Here's Claude Pro, ChatGPT's Competitor

and other top AI News stories for the day

Author

AI Breakfast
September 08, 2023

Good morning. It’s Friday, September 8th.

Did you know: You can pre-load prompts into ChatGPT instead of re-writing the same ones over and over? Check out how to do this on OpenAI’s Blog.

In today’s email:

  • AI in Drug Discovery & Healthcare

  • Chatbots & Conversational AI [Claude Pro]

  • Legal & Ethical Implications of AI

  • AI Features in Productivity & Collaboration Tools

  • AI Fairness & Bias

  • AI Platforms & Services Growth

  • AI in Defense & Geopolitics

  • Tech Giants & AI Investments

  • Advanced AI Models & Research

  • AI in Design & Creativity Tools

  • 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 trending AI news stories

AI in Drug Discovery & Healthcare

ChatGPT pioneer launches ‘biological software’ company Inceptive, a startup founded by Jakob Uszkoreit, one of the co-inventors of ChatGPT technology, has secured £80 million ($100 million) in funding from investors, including Nvidia, to pioneer the use of generative AI in drug discovery. Uszkoreit's company aims to develop “biological software” for inventing new medicines. Inceptive new medicines. Inceptive utilizes AI to design novel molecules and collaborates with major pharmaceutical firms for testing. The investment underscores the growing interest in AI-driven drug development, although challenges such as lengthy laboratory testing and clinical trials persist.

Chatbots & Conversational AI

[Claude Pro]

Anthropic launches a paid plan for its AI-powered chatbot: Anthropic, a startup co-founded by ex-OpenAI employees, has introduced its first premium subscription plan, Claude Pro, for its AI-powered chatbot, Claude 2. Priced at $20 per month in the U.S. and £18 in the UK, subscribers gain “5x more usage” than the free tier, additional message capabilities, priority access during peak times, and early feature access. Anthropic positions Claude Pro in competition with OpenAI’s ChatGPT Plus.

Legal & Ethical Implications of AI

Microsoft to defend customers on AI copyright challenges: Microsoft has announced that it will cover legal damages on behalf of its customers who use its AI products if they are sued for copyright infringement related to the content generated by these systems. The company will take responsibility for potential legal risks, provided customers use the content filters and guardrails built into Microsoft’s products, which are designed to reduce the likelihood of AI-generated content infringing copyrights. Microsoft has been heavily investing in AI, particularly generative AI, and is expanding its use in various products and services, search, and enterprise productivity software.

Google sets new transparency rule for AI content in election ads: Google is introducing a new transparency rule for AI-generated content in election ads. Starting in mid-November, political advertisers will need to prominently disclose whether and how they are using “synthetic content that inauthentically depicts real or realistic-looking people or events” in their ads. This rule applies to images, videos, and audio files but exempts minor changes like image resizing or color correction. The move is aimed at addressing concerns about deep fake content in political advertising and ensuring greater transparency in the use of AI-generated content in politics.

The AI Drake ‘Ghostwriter’ is back with a new song and is chasing a Grammy: The AI creator known as Ghostwriter, famous for generating an AI-produced Drake song, has unveiled a new track featuring AI-generated voices of Travis Scott and 21 Savage. Title “Whiplash,” the song was shared on TikTok with Ghostwriter expressing interest in the future of music powered by AI. Ghostwriter’s AI-generated Drake song raised copyright concerns and was removed from streaming platforms. Interestingly, the creator is also submitting “Heart on My Sleeve,” for a Grammy, despite copyright challenges, aiming to qualify through creative criteria.

AI Features in Productivity & Collaboration Tools

Slack’s AI tool that can recap channels and threads starts testing this winter: Slack is set to launch Slack AI, an AI tool that can summarize channels and threads, and search for answers within messages, files, and channels. Starting testing this winter, it aims to help users sift through discussions, generate summaries, and find relevant information more efficiently. Slack AI also offers channel recaps and the ability to track projects across teams, taking on planning apps like Asana and Airtable. Slack continues to integrate AI features into its platform, enhancing productivity and collaboration.

AI Fairness & Bias

Meta releases FACET dataset for evaluating AI fairness: Meta Platforms Inc. has released the FACET dataset, designed to assist researchers in evaluating AI fairness in computer vision models. This dataset aims to make it easier for researchers to determine if a computer vision model exhibits bias by providing a substantial evaluation dataset containing images with labeled demographic attributes. Researchers can use FACET to check for fairness issues across four types of computer vision models, including those optimized for classification, object detection, instance segmentation, and visual grounding.

AI Platforms & Services Growth

ChatGPT is back on track for growth in the USA: ChatGPT’s global traffic has stabilized after experiencing significant declines during the summer months, with a slight increase observed in the US in August. This drop in usage during the summer is attributed to lower engagement by students. ChatGPT remains one of the largest websites globally, with over 1.4 billion visits. The platform’s popularity among students is expected to rise as they return for learning and writing purposes in September. OpenAI recently launched ChatGPT Enterprise, a business version not included in tracking, which may impact future reports.

OpenAI will host its first developer conference on November 6: The one-day event will feature keynote speeches and breakout sessions by OpenAI’s technical team, where they plan to unveil “new tools and exchange ideas.” While details about specific announcements are scarce, it’s an opportunity to gain insights into OPenAI’s plans for its AI models, including GPT-4, image understanding capabilities, and potentially addressing concerns like AI-generated content misappropriation. The conference will have both in-person and online components, catering to its developer community of over 2 million users. OpenAI seeks to expand its reach and explore commercial opportunities.

AI in Defense & Geopolitics

Pentagon Plans Vast AI Fleet to Counter China Threat: The Pentagon is reportedly planning to establish an extensive network of AI-powered technology, drones, and autonomous systems in the next two years to counter potential threats from China, according to the Wall Street Journal. This move underscores the growing importance of AI in defense strategies and the ongoing technological competition between nations.

Tech Giants & AI Investments

Apple is reportedly spending ‘millions of dollars a day’ training AI: Apple is reportedly making substantial daily investments in AI, developing multiple AI models across different teams. The “Foundational Models” unit focuses on conversational AI, while another unit, Visual Intelligence, is working on image generation models. Additionally, Apple is researching multimodal AI capable of recognizing and generating text, images, or video. One notable model, Ajax GPT, trained on over 200 billion parameters, is considered more potent than OpenAI’s GPT3.5. These AI efforts may enhance Siri and introduce new features like chatbots for customer interactions.

Tencent releases AI model for businesses as competition in China heats up: Tencent has been internally testing its AI model, Hunyuan on advertising and fintech and plans to integrate capabilities with existing products for video conferencing and social media. The release follows new Chinese regulations on generative AI, which has also led to other companies, such as Baidu, launching AI-powered chatbots in China. However, Chinese firms face U.S. restrictions on obtaining advanced semiconductors for AI development. Tencent’s CEO Dowson Tong emphasizes the need for guardrails to ensure AI quality and accuracy.

Advanced AI Models & Research

Spread Your Wings: Falcon 180B is here: Hugging Face introduces TII's Falcon 180B, the largest publicly available language model with 180 billion parameters. Trained on 3.5 trillion tokens using TII's RefinedWeb dataset, it achieves state-of-the-art results across various natural language tasks, rivaling proprietary models. The blog posts detail Falcon 180B's architecture, capabilities, hardware requirements, and prompt format. It’s available in the Hugging Face ecosystem, offering tools, scripts, and examples for utilization. Commercial usage is allowed with certain restrictions, and users are required to adhere to the license and terms of use.

AI in Design & Creativity Tools

Adobe Substance 3D Sampler 4.2 introduces a new AI-powered version of Image to Material and a new AI Upscale feature: This version provides complete control of the resolution per layer, enhancing material types, such as fabric and wood. The AI Upscale feature increases the quality and detail of low-resolution textures while maintaining feature coherency between maps. Additionally, the release includes various improvements, such as enhanced resolution dropdowns, new icons, and scripting support.

5 new AI-powered tools from around the web

Dokkio Organizes and manages your digital clutter effortlessly. Dokkio brings together data from cloud storage, emails, messages, desktop folders, and the web, using AI to tag and organize content to match your workflow and gain control over digital chaos.

SuperQuiz empowers educators, companies, and families to effortlessly create custom multiplayer quizzes with the help of AI, revolutionizing learning and engagement.

Charmed AI Texture Generator accelerates 3D game texture creation. Offers rapid design exploration and artist control for creating assets, ideal for seasonal events and themed packs. Designed to simplify the asset creation process for game developers and artists, fostering consistent styles across game meshes.

Jupitrr is an AI video maker that auto-generates B-rolls, including stock footage, charts, and memes, in a few clicks—empowering creators to save time on video editing and focus on content creation.

JoyAI for Salesforce connects businesses with generative AI, offering AI assistants, AI platform, Apps, Co-Pilots, and Plugins to augment processes and gain a competitive edge.

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

ImageBind-LLM introduces an innovative approach to LLMs by enabling multi-modality instruction tuning via ImageBind. Unlike previous models that focused on language and image instruction-tuning, ImageBind-LLM responds to diverse modalities like audio, 3D point clouds, video, and their embedding-space arithmetic through image-text alignment training. It utilizes a learnable bind network to align the embedding space between LLaMA and ImageBind’s image encoder, progressively injecting visual instructions via an attention-free and zero-initialized gating mechanism. In addition, a training-free visual cache model enhances cross-modal embedding during inference. This approach allows ImageBind-LLM to respond effectively to multi-modal instructions with high-language generation quality.

The Contrastive Feature Masking Vision Transformer (CFM-ViT) is a novel image-text pre-training method for open-vocabulary object detection. It combines masked autoencoder (MAE) objectives with contrastive learning to improve representation for localization tasks. Unlike traditional MAE, CFM ViT, performs reconstruction in the joint image-text embedding space, enhancing region-level semantics learning. Positional scale variation, improving detection performance and enabling the use of frozen ViT backbone as a region classifier. CFM-ViT achieves state-of-the-art results on open-vocabulary detection benchmarks, outperforming previous approaches. This work introduces a promising approach to open-vocabulary object detection with vision transformers.

The paper presents a solution to the visual disambiguation problem, distinguishing whether visually similar images represent the same or distinct 3D surfaces. These ambiguities, referred to as “doppelgangers,” challenge human perception and 3D reconstruction algorithms. The authors introduce the Doppelgangers Dataset, containing labeled image pairs, and propose a neural network architecture to address this challenge. Their method effectively distinguishes illusory matches, improving 3D reconstructions. The dataset is collected from Wikimedia Commons, and image pairs are labeled based on both geometric and visual cues. Data augmentation includes flipping images to create negative pairs for training.

In this paper, the authors explore the Feed Forward Network (FFN) component of the Transformer architecture and its redundancy. They find that sharing a single FFN across the encoder layers while eliminating the FFN in the decoder layers results in significant parameter savings without sacrificing accuracy. They also suggest using wider FFNs in the encoder, achieving improved accuracy and reduced latency compared to the original Transformer Big model. The study includes experiments on various language pairs, and the authors analyze representational similarity between different models. The results highlight opportunities for optimizing model architectures for practical deployment.

In this work, the authors introduce Optimizing by PROmpting (OPRO), a novel approach that leverages large language models (LLMs) as optimizers. OPRO utilizes natural language prompts to guide the optimization process. The authors demonstrate its effectiveness in solving optimization problems, including linear regression and the Traveling Salesman Problem (TSP). OPRO’s flexibility allows it to adapt to various tasks by changing the problem description in the prompt. It excels in prompt optimization, achieving up to 8% higher accuracy than human-designed prompts on GSM8K and up to 50% improvement on Big-Bench Hard tasks. OPRO’s potential is showcased through case studies and extensive evaluation using different LLMs.

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.