Meta Releases ChatGPT Competitor Model

...and partnered with Microsoft to do it.

Author

AI Breakfast
July 19, 2023

Good morning. It’s Wednesday, July 19th.

Did you know: On this day in 2000, Apple announced the Power Mac G4 Cube.

In today’s email:

  • Meta, Microsoft launch Llama 2 open-source model

  • Microsoft introduces secure Bing Chat Enterprise

  • Stability AI delays Stable Diffusion XL v1.0 release

  • Meta, Qualcomm to integrate Llama 2 in devices from 2024

  • Futureverse raises $54M for AI metaverse venture

  • NY police use AI for successful drug bust

  • Microsoft announces Vector Search, voice cloning launch

  • Musk develops curious AI model, xAI

  • Wix unveils AI-powered Site Generator

  • WormGPT AI tool enables sophisticated cyber attacks

  • MIT's PIGINet boosts household robots' efficiency

  • Calls for ban on invasive automated legal enforcement

  • Bank of America uses AI, VR for training

  • Samsung tests ChatGPT in browser integration

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s edition is brought to you by:

Attending Ai4 this year? We will be!

Today’s trending AI news stories

Meta and Microsoft Reveal Llama 2 Open-Source Language Model: Meta and Microsoft have jointly launched Llama 2, the latest in open-source language model development. Available for free use in “research and commercial” applications, the powerful model has been touted to be an open-source contender to GPT-4. Microsoft is positioned as the preferred partner for Llama 2, which can be accessed through Microsoft Azure and other platforms like Amazon Web Services and Hugging Face. The partnership seeks to democratize AI access while fostering responsible development through transparency, safety testing, and shared resources. Read the full research paper here.

Microsoft Unveils Secure AI Chat Tool for Businesses, Bing Chat Enterprise: Microsoft has announced Bing Chat Enterprise, a secure business-orientated version of its AI chat tool. The new tool ensures users' chat data remains private and not used for AI training, mitigating potential data leak concerns.

Stability AI Delays Launch of Stable Diffusion XL v1.0: Stability AI has postponed the release of its anticipated upgrade, Stable Diffusion XL v1.0, by approximately a week to evaluate different fine-tuned models. The delay has sparked interest and speculation in the AI art community.

Meta and Qualcomm Collaborate to Run Large AI Models on Phones: Qualcomm and Meta are joining forces to run Meta’s Llama 2 on Qualcomm chips in phones and PCs starting 2024. The collaboration aims to integrate AI capabilities on devices, reducing reliance on large server farms and lowering AI execution costs.

Futureverse Raises $54 Million for AI Metaverse Venture: AI metaverse startup Futureverse, a merger of 11 companies, has secured $54 million in funding in a Series A round led by 10T Holdings and participated by Ripple Labs Inc.

AI Helps New York Police Apprehend Drug Trafficker: New York police utilized an AI-based "roadway intelligence" platform developed by Rekor to apprehend a drug trafficker based on analysis of driving patterns.

Microsoft Announces Vector Search Preview, Voice Cloning Launch: At its annual Inspire conference, Microsoft announced Vector Search, an AI feature in preview for Azure Cognitive Search, and the launch of voice cloning capabilities, signifying the company's dedication to AI leverage in document processing and speech services.

Musk’s xAI Trained to be 'Maximally Curious': Elon Musk is developing a new AI system, xAI, trained to be "maximally curious" to align with human values and manage potential risks. xAI is an Explainable AI model aimed to provide transparency in AI decision-making.

Wix Unveils AI Site Generator: Wix has introduced its AI Site Generator tool, enabling users to generate fully-fledged websites based on their prompts. The tool is a part of Wix’s wider AI initiative, streamlining website building and enhancing customer experience.

WormGPT: A New AI Tool Enabling Sophisticated Cyber Attacks: The emergence of WormGPT, a generative AI tool available on underground forums, is facilitating sophisticated phishing and business email compromise attacks by automating the creation of personalized, convincing fake emails.

MIT's AI System Enhances Household Robots' Problem-Solving Skills: MIT's AI system, PIGINet, enhances household robots' problem-solving capabilities by reducing planning time by 50-80%, thus revolutionizing how robots are trained and deployed.

AI Microdirectives May Be Used for Legal Enforcement: Legal experts suggest expanding surveillance tech bans to include invasive automated legal enforcement, warning of AI-driven threats to freedom and potential challenges in legal interpretations.

Bank of America Implements AI and VR for Employee Training: Bank of America is utilizing AI, VR, and the metaverse for immersive employee training, reflecting a trend in the banking industry towards technology-driven performance enhancement and cost reduction.

Samsung Tests ChatGPT Integration in Browser: Samsung is reportedly testing the integration of ChatGPT in its Internet Browser app, indicating its interest in extending AI capabilities to customers, although the release details have yet to be confirmed.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Featured Tool:

ChatDOC: Chat with documents, get instant answers with cited sources. A sophisticated document processing tool that employs a PDF parser to extract data with high accuracy from tables and text. It allows users to ask targeted questions by selecting specific content. Compatibility extends to scanned files, with OCR for English and Chinese languages, as well as .docx and .doc formats. Every response can be accurately traced back to its original fragment, ensuring precise traceability.

BrainstormGPT utilizes multi-agent LLM, and auto-search to convert topics into meeting reports seamlessly. Custom topics, user-defined roles, autonomous agent discussion, and rapid report generation make it a powerful tool for efficient collaboration and analysis.

Snov.io Email AI assistant is a powerful tool that helps you generate effective and compelling reply boosting emails in a matter of seconds. With a unique set of prompt enhancers developed by their team, even users with no copywriting skills can create high-quality emails easily.

CoPrompt is a productivity SaaS platform that allows teams to collaborate using ChatGPT. With its simple and flexible interface, teams can iterate prompts, achieve better results, and learn from each other in a multiplayer ChatGPT jamming session.

Permar AI is an innovative tool that enables you to create conversion rate optimized landing pages in seconds. By simply uploading a text prompt, Permar AI generates a fully functional landing page for you.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

The study examines the faithfulness of the stated reasoning in LLMs when generating step-by-step “Chain-of-Thought” (CoT) explanations. The research investigates possible reasons for unfaithfulness, such as post-hoc reasoning, test-time computation, and encoded reasoning. Experimental tests are conducted to measure the impact of truncating the CoT, adding mistakes, using filler tokens, and paraphrasing. The findings indicate that while CoT reasoning is not always faithful, it is possible to find conditions where it is more faithful, suggesting avenues for designing LLMs to produce more reliable reasoning.

In a study conducted by researchers from Stanford University and UC Berkeley, the behavior of ChatGPT’s 3.5 and GPT-4 models over time was evaluated. The study compared the March 2023 and June 2023 versions of both models on tasks such as math problem-solving, answering sensitive questions, code generation, and visual reasoning. The findings revealed significant variations in the performance and behavior of the models over time. For example, GPT-4’s accuracy in identifying prime numbers dropped from 97.6% to 2.4% between March and June 2023. The study emphasizes the need for continuous monitoring of large language models to ensure their quality and reliability.

In this research paper, authors from Meta AI, Microsoft Research, and Samsung AI Center, University of Maryland, and Cambridge University propose a sample-efficient method called SDS-CLIP. This method is intended to enhance the compositional visio-linguistic reasoning capabilities of CLIP. By leveraging distillation objectives from text-to-image generative models like Stable-Diffusion, they fine-tune CLIP with differentiable image parameterizations. The method demonstrates improved performance on challenging visio-linguistic benchmarks, such as Winoground, and also marginally enhances the zero-shot classification capabilities of CLIP. The study highlights the potential of leveraging generative models to extend contrastive image-text models for better visio-linguistic reasoning.

INVE (Interactive Neural Video Editing) is a real-time video editing solution that propagates sparse frame edits to the entire video clip. It improves upon Layered Neural Atlas (LNA) by addressing its limitations of slow editing speed and limited support for certain editing use cases. INVE achieves faster training and inference times, supports various video editing operations, and enables direct frame editing and rigid texture tracking. It utilizes efficient network architectures and bi-directional mapping between image-atlas, providing a more interactive and versatile video editing experience. INVE demonstrates superior performance over LNA through quantitative and qualitative analysis.

SEED, a project by Tencent AI Lab and ARC Lab, introduces a discrete image tokenizer that enhances LLMs with the capacity to comprehend and generate visual content. By focusing on high-level semantics and causal dependency, SEED facilitates seamless integration of visual and textual representations. The partnership advances the field of AI by bridging the gap between language and vision, empowering LLMs to comprehend and generate visual content with greater efficiency.

Thank you for reading today’s edition.

Your feedback is valuable.


Respond to this email and tell us how you think we could add more value to this newsletter.