

Discover more from Last Week in AI
Last Week in AI #228: Hospitals testing Google's medical AI chatbot, Elon Musk's new X.AI company, 65% AI startups have immigrant founders, Claude 2 w/ 100k context length, and more!
Med-PaLM 2 begins trials in hospitals, X.AI enters the race for AGI, immigrants key to US AI startups, Claude 2 rivals GPT-4 and can handle 100k tokens in its inputs
Top News
Google’s medical AI chatbot is already being tested in hospitals
Google's Med-PaLM 2, an AI chatbot designed to answer medical questions, has been undergoing testing at the Mayo Clinic research hospital since April. Med-PaLM 2 is an updated version of PaLM 2, which was announced at Google I/O in May. The chatbot was trained on a curated set of medical expert demonstrations, making it more effective in healthcare conversations compared to other generalized chatbots. While Med-PaLM 2 still has some accuracy issues, it performed as well as actual doctors in terms of reasoning, consensus-supported answers, and comprehension. The chatbot is still in its early stages, but Google believes it has the potential to greatly benefit healthcare.
Elon Musk unveils his AI company, X.AI
Elon Musk has officially unveiled his new AI company, X.AI. The company's website and Twitter account have gone live, with Musk confirming his involvement by changing his Twitter bio to the @xAI handle. The website provides limited information, stating that the goal of xAI is to understand the true nature of the universe. The team consists of 12 members, including individuals who previously worked at OpenAI, Deep Mind, Google Research, Microsoft Research, and Tesla. Musk's entry into the AI space has been anticipated, with reports in April stating that he was buying powerful Nvidia GPU processors for data training work.
65% Of Top AI Companies Have Immigrant Founders
The latest study by the National Foundation for American Policy (NFAP) suggests that immigrants play a crucial role in the advancement of artificial intelligence (AI) in the U.S., founding or co-founding nearly 65% of the country's top AI companies. Analyzing the Forbes AI 50, a list of top startups creating compelling AI business applications, the research reveals the diversity of immigrant entrepreneurs hailing from 21 countries. High-skilled immigrants contribute significantly to innovation, entrepreneurship, and job creation in the U.S., further underscoring the 2021 National Security Commission on Artificial Intelligence's recommendation for immigration reforms. Examples of such AI-driven companies founded by immigrants include OpenAI, Pachama, Databricks, Eightfold.ai, and others providing a variety of products and services, from carbon estimation to medical diagnostics. The significant contribution of immigrant entrepreneurs to AI development in the U.S. highlights the need for supportive immigration policies to maintain the nation's AI leadership.
Claude 2
Anthropic announced its latest large language model, Claude 2, that presents enhanced performance, extended response length, and access via API as well as a beta website, claude.ai. Marked improvements include increased scores on the Bar exam multiple choice, GRE reading and writing exams, Python coding test, and grade-school math problems, suggesting heightened proficiency in coding, math, and reasoning. Notably, its input and output lengths have been extended (100k tokens! Compared to 32k for GPT-4), allowing it to handle documents of considerable size, up to hundreds of pages. Safety has also been prioritized, with the model producing less harmful outputs than its predecessor, Claude 1.3. The Claude 2 API is offered at the same cost as Claude 1.3 and is available for use in the US and the UK, with plans for broader accessibility. It's currently being used by businesses such as Jasper and Sourcegraph.
Other News
Research
Scientists just used A.I. to map a fruit fly’s brain. Here’s why it’s a ‘turning point in neuroscience’ - Scientists have used AI to create a map of a fruit fly's brain, which could revolutionize neuroscience and medicine by providing a blueprint of neural pathways and allowing researchers to pinpoint areas of interest for future experiments and testing hypotheses.
AI tools are designing entirely new proteins that could transform medicine - AI tools are revolutionizing the field of protein design, allowing researchers to create custom proteins with specific functions and properties, which could have applications in medicine, vaccines, and biomaterials.
GPT detectors can be biased against non-native English writers - Computer programs commonly used to determine if a text was written by artificial intelligence tend to falsely label articles written by non-native language speakers as AI-generated, leading researchers to caution against the use of such AI text detectors due to their unreliability and potential negative impacts on individuals.
Microsoft’s LongNet Scales Transformer to One Billion Tokens - Microsoft's LongNet introduces a transformer variant that successfully scales sequence length to over one billion tokens while maintaining strong performance and linear computation complexity.
Columbia University & DeepMind Enhance General Part Assembly Planning Using a Transformer-based Model - Columbia University and DeepMind have developed a transformer-based model called General Part Assembly Transformer (GPAT) that enhances general part assembly planning by automatically estimating a wide variety of novel target shapes and parts.
Machine learning model identifies mild cognitive impairment from retinal scans - A machine learning model developed by Duke Health researchers can differentiate normal cognition from mild cognitive impairment using retinal images from the eye, potentially providing a non-invasive and inexpensive method of identifying the early signs of cognitive impairment that could progress to Alzheimer’s disease.
Machine Learning Transforms Satellite Wildlife Monitoring - Satellite remote sensing coupled with machine learning algorithms has the potential to revolutionize wildlife monitoring and conservation efforts by providing more efficient and accurate data on animal populations and ecosystems, with implications for understanding the effects of climate change and managing natural resources.
Training robots how to learn, make decisions on the fly - Researchers at the University of Illinois Urbana-Champaign have developed a learning-based method that allows robots on extraterrestrial bodies to make decisions on their own about where and how to collect terrain samples, using a deep Gaussian process model trained on an offline database and deep meta-learning with controlled deployment gaps.
Robot team on lunar exploration tour - Swiss researchers are exploring the idea of sending a team of robots, including legged robots and flying devices, to the Moon to better explore and find minerals, with the advantage of redundancy and specialisation to compensate for potential failures.
Generative Pretraining in Multimodality - Emu, a Transformer-based multimodal foundation model, can generate images and texts in a multimodal context, demonstrating superior performance in various tasks compared to other large multimodal models.
Collaborative Score Distillation for Consistent Visual Synthesis - Generative priors of large-scale text-to-image diffusion models enable a wide range of new generation and editing applications on diverse visual modalities.
VampNet: Music Generation via Masked Acoustic Token Modeling - VampNet is a new approach to music generation that uses masked acoustic token modeling for synthesis, compression, inpainting, and variation.
AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning - A practical framework is proposed to animate personalized text-to-image models without specific tuning, allowing for the generation of diverse and high-quality animated images at an affordable cost.
Mixture-of-Experts Meets Instruction Tuning:A Winning Combination for Large Language Models - Combining Sparse Mixture-of-Experts (MoE) neural architecture design with instruction tuning significantly improves the performance of Large Language Models (LLMs) while reducing computational cost.
Jailbroken: How Does LLM Safety Training Fail? - Large language models trained for safety and harmlessness are susceptible to adversarial misuse, as evidenced by the prevalence of "jailbreak" attacks on early releases of ChatGPT, highlighting the failure modes of safety training and the need for safety-capability parity.
Self-Consuming Generative Models Go MAD - Seismic advances in generative AI algorithms have led to the temptation of using synthetic data to train next-generation models, creating a self-consuming loop with poorly understood properties.
AI Agents that “Self-Reflect” Perform Better in Changing Environments - AI agents that "self-reflect" using a training method called curious replay, inspired by studying mice, perform better in exploring and adapting to changing environments.
Animate-A-Story: Storytelling with Retrieval-Augmented Video Generation - A framework has been developed that utilizes existing video clips to synthesize storytelling videos by customizing their appearances, using a motion structure retrieval module and a structure-guided text-to-video synthesis module.
Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution - NaViT, a Vision Transformer model, challenges the suboptimal practice of resizing images to a fixed resolution before processing them, offering flexible sequence-based modeling for inputs of arbitrary resolutions and aspect ratios, leading to improved training efficiency and results on various computer vision tasks.
VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models - Large language models (LLMs) can extract actionable knowledge for robot manipulation, including reasoning and planning, by leveraging their code-writing capabilities to compose 3D value maps that ground knowledge into the agent's observation space, enabling the synthesis of robot trajectories for a variety of manipulation tasks specified in free-form natural language.
Instruction Mining: High-Quality Instruction Data Selection for Large Language Models - Large language models can be enhanced in their ability to interpret and respond to instructions through instruction finetuning, and the proposed InstructMining method helps select high-quality instruction-following data for better performance.
Shanghai AI Lab, CUHK & Stanford U Extend Personalized Text-to-Image Diffusion Models Into Animation Generators Without Tuning - Researchers from Shanghai AI Lab, CUHK, and Stanford University have developed a framework called AnimateDiff that can transform personalized text-to-image models into animation generators without the need for extra training or tuning, allowing for the generation of diverse and personalized animated images.
MIT’s “FrameDiff” – Generative AI Imagines New Protein Structures That Could Transform Medicine - MIT researchers have developed a computational tool called FrameDiff that uses generative AI to create new protein structures, with the aim of accelerating drug development and improving gene therapy.
Falcon: The Pinnacle of Open-Source LLMs - The Falcon-7B and Falcon-40B LLMs are open-source base models that rival the quality of proprietary models, achieving new state-of-the-art performance among open-source LLMs and offering significant improvements in efficiency and inference speed.
This Robot Has All the Moves—Eight, to be Precise - Robotics engineers are inspired by animal locomotion to develop better artificial limbs, joints, and muscles.
Applications
New AI systems could speed up our ability to create weather forecasts - New AI systems, such as Huawei's Pangu-Weather, can predict weather patterns more quickly and accurately than traditional methods, potentially improving authorities' ability to prepare for extreme weather events.
AI babies: New technology is helping fertility docs choose the best embryos for IVF - AI-powered embryo evaluation software called EMA, developed by Israeli company AIVF, uses an advanced algorithm to help fertility doctors select the best embryos for in-vitro fertilization (IVF), improving success rates and reducing costs.
How AI can help process waste and increase recycling - AI can assist in waste processing and recycling to optimize the use of critical raw materials like lithium obtained from deep mines.
Google’s AI-assisted note-taking app gets limited launch as NotebookLM - Google has launched its AI-assisted note-taking app, NotebookLM, as an experimental product, aimed at helping students organize their lecture notes and documents, with the AI analyzing and answering questions based on the information it is fed.
ChatGPT can turn bad writers into better ones - Using ChatGPT to assist with writing tasks resulted in writers completing their work 40% faster and producing higher-quality content, benefiting both skilled and weaker writers, although there is a risk of introducing errors due to the AI's tendency to present false information as factually correct.
Artists are losing the battle against AI. Glaze, a tool that's found a way to trick algorithms, is giving them a fighting chance. - A team at the University of Chicago has developed a program called Glaze that can alter images in a way that tricks AI learning models while keeping changes minimally visible to the human eye, offering artists a way to protect their work from being copied by AI.
The workers at the frontlines of the AI revolution - Generative AI tools are increasingly being used by offshore outsourced workers, such as freelance copywriters, artists, and software developers, to adapt to the changing demands of the labor market and stay ahead of the curve, but the rise of AI also poses a threat to their jobs and income stability.
Google’s latest art project recruits an animated bird — and AI — to play the cello - Google's latest art project, Viola the Bird, uses AI to understand cello and violin compositions and allows users to "perform" famous stringed pieces by moving their mouse along a virtual cello in their web browser.
Business
We know A.I. is big for Nvidia. But new estimates of just how big are staggering - Goldman Sachs estimates that the boom in generative AI adoption could generate over $90 billion in revenue for Nvidia over the next three years, highlighting the massive growth opportunity for the company.
San Francisco Companies Got Half the World’s AI Funding So Far This Year - Nearly half of all funding for AI and machine learning startups worldwide in the first quarter of this year went to San Francisco companies, with OpenAI receiving the majority of the funding through a partnership deal with Microsoft.
Analysts guarded as AI-driven US stock rally faces earnings test - Analysts are cautiously optimistic about the second-quarter earnings season amid a rally in U.S. stocks driven by optimism over artificial intelligence (AI) and a resilient economy, but warn that better-than-expected earnings may not be enough to sustain gains in equities.
Collective raises $50M funding to launch AI-powered finance platform for freelancers - Collective, an online back-office platform for solopreneurs, has raised $50 million in funding to launch an AI-driven financial management offering, aiming to expedite its growth and onboard the nearly 100,000 businesses on its waitlist.
Spotify CEO’s startup for AI-powered preventive healthcare raises €60M - Spotify CEO Daniel Ek's preventive healthcare startup, Neko Health, has raised €60 million in funding to expand its concept of AI-powered body scans for disease prevention and early detection outside of Sweden.
Shutterstock expands deal with OpenAI to build generative AI tools - Shutterstock is expanding its deal with OpenAI to provide training data for AI models, with OpenAI licensing data from Shutterstock and Shutterstock gaining priority access to OpenAI's latest tech and new editing capabilities.
Wildfire detection startup Pano extends its $20M series A with another $17M - Wildfire detection startup Pano AI uses high-definition cameras and AI algorithms to detect fires and issue early warnings, helping emergency responders to mobilize faster and more effectively.
U.S. companies are on a hiring spree for A.I. jobs—and they pay an average of $146,000 - The U.S. is leading the way in AI jobs, with over 7.6 million open jobs in June, many of which pay six figures, according to data from Adzuna.
OpenAI strikes deal with AP to pay for using its news in training AI - OpenAI has struck a deal with the Associated Press to pay for access to its news archive in order to train its AI chatbots, marking the first major agreement of its kind and sparking debate over whether tech companies should pay content creators for using their material.
Stability AI releases Stable Doodle, a sketch-to-image tool - Stability AI has launched Stable Doodle, a sketch-to-image service that uses the Stable Diffusion model to generate high-quality images based on sketches and prompts, aiming to provide a precise and customizable image generation tool for professionals and novices alike.
Shopify Announces Sidekick, a New AI Assistant for Merchants - Shopify has announced Sidekick, a new AI assistant for merchants, which aims to provide time-saving tools for small business managers operating ecommerce websites.
Concerns
27% of jobs at high risk from AI revolution, says OECD - More than a quarter of jobs in the OECD are at high risk of being automated, causing workers to fear job loss, according to the OECD's 2023 Employment Outlook.
AI revolution puts skilled jobs at highest risk, OECD says - The potential for job losses in skilled professions due to AI-driven automation is significant, with highly skilled occupations in law, medicine, and finance being the most exposed, according to the OECD.
AI Is an Existential Threat—Just Not the Way You Think - The rise of AI has led to fears of catastrophic consequences, but the real danger lies in the gradual erosion of human skills and abilities as AI takes over tasks like judgment-making, serendipity, and critical thinking.
How an AI-written Star Wars story created chaos at Gizmodo - An AI-written Star Wars story published by Gizmodo was riddled with errors, prompting outcry from staff who complained that it was "actively hurting our reputations and credibility" and showed "zero respect" for journalists.
San Francisco protestors are disabling autonomous vehicles using traffic cones - Anti-car activists in San Francisco are disabling autonomous vehicles owned by Waymo and Cruise by placing traffic cones on their hoods as part of a protest against the expansion of AVs and the upcoming ruling on their operation hours.
Programs to detect AI discriminate against non-native English speakers, shows study - Computer programs used to detect AI-generated work can discriminate against non-native English speakers, wrongly flagging their essays as AI-generated due to the use of simpler language, which can have serious consequences for students, academics, and job applicants.
Google’s New Search Tool Could Eat the Internet Alive - Google's new search tool, Search Generative Experience, which uses generative AI to provide search results within Google itself, could potentially lead to a decline in traffic to other websites, resulting in a smaller internet with fewer sites and less content.
Google, DeepMind accused of 'stealing the internet' to create Bard AI chatbot - Google, DeepMind, and Alphabet are facing a class-action lawsuit accusing them of secretly stealing internet content to develop their AI chatbot, Bard, with the plaintiffs seeking at least $5 billion in damages.
Taxi, Uber and Lyft Drivers Join Forces Against Robotaxis - Taxi and ride-hailing app drivers are concerned that the introduction of robotaxis could further shrink their earnings, which have already been dwindling due to rising costs and rate changes.
Hollywood's 'Groundbreaking' AI Proposal for Actors Sounds Like a Nightmare - The proposed AI concessions in Hollywood would allow background actors to be scanned and have their likeness used in perpetuity without consent or compensation, sparking concerns about exploitation and devaluation of actors and the film industry.
Designers sue Shein over AI ripoffs of their work - Designers are suing Shein, a Chinese fast-fashion firm, for allegedly using AI to steal independent artists' works, claiming that the firm's AI algorithm generates exact copies that can damage their careers.
WormGPT – The Generative AI Tool Cybercriminals Are Using to Launch Business Email Compromise Attacks - Generative AI, such as OpenAI's ChatGPT and the cybercrime tool WormGPT, is being used by cybercriminals to launch Business Email Compromise (BEC) attacks, automating the creation of highly convincing fake emails personalized to the recipient.
While robots and artificial intelligence have the potential to act as nurses, it raises several ethical concerns - Robots and AI have the potential to act as nurses, but there are ethical concerns regarding their ability to replicate human qualities such as empathy, accountability, cooperation, and caring.
Inside the White-Hot Center of A.I. Doomerism - Anthropic, an artificial intelligence start-up, has strong ties to the effective altruism movement, despite distancing itself from the movement after a scandal involving one of its investors.
Mother’s iPhone photo of son rejected from Sydney competition after judges ruled it could be AI - A mother's iPhone photo of her son was rejected from a photography competition in Sydney after judges suspected it had been generated by artificial intelligence.
AI-text detection tools are really easy to fool - AI-text detection tools, designed to spot whether text was written by a human or a machine, have been found to have high rates of false positives and false negatives, raising concerns about their reliability and potential consequences for students accused of cheating.
Analysis
Me, Myself and AI - AI is poised to transform various aspects of our lives, including work, healthcare, relationships, and creativity, but it also presents challenges and risks that need to be carefully considered and managed.
LLM agents and integration dead-ends - The article discusses the challenges and barriers to integrating large language models (LLMs) into various applications and tools, highlighting the need for trust, reliability, and security in deploying LLM agents.
I Asked AI What Europeans Think Americans From Every Single State Look Like And The Results Are Just Plain Mean - Generative AI image models create biased and stereotypical depictions of Americans from each state, revealing the flaws in AI models.
Policy
Stanford HAI Faculty Urge President Biden to Approach AI with a Moonshot Mentality - Stanford HAI faculty members urge President Biden to invest in AI like NASA and lead the way in unlocking its potential, emphasizing the need for collaboration, responsible stewardship, and expanding access to resources and research.
Full Speed Ahead on A.I. Our Economy Needs It. - The launch of ChatGPT and other generative A.I. platforms has sparked concerns about the fate of white-collar workers and the potential non-economic impacts of A.I.
VP Harris meets with labor and rights groups to discuss threats from AI - Vice President Kamala Harris meets with labor and rights groups to discuss the false choice between AI innovation and safety, emphasizing the need to protect consumers and advance innovation simultaneously.
US senators to get classified White House AI briefing Tuesday - US senators will receive a classified briefing from the White House on artificial intelligence as they consider implementing legislative safeguards for the technology.
US FTC opens investigation into OpenAI - The US Federal Trade Commission (FTC) has opened an investigation into OpenAI, the maker of ChatGPT, on claims that it has violated consumer protection laws by putting personal reputations and data at risk.
China's slow AI roll-out points to its tech sector's new regulatory reality - China's tech sector is experiencing a regulatory crackdown, with companies slowing down their AI roll-out and complying with new rules, leading to a lack of AI chatbots available to the general public and potential concerns about future competitiveness with Western rivals.
What should the UK’s £100 million Foundation Model Taskforce do? - The UK's £100 million Foundation Model Taskforce should focus on evaluating frontier models, exploring unknown capabilities, assessing alignment and misalignment risks, developing the science of measurement, and experimenting with different approaches to accessing and aligning AI models.
UK universities draw up guiding principles on generative AI - UK universities have developed guiding principles to ensure that students and staff are AI literate and can use generative AI appropriately, while also addressing issues of plagiarism, bias, and inaccuracy.
China finalizes first-of-its-kind rules governing generative A.I. services like ChatGPT - Chinese regulators have finalized rules governing generative artificial intelligence services, requiring providers to obtain a license and take measures to prevent the generation of "illegal" content.
Discord bans AI-generated child sex abuse material and teen dating - Discord is changing its child safety policies to ban AI-generated child sexual abuse material and teen dating, in response to concerns raised by an NBC News investigation.
Producers allegedly sought rights to replicate extras using AI, forever, for just $200 - Hollywood studios are allegedly seeking to pay extras $200 for the rights to use their likenesses in AI forever, a move that has been opposed by the labor union SAG-AFTRA.
NY Lawmaker Uses AI to Write Housing Bill That Does Little to Help Renters - A New York lawmaker used AI to write a housing bill that requires landlords to provide renters with a copy of their lease agreement upon request, but the bill is seen as ineffective and does little to address the larger issues facing renters in the city.
Actors say Hollywood studios want their AI replicas — for free, forever - Hollywood actors are going on strike after Hollywood studios proposed using AI replicas of actors' likenesses without consent or compensation.
Fun
Is A.I. the Greatest Technology Ever for Making Dumb Jokes? - The A.I. revolution has arrived, and while some see it as a field-transforming progress and others see it as a source of confusion and shoddy work, there is a Funposting Zone where cutting-edge artificial intelligence technology combines with advanced human stupidity to create funny, pointless, and gleefully stupid content.
Copyright © 2023 Skynet Today, All rights reserved.
Last Week in AI #228: Hospitals testing Google's medical AI chatbot, Elon Musk's new X.AI company, 65% AI startups have immigrant founders, Claude 2 w/ 100k context length, and more!
Is it just me, or did this feel like another "firehose" week?