays real-time locations of devices and AI characters representing family members and pets. The features are accessible on select Samsung TVs, the Family Hub smart fridge, and M8 monitors, enhancing the smart home experience.
Samsung has reintroduced Ballie robot with several AI upgrades, with better navigation and lidar sensors, and a smart projector that adapts to room conditions and user positions. Its AI also allows voice and text control, and it learns user habits to personalize its functions and control various smart and traditional home devices.
Volkswagen is integrating ChatGPT into their vehicles, starting in Europe, to enable drivers to control various car functions and access information through natural language interaction. This feature, offered in models like the ID.7 and Tiguan, enhances the in-car experience by allowing hands-free conversations.
Getty Images and Nvidia have collaborated to create "Generative AI by iStock," an AI-powered text-to-image platform for stock photo generation, utilizing Nvidia's Picasso model and Getty’s huge library. This platform is tailored for small and medium businesses and also includes innovative features like Inpainting and Outpainting.

How Everyday Language Can Jailbreak LLMs 🔓

Recent studies have shifted the focus from traditional algorithm-focused attacks to exploring the risks posed by non-expert users in everyday interactions with LLMs. This paper delves into the concept of jailbreaking LLMs by treating them as human-like communicators, a largely unexplored intersection in AI safety. Utilizing a persuasion taxonomy derived from social science research, the study introduces the use of Persuasive Adversarial Prompts (PAP) to effectively jailbreak LLMs and the results are significant! This new perspective not only highlights the vulnerabilities of LLMs in everyday language use but also underscores the need for more comprehensive defense strategies for highly interactive AI systems.

Key Highlights:

The study introduces a novel approach using Persuasive Adversarial Prompts (PAPs) that achieved an astonishing 92% attack success rate in jailbreaking aligned LLMs, including models like Llama 2-7b Chat, GPT-3.5, and GPT-4. This was accomplished without relying on any specialized optimization techniques, emphasizing the effectiveness of the persuasion-based approach.
Intriguingly, the research revealed that more advanced models, particularly GPT-4, are more susceptible to these PAPs. Concurrently, adaptive defenses tailored to neutralize PAPs proved to be effective against a broader spectrum of attacks. However, even the most effective defense strategies could only reduce the attack success rate on GPT-4 to 60%, which is still higher than the success rate of the best baseline attacks.
The paper presents an extensive taxonomy of 40 persuasion techniques and examines their efficacy across OpenAI's 14 risk categories in GPT-3.5. This detailed analysis highlights the intricate relationship between different persuasion techniques and their impact on various risk categories, underscoring the need for ongoing research and development in AI safety and ethics.

Tools of the Trade ⚒️

Offline Chat: Use Mistral 7B 0.2 on your phone locally! The AI ChatBot is powered by the Mistral model that runs entirely on your device without the Internet. You can use it anywhere, and your data stays private and secure.

DeployFast: Offers a machine learning boilerplate for quick AI model deployment using FastAPI, Docker, and Streamlit, integrated with OpenAI and ElevenLabs APIs. It simplifies setup and deployment, making it ideal for AI novices and professionals alike.
Slack AI: Contains code on how to build your own Slack AI to chat with the unstructured data lying in your Slack channels. Set up your AI that runs locally, in just a few lines of code, and start chatting with your Slack data.
ByeByeCarta.com: An open-source cap table manager that streamlines equity management with real-time cap table updates and error-free equity tracking. Automated processes ensure accuracy and eliminate reconciliation worries. The platform facilitates seamless collaboration among stakeholders with shared access and transparent equity management.

😍 Enjoying so far, TWEET NOW to share with your friends!

Hot Takes 🔥

Heard someone say they are building a "LLM for law"?! I don't really get this? Is there a special architecture change that makes it better for "law"? Like that's not how it works imo ~ anton
Let’s hope 2024 is the year when elonmusk finally open sources an AI model. ~ Julien Chaumond

Meme of the Day 🤡

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!

Share Unwind AI

Unwind AI