• AI Breakfast
  • Posts
  • AlphaFold AI Discovers Thousands of Possible Psychedelics

AlphaFold AI Discovers Thousands of Possible Psychedelics

Good morning. It’s Friday, January 19th.

Did you know: On this day in 1995, the yahoo.com domain was registered?

In today’s email:

  • AI in Science and Technology

  • Advances in AI Technologies & Models

  • AI in Industry and Commerce

  • AI in Healthcare and Wellness

  • AI and Ethics

  • AI in Media and Creative Industries

  • 6 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Today’s trending AI news stories

AI in Science and Technology

> DeepMind’s Alphafold has identified potential new psychedelic molecules, opening doors for antidepressant development. This breakthrough indicates that AlphaFold's rapid predictions can rival the accuracy and utility of conventional, more time-intensive methods. The pharmaceutical sector is taking keen interest in AlphaFold. Major companies are integrating it into their drug development pipelines, hoping it will expedite the discovery of novel therapeutics. AlphaFold is not expected to fully replace conventional methods, but is undoubtedly transforming the landscape of drug discovery, offering a faster and potentially more efficient route.

> Microsoft has launched Reading Coach, a free AI-powered tool designed to enhance reading practice. Accessible on the web and soon Windows app, it integrates with learning management systems like Canva. Building on Reading Progress from Teams for Education, Reading Coach offers personalized support, allowing educators to track progress and share insights. It features a "choose your own story" option, using AI to tailor narratives to the learner's needs, and provides interactive tools for vocabulary and pronunciation practice.

> BMW is set to deploy Figure's first humanoid robot at its South Carolina manufacturing plant, marking a significant step in integrating advanced robotics into automotive production. This "commercial agreement" brings cutting-edge humanoid automation to BMW's Spartanburg facility, showcasing a progressive move towards optimizing manufacturing processes and setting a new benchmark for the industry's adoption of robotic technology.

Advances in AI Technologies & Models

> Meta is advancing towards Artificial General Intelligence (AGI), developing Llama 3 for enhanced code generation and planning and is set to challenge the sophistication of GPT-4. Meta, while championing the principles of open-source, navigates this terrain with a nuanced balance of pioneering innovation and stringent ethical standards. Its AI teams merge to realize the AGI vision, amassing 600,000 GPUs for computational power.

> Arizona State University teams up with OpenAI, integrating ChatGPT into their curriculum. This groundbreaking partnership focuses on enhancing student success, driving innovative research, and optimizing organizational processes. ASU’s ambitious plans include creating personalized AI tutors for subjects like STEM and offering creative AI study aids. This initiative reflects a growing comfort with AI in education, positioning OpenAI as a key tech provider across various sectors.

> Google DeepMind’s new AI, AlphaGeometry, demonstrates a leap in AI’s reasoning capabilities by solving complex geometry problems, matching top high school mathematicians. This hybrid AI combines a language model with a symbolic engine for logical deductions, mimicking human problem-solving. AlphaGeometry’s success in the International Mathematical Olympiad underscores its potential beyond mathematics, impacting fields like computer vision and theoretical physics.

> Anthropic, an AI startup counters a copyright lawsuit from music publishers, arguing its use of lyrics to train AI models like Claude is transformative and constitutes a small fraction of its data. It challenges claims of direct infringement and irreparable harm, asserting that any lyric outputs were unintentional and now rectified with safeguards.

> Apple’s AIM (autoregressive image models) ushers in a new era in vision model training, drawing inspiration from the principles of large language models. This autoregressive image model suite showcases the potential of scalable, unsupervised pre-training, coupled with key technical advancements tailored for downstream applications. A standout achievement is the 7 billion parameter AIM model, pre-trained on 2 billion images, reaching an impressive 84.0% on ImageNet-1k, all while displaying no performance plateau, indicating vast possibilities for future model enhancement and extended training periods.

AI in Industry and Commerce

> Tesla CEO Elon Musk has expressed a desire for increased voting control at the company, aiming for at least 25% before advancing Tesla's role in AI and robotics. Musk, holding about 13% of Tesla's shares, indicated he might develop AI and robotics projects outside Tesla if his voting control does not meet his threshold. Despite Tesla's primary revenue from automotive sales, Musk has been vocal about Tesla's potential in AI, including the development of its "Full Self-Driving" software and humanoid robots, as well as the Dojo supercomputer for AI model training.

AI in Healthcare and Wellness

> The FDA has authorized an AI-powered medical device by Miami's DermaSensor Inc., to assist doctors in detecting prevalent skin cancers. This handheld device, not intended for initial screening, employs AI-driven spectroscopy to analyze skin lesions' cellular and sub-surface traits. DermaSensor evaluates moles or lesions against a database of over 4,000 cases, providing a 'spectral similarity score' to aid physicians' assessments.

> Forta, an AI-enhanced autism therapy platform, secured $55M in Series A funding to expand its services, advance AI and LLM research, and refine clinical algorithms. The San Francisco-based company empowers families with a 50-hour training course, enabling parents to deliver personalized ABA therapy for autism at home. Addressing care shortages, Forta aims to extend its caregiver model to memory care and other chronic conditions, amidst a growing market for digital autism care solutions.

AI and Ethics

> OpenAI CEO Sam Altman revealed that future AI models, including the anticipated GPT-5, will enhance its current capabilities noting that it necessitates “uncomfortable” decisions regarding individual customization and the handling of diverse cultural values. This adaptation may lead to AI providing different responses based on individual preferences or regional norms, posing ethical and practical challenges.

> Meta, under Zuckerberg’s direction, is massively investing in NVIDIA H100 Tensor Core GPUs, vital for AI research, signaling its commitment to Artificial General Intelligence. By 2024, Meta’s infrastructure will boast 350,000 H100 cards, with spending nearing $9 billion. This strategic deployment of advanced computing resources underscores Meta’s ambitious stance in the rapidly evolving AI domain, focusing on AGI initiatives and the development of the Llama 3 language model.

AI in Media and Creative Industries

> Rabbit partners with Perplexity, integrating its AI ‘answer engine’ into the Rabbit R1, a $199 AI gadget designed by Teenage Engineering. The R1 offers live, up-to-date information, breaking the knowledge cutoff barrier of traditional LLMs. Early buyers receive a year of Perplexity Pro, featuring access to advanced LLMs like GPT-4. The R1, with a touchscreen and camera, acts as a universal controller, while Perplexity AI blends search with LLMs, challenging giants like Google.

> RunwayML introduces Multi Motion Brush, enhancing AI video generation by allowing users to animate up to five separate objects within one image, each with independent movements. This advanced tool builds on the existing Motion Brush, offering unparalleled creative control in video creation and editing, and is available for Runway’s Gen-2 video model, marking a significant leap in AI-driven video manipulation technology.

6 new AI-powered tools from around the web

Sparksocial is an AI-driven platform changing lead generation through advanced social listening and targeted outreach.

riyo.ai is an AI-driven platform enhancing UX and conversions, providing marketing teams with tools like Dynamic Heatmaps, Session Recordings, and AI Chatbots for in-depth visitor engagement and insightful decision-making.

Helloii is a Chrome extension transforming your homepage into a ChatGPT interface, offering quick, insightful conversations and replacing traditional Google searches for smooth, AI-powered browsing experience.

Modelize.ai 1.0 streamlines AI workflows, enabling quick generation and customization for various projects. It’s an intuitive, comprehensive platform blending expert insights with AI capabilities, simplifying tasks and enhancing creativity.

iFoto.AI enhances e-commerce imagery with AI, simplifying product photography into a one-click process. It offers cost-effective background removal and high-definition, watermark-free downloads, with a generous limit of 10,000 images per user monthly.

Broadcast 2.0, tailored for managers, streamlines meetings by automating notes, syncing actions with project tools, and providing AI transparency. It offers decision tracking, integrates with workflows, and delivers weekly personalized reports, all while ensuring privacy.

arXiv is a free online library where researchers share pre-publication papers.

The Baidu team introduces UniVG, a versatile video generation framework that transcends traditional single-task models, adapting to diverse user inputs like text, images, or their combinations. This comprehensive approach categorizes video generation into high-freedom and low-freedom tasks, employing techniques like Multi-condition Cross Attention and Biased Gaussian Noise to ensure content alignment and preserve input details. Achieving impressive benchmarks in video generation, UniVG proves to be a promising solution for generating dynamic, contextually accurate videos, marking a significant stride in the field.

DeepSpeed-FastGen, developed by Microsoft, optimizes LLM serving with Dynamic SplitFuse, enhancing throughput and reducing latency in text generation tasks. This system integrates DeepSpeed-MII and DeepSpeed-Inference, offering a high-performance, user-friendly platform for diverse LLM applications. DeepSpeed-FastGen excels in performance metrics, supports various models, and provides flexible deployment options. The system's code is openly available for community use and contribution, indicating a commitment to collaborative improvement and expansion.

The Self-Rewarding Language Models (SRLM) study by Meta and NYU pioneers a new approach in AI learning, enabling models to generate and assess their own training data, breaking free from limitations set by human-level feedback. SRLMs utilize Iterative DPO training, where the model acts as both an instruction follower and a judge, improving its instruction following and reward modeling abilities with each iteration. The study demonstrated that the model not only adheres to instructions more effectively but also refines its ability to provide quality rewards to itself, indicating potential for continuous self-improvement in AI systems. The research opens new pathways for developing increasingly autonomous and self-enhancing AI agents.

The study introduces ChatQA, a new conversational QA model series achieving GPT-4 level accuracies through a unique two-stage instruction tuning and a dense retriever fine-tuning process. Unlike traditional methods, ChatQA utilizes context-enhanced instruction tuning to integrate user-provided or retrieved context for zero-shot conversational QA tasks, outperforming regular instruction tuning and RLHF-based recipes. Fine-tuning a dense retriever on a multi-turn QA dataset shows comparable results to state-of-the-art query rewriting models, significantly reducing deployment costs. ChatQA-70B, the leading model, surpasses GPT-4 in average scores across 10 conversational QA datasets without using synthetic data from OpenAI GPT models. Additionally, the inclusion of a small number of "unanswerable" samples in instruction tuning helps reduce hallucinations, making ChatQA a robust model for real-world applications.

SPARC, a novel method for multimodal pre-training, introduces a fine-grained contrastive alignment between image-text pairs, enhancing detailed information capture beyond traditional global alignment. Through a sparse similarity metric, SPARC groups relevant image patches to corresponding text tokens, forming language-grouped vision embeddings. These embeddings, contrasted with token embeddings, foster a detailed understanding of the image content. SPARC's innovative approach significantly outperforms existing models in various tasks, including classification, retrieval, object detection, and segmentation, without relying on extensive computational resources. The method's ability to scale to large batch sizes and avoid softmax's drawbacks further emphasizes its practicality and efficiency.

ChatGPT Creates Comics

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.