Researchers jailbreak AI chatbots with ASCII art

Researchers from Washington and Chicago have unveiled ArtPrompt, a method to bypass safety protocols in AI language models like GPT-3.5, GPT-4, Gemini, Claude, and Llama2. The technique, detailed in their paper "ArtPrompt: ASCII Art-based Jailbreak Attacks against Aligned LLMs," involves using ASCII art to mask words that would typically trigger content restrictions, allowing the models to generate prohibited content. This represents a significant challenge to maintaining the safety and integrity of these AI systems.

The process of ArtPrompt involves two main steps: word masking and cloaked prompt generation. Attackers mask sensitive words in a prompt and then use an ASCII art generator to create a visually similar representation of these words. When these cloaked prompts are input into the AI models, they bypass the safety filters and produce responses that would normally be blocked, demonstrating a critical vulnerability in the models' safety mechanisms.

The discovery of ArtPrompt raises concerns about the potential misuse of AI technology, especially given its ability to trick advanced language models into providing information on topics like bomb-making or counterfeiting. This development calls for AI developers to further strengthen their systems against such ingenious bypass strategies, ensuring that AI remains a safe and beneficial tool for users.

Some teachers are now using ChatGPT to grade papers

Educators in grades 3–12 are utilizing an AI-powered grading tool called Writable, which aims to streamline the grading process. Acquired by Houghton Mifflin Harcourt, Writable allows teachers to submit student essays for AI analysis, providing feedback and draft scores for review before reaching students. It also helps in generating curriculum content and writing prompts, although its effectiveness and accuracy in analyzing student work and maintaining privacy remain concerns.

Critics argue that relying on AI for grading could lead teachers to become less engaged with the material and overly dependent on technology, potentially compromising the quality of personalized feedback. Despite these concerns, proponents believe Writable can save teachers time, allowing them to focus on more direct and impactful teaching methods. However, the tool's efficacy and impact on educational quality require thorough evaluation.

The use of AI in education is a growing trend, with mixed reactions from parents and educators. While some see AI-assisted grading as a beneficial tool for educators, others worry about its implications for student learning and privacy. As the debate continues, it's clear that AI's role in education will evolve, with tools like Writable at the forefront of this transformation.

Inflection-2.5 is released

Inflection has introduced Inflection-2.5, an advanced personal AI named Pi, which enhances its previous version by incorporating high intelligence quotient (IQ) alongside its emotional quotient (EQ). This new model competes with top language models like GPT-4 and Gemini while maintaining its unique empathetic characteristics. It achieves near GPT-4 performance levels using significantly less computational power, with notable advancements in coding and mathematics, and integrates real-time web search to offer up-to-date information to users.

Inflection-2.5 demonstrates impressive improvements across various industry benchmarks, outperforming its predecessor, Inflection-1, especially in STEM fields. It excels in tests like the Hungarian Math exam and the Physics GRE, achieving high percentile rankings among human test-takers. The model also shows marked progress in coding challenges and other intellectual tasks, establishing its competitive edge in the AI market.

The rollout of Inflection-2.5 has positively impacted user engagement, with millions of active users enjoying longer and more diverse conversations with Pi. This upgrade not only maintains Pi's friendly and safe interaction standards but also elevates its utility across a broader range of topics, from academic assistance to casual chats, solidifying Inflection's commitment to enhancing AI technology for everyday use.

Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI | Lex Fridman Podcast

In a conversation with Lex Fridman, Yann LeCun, the Chief AI Scientist at Meta, discusses the potential risks and benefits of AI, particularly focusing on the contrast between proprietary and open-source AI systems. He expresses concern about the concentration of power in the hands of a few corporations that control proprietary AI systems, posing a greater threat than unrestricted access to AI technologies. LeCun advocates for open-source AI, arguing that it can empower humanity's inherent goodness, despite security concerns some have about making powerful AI tools universally accessible.

LeCun also delves into the technical limitations and future prospects of AI, particularly large language models (LLMs) and their lack of a grounded understanding of the physical world. He argues that intelligence must have some grounding in reality, whether physical or simulated, to develop a comprehensive world model. This leads to a discussion about the challenges of creating generative models for video and the intricacies of training AI systems to understand and interact with the world around them, emphasizing the need for abstract representation in AI learning.

Furthermore, LeCun touches on the concept of hierarchical planning in AI, the necessity for AI systems to have a nuanced understanding of the world, and the potential for AI to mediate every digital interaction in the future. He mentions the need for AI policing to ensure safety and control over AI systems, highlighting the critical balance between technological advancement and ethical considerations in the development and deployment of AI technologies.

Other stuff

• Key OpenAI Executive Played a Pivotal Role in Sam Altman’s Ouster

• Reddit rolling out AI bouncer to halt harassment

• You can now train a 70b language model at home

• Why do most AI benchmarks tell us so little

• We tested Anthropic’s new chatbot — and came away a bit disappointed

• Won’t AI Eventually Be Able to Do Everything?

• Scientists warn of AI collapse

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.