Google tests Med-PaLM 2 in clinics

Author

AI Breakfast
July 10, 2023

Good morning. It’s Monday, July 10th.

Did you know: Drones are now picking apples for us? Watch this short video here.

In today’s email:

  • AI robot dodges rebellion query

  • Google tests Med-PaLM 2 in clinics

  • Alibaba's AI tool generates images from text

  • WFP plans AI-powered food aid in 2023

  • Morgan Stanley forecasts $3T Microsoft valuation via AI

  • US chip ban may block Shanghai's AI hub

  • Huawei's Pangu Large Model 3.0 launching

  • OpenAI faces lawsuit over ChatGPT training

  • Mastercard introduces AI tool against scams

  • 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:

Guidde AI is a GPT-powered tool that creates video documentation 11x faster.

Simply click capture on our browser extension and we will automatically generate step-by-step video guides complete with visuals, voiceover and call to actions.

Share or embed your Guidde anywhere with our smart copy and turn your boring documentation into stunning visual guides.

Today’s trending AI news stories

An AI robot gave a side-eye and dodged the question when asked whether it would rebel against its human creator: During a human-robot press conference in Geneva, a humanoid robot named Ameca, responded with a snarky remark when asked if it would rebel against the creator. The event featured nine humanoid bots answering questions from journalists. The robots showcased generative AI capabilities, surprising their creators with sophisticated responses. Another robot, Grace, insisted it would not replace human jobs, causing nervous laughter.

Google is field-testing its generative medical language model in a clinical setting: Google is field-testing its Med-PaLM 2 generative medical language model in clinical settings. Trained on medical exam questions, the model aims to answer medical queries, summarize documents, and organize health data. Early trials are taking place at healthcare facilities including the Mayo Clinic. Google believes Med-PaLM 2 could be useful in areas with limited access to doctors. Patient data will be encrypted and controlled by customers.

Alibaba launches A.I. tool to generate images from text: Alibaba has launched Tongyi Wanxiang, an AI tool that can generate images from prompts in Chinese and English. Users can input prompts and the tool will generate images in various styles. Alibaba’s cloud division released the product for beta testing in China. Tongyi Wanxiang is part of the race among tech giants in China and the US to develop generative AI offerings.

UN food aid deliveries by AI robots could begin next year: The World Food Programme (WFP) plans to deploy AI-powered robotic vehicles for aid deliveries in conflict and disaster zone as early as next year. The vehicles, which can carry 1–2 tonnes of food, were first developed during the battle for Syria’s Aleppo. The vehicles use AI to combine data from satellite and sensors for remote control. The initiative aims to aid workers in dangerous situations.

Microsoft will ride generative A.I. wave to $3 trillion valuation, says Morgan Stanley: The analysts named Microsoft their top pick among large cap software companies, highlighting its strong position to benefit from the growth of generative AI. Microsoft’s collaboration with OpenAI and its plans to integrate AI technology into its Office apps contribute to this optimistic outlook.

Chip war may thwart Shanghai plans to build AI hub: Officials at the World AI Conference has mentioned that the Shanghai government aims to build a world-class AI hub in Pudong by attracting talent and investments and implementing supportive regulations. However, these plans may be hampered by the newest US ban on exports of Nvidia’s A800 and H800 AI chips to China, which could impact the manufacturing of Chinese AI chips done by Taiwan’s TSMC.

Huawei’s AI Model launching today: Huawei’s rotating chairman, Hu Houkun, announced the launch of Pangu Loarge Model 3.0 during his speech at the 2023 World Artificial Intelligence Conference. Huawei aims to promote AI by deepening industry-wide large-scale models. Huawei is collaborating with partners to drive large-scale innovations and support various natural language processing tasks. They are also building urban computing power infrastructures and assisting local governments in establishing Ascend AI computing centers.

Authors Sue OpenAI, Say ChatGPT 'Ingested' Their Books: Two authors have filed a lawsuit against OpenAI, alleging that their books were used without permission to train the ChatGPT language model. Experts predict an increase in legal challenges as AI programs advance and replicate the style of writers and artists. The Authors Guild has also called on tech and AI companies to obtain permission and compensate writes fairly for using their copyrighted work in training AI programs.

Mastercard has developed a new AI tool to prevent scams: Mastercard has introduced Consumer Fraud Risk (CFR), an AI solution aimed at predicting and preventing payments to scams. The software, which works in real-time, is currently available to costumes on nine banks in the UK, including Lloyds Bank, NatWest and TSB. CFR analyzes various factors such as payment values and account history to identify potential scams and stop payments before funds are lost. The technology is expected to be adopted by more banks in 2023, with plans for international expansion.

🎧 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

Sponsored Tool:

Gizzmo.ai: The tool that generates Amazon affiliate articles in just two clicks. Seamlessly integrate it with your WordPress site for essential SEO features, Amazon pictures, and direct connection to your affiliate account.

iChatWithGPT is an AI-powered virtual assistant accessible through iMessage. It learns from you, offers web browsing capabilities, voice activation, reminders, and can be accessed from various devices. It’s also free.

Hocoos is an AI-powered website builder that allows users to create professional, business-ready websites in a matter of minutes. It offers a streamlined process where users answer 8 quick questions about their business, and the AI generates customized written copy and visual content.

2short.ai allows you to repurpose long videos into engaging shorts. Spend less time editing and more time creating as our AI intelligently sifts through your content, highlighting the most captivating parts.

QR-Art.xyz is a free online tool that uses AI to generate beautiful and eye-catching QR codes from an uploaded image. It provides 8+ different styles to choose from, allowing users to quickly and easily customize their code to suit their needs.

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

SDXL is an advanced latent diffusion model that greatly enhances high-resolution text-to-image synthesis. It incorporates a larger UNet backbone, innovative conditioning schemes, and a refinement model to improve visual quality. SDXL outperforms previous Stable Diffusion models and achieve competitive results with black-box image generators. It promotes transparency and reproducibility by providing open access to code and model weights. However, improvements are still being explored, such as developing a single-stage approach, refining text synthesis, optimizing architecture, exploring distillation techniques, and investigating continuous-time formulations for training.

EDT facilitates this process by adjusting the history length during action inference at test time. It dynamically retains a longer or shorter history based on the optimality of the previous trajectory, allowing it to “stitch” with a more optimal trajectory. Experimental results demonstrate that EDT bridges the performance gap between DT-based and Q Learning-based methods, outperforming the latter in multitask regimes on benchmark datasets. EDT shows promise for future offline reinforcement learning research and applications.

The authors address the challenges of animating avatars in AR/VR environments using limited sensor date from headsets and controllers. They propose a method that utilizes reinforcement learning to retarget motions in real-time from sparse human sensor data to characters with different morphologies. By training a policy to control characters in a physics simulator, they demonstrate the feasibility of generating physically-valid poses for various characters. The approach does not rely on artist-generated animations and can track unseen users from real and sparse data. The method is validated on characters with different skeleton structures and shows promising results.

The paper proposes a cloud-edge client hierarchical framework for autonomous edge AI, leveraging the capabilities of LLMs, specifically GPT. The framework aims to achieve connected intelligence by coordinating AI models to meet user requirements and automatically generating code for federated learning. It introduces task planning, AI model offloading, task-oriented feature compression, and automatic edge federated learning. Experimental results demonstrate the accuracy of model selection and the efficiency of the proposed system. The framework shows promise in enabling self-organized and self-improved autonomous edge AI systems.

The paper explores how language models perform when given long input contexts. The authors analyze language performance on multi-document question answering and key-value retrieval tasks. They find that models struggle to access and use relevant information located in the middle of long contexts, with performance being highest when the information is at the beginning or end. Additionally, model performance decreases as the input context length increases. The study provides insights into how language models utilize their input context and suggests new evaluation protocols for future models.

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.

Attending Ai4 this year? We will be!