Mistral CEO confirms ‘leak’ of new open source AI model nearing GPT-4 performance

The AI community has been buzzing with the release of "miqu-1-70b", an open-source large language model (LLM) introduced by "Miqu Dev" on HuggingFace, which has shown impressive performance closely matching or potentially surpassing OpenAI's GPT-4 in certain benchmarks. This model, speculated to be a quantized version of Mistral's work, based on Meta's Llama 2, sparked discussions across various platforms, from 4chan to LinkedIn, suggesting it could be an official leak or an independent release by an enthusiastic individual. Mistral's CEO, Arthur Mensch, confirmed that an early access customer leaked a quantized and watermarked version of an old model, hinting at more advanced versions in development. This development marks a significant moment in the open source AI landscape, indicating a rapid catch-up to proprietary models like GPT-4, potentially reshaping the competitive dynamics in AI technology and application.

OpenAI Says ChatGPT Probably Won’t Make a Bioweapon

OpenAI released a research result that outlines an evaluation framework developed to assess the risk of large language models (LLMs), like GPT-4, aiding in the creation of biological threats. This framework is part of a broader Preparedness Framework aimed at understanding AI-enabled safety risks and improving evaluation methods.

This research, prompted by concerns outlined in President Biden's AI Executive Order, aimed to assess if AI could lower the barrier for creating bioweapons. The study involved 100 participants, divided into groups with and without GPT-4 access, tasked with planning bioweapons.

The evaluation involved 100 participants, split evenly between biology experts and students, to see if access to GPT-4, in addition to the internet, would significantly increase their ability to create biological threats across five stages of threat creation. The study found only mild uplifts in accuracy and completeness for tasks related to biological threat creation, with no statistically significant effect sizes, suggesting GPT-4 provides at most a minor advantage over existing internet resources.

Results showed an 8.8% increase in accuracy for biology experts and a 2.5% increase for students using GPT-4. Despite slight improvements in plan accuracy and completeness, OpenAI emphasized that access to information alone doesn't significantly elevate the risk of actual bioweapon creation. The company calls for more research, noting that bioweapon information is already accessible online, implying that AI contributes little to the overall threat.

Google released MobileDiffusion: Rapid text-to-image generation on-device

"MobileDiffusion: Subsecond Text-to-Image Generation on Mobile Devices" introduces a novel, efficient latent diffusion model designed for rapid, on-device text-to-image generation on mobile devices. With only 520 million parameters, MobileDiffusion is capable of generating high-quality 512x512 images in less than half a second on both iOS and Android premium devices.

This breakthrough is achieved by optimizing the architecture for mobile deployment, focusing on the diffusion UNet's efficiency and a lightweight image decoder, and adopting a one-step sampling technique through DiffusionGAN. This approach addresses the challenges posed by the high computational demand and inefficiency of current text-to-image models, offering a solution that is both fast and resource-light for mobile use.

The model's design includes a compact text encoder, an efficient UNet architecture that incorporates transformer and lightweight convolution blocks, and a pruned decoder that improves latency and image quality. MobileDiffusion's performance indicates significant potential for enhancing mobile user experiences with rapid image generation capabilities, aligning with responsible AI practices.

Shopify’s ‘Magic’ AI image editor can make any product pics look professional

Shopify is enhancing its e-commerce platform with new AI tools, including a media editor capable of generating images and a conversational search feature, as part of its Shopify Magic suite. This update aims to assist merchants, especially those without access to professional photography or studios, in creating visually appealing product presentations by enabling easy background changes for product images.

The media editor offers a variety of styles, such as minimal, vibrant, and surreal, to fit different themes. Additionally, Shopify has integrated other AI-powered functionalities like automated product descriptions and chatbot services. These tools are designed to improve the online shopping experience by making product searches more intuitive and providing support for both merchants and customers.

Shopify's AI features are built on foundation models like GPT and Llama 2, ensuring they include necessary watermarking and moderation capabilities. This initiative reflects Shopify's commitment to lowering entry barriers for entrepreneurs and competing with other platforms offering AI-driven enhancements, such as Amazon.

Other stuff

• Sam Altman: A one-person company can hit a billion-dollar valuation soon

• How hard is it to cheat with ChatGPT in technical interviews? We ran an experiment.

• Give AI curiosity, and it will watch TV forever

• Amazon's AI model lets you virtually try on clothes and furniture

• It's true, LLMs are better than people – at creating convincing misinformation

• The promise and challenges of crypto + AI applications

Maximize Efficiency with Custom Prompts in SuperPower ChatGPT: A Comprehensive Guide

Hope you enjoyed today's newsletter

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