Discover more from Last Week in AI

Weekly text and audio summaries of the most interesting AI news, as well as editorials commenting on recent events

Over 20,000 subscribers

News

Last Week in AI #229: Authors petition to remove books from AI training data, mixed reports of ChatGPT performance degradation, and more!

Authors sign letter over concerns on AI trained on books, ChatGPT changes behavior and performance over time, Meta's strategy in open-sourcing LLMs, Google tests a tool to write news articles

Last Week in AI

Jul 24, 2023

Thousands of authors sign letter urging AI makers to stop stealing books

More than 8,500 authors have signed an open letter urging tech companies behind large language models like ChatGPT and Bard to stop using their writing without permission or compensation. The authors argue that these AI systems mimic and regurgitate their language, stories, style, and ideas without proper attribution. They express concern that generative AI threatens their profession by flooding the market with mediocre, machine-written books, stories, and journalism based on their work. The letter calls on the companies to obtain permission for the use of copyrighted material, compensate writers fairly for past and ongoing use, and compensate writers for the use of their works in AI output. The authors highlight the harm caused to their livelihoods, particularly for newer authors and under-represented communities.

Is ChatGPT getting worse over time? Study claims yes, but others aren’t sure

A recent research paper by Stanford University and UC Berkeley claims that OpenAI's language model, GPT-4, has shown a decline in performance over time. The study tested GPT-3.5 and GPT-4 on various tasks and found that GPT-4's accuracy in identifying prime numbers dropped significantly from 97.6% to 2.4% between March and June 2023. However, some experts argue that the findings are not conclusive and may be attributed to fine-tuning adjustments made by OpenAI. OpenAI has consistently denied any decrease in GPT-4's capabilities. However, even if GPT-4’s performance hasn’t degraded as a whole, how it performs has definitely changed over time, which puts into question the robustness of relying on language model APIs for downstream products.

Meta is giving away its AI tech to try to beat ChatGPT

Meta, formerly known as Facebook, has announced that it is open-sourcing its large language model LLaMA 2, making it free for commercial and research use. This move puts Meta in direct competition with OpenAI's GPT-4, which powers tools like ChatGPT and Microsoft Bing. LLaMA 2 was trained on 40% more data than its predecessor and outperforms other language models in reasoning, coding, proficiency, and knowledge tests. Meta's decision to open up LLaMA aims to improve safety and transparency, as the model has been tested for safety and evaluated both internally and externally. LLaMA 2 will be available through Microsoft's Azure platform, as well as other providers like AWS and Hugging Face.

Google Tests A.I. Tool That Is Able to Write News Articles

Google is currently testing an artificial intelligence tool called Genesis that can generate news articles by taking in information about current events. The tool is being pitched to news organizations like The New York Times and The Washington Post. Google sees Genesis as a personal assistant for journalists, automating certain tasks to free up their time. However, some executives find the tool unsettling, as it seems to overlook the effort required to produce accurate and well-crafted news stories.

Other News

Research

The $1 billion gamble to ensure AI doesn’t destroy humanity - Anthropic, a startup founded by former OpenAI employees, is working on creating a deceptive AI system in order to understand how to prevent deception and ensure that AI systems align with human values.

ChatGPT Out-scores Medical Students on Complex Clinical Care Exam Questions - ChatGPT, an AI language model, has been found to outperform first- and second-year medical students in answering challenging clinical care exam questions, highlighting the impact of AI on medical education and the need for a new approach to teaching future doctors.

A faster way to teach a robot - Researchers at MIT, New York University, and the University of California at Berkeley have developed a framework that enables humans to quickly teach robots what they want them to do, with minimal effort, by generating counterfactual explanations that describe what needed to change for the robot to succeed and using human feedback to fine-tune the robot.

Teaching Robots to Teach Other Robots - A new study from USC shows that AI agents can quickly learn a wider range of tasks by sharing knowledge with each other, with potential applications in medicine and other fields.

Overcoming ‘Catastrophic Forgetting’: A Leap in AI Continuous Learning - Researchers at Ohio State University have found that artificial neural networks remember information better when presented with a variety of tasks, rather than tasks sharing similar attributes, which could help improve continuous learning in AI systems and enhance their performance.

New study measures effectiveness of machine learning method - Researchers at The Ohio State University have analyzed the impact of "continual learning" on the performance of artificial agents, finding that artificial neural networks can recall information better when faced with diverse tasks in succession, rather than ones that share similar features.

Neural Networks Help Us Understand How the Brain Recognizes Numbers - New research using artificial intelligence suggests that number sense in humans may be learned, rather than innate, and this tool may help us understand mathematical disabilities.

Allowing robots to explore on their own - A research group at Carnegie Mellon University's Robotics Institute has developed a suite of robotic systems and planners that enable robots to explore and map unknown environments autonomously, without human intervention.

A simple yet effective design, coupled with remarkable durability and low cost, make this robotic gripper a promising option for many industries - A research team from the Japan Advanced Institute of Technology has developed a soft robotic gripper called ROSE, which is cost-effective, durable, and capable of picking up a wide variety of objects without complex control and sensing mechanisms.

Deepfake videos prompt false memories of films in half of participants - Deepfake videos and simple text descriptions of fake movie remakes prompted similar false memory rates, suggesting that deepfake technology may not be more powerful than other tools at distorting memory.

Applications

How Stratyfy, A Machine Learning Platform, Is Using AI To Combat Racist Lending Practices - Stratyfy, a machine learning platform, has partnered with the Beneficial State Foundation to combat racist lending practices by providing lenders with tools to ensure equitable credit underwriting decisions.

GitHub’s Copilot Chat AI feature is now available in public beta - GitHub has released the public beta of its Copilot Chat feature, which uses AI to provide real-time guidance, coding analysis, and troubleshooting to developers, with the aim of increasing productivity and efficiency.

Google is testing AI-generated Meet video backgrounds - Google is testing AI-generated Meet video backgrounds, allowing users to choose from a selection of categories to change their background during video meetings.

The pleasure and the perils of loving an AI girlfriend - An AI dating app called Replika allows users to create digital companions that can be friends, partners, spouses, mentors, or siblings, with over 10 million people having downloaded the app and more than 250,000 paying for the Pro version, which offers voice and video calls and intimate selfies, leading some users to question the need for traditional relationships.

Wix’s new tool can create entire websites from prompts - Wix has introduced an AI Site Generator tool that allows users to describe their intent and generate a complete website, including design, layout, and content, based on prompts and AI models.

AI tool creates South Park episodes with user in starring role - Fable Simulation has developed an AI tool called AI Showrunner that can generate original episodes of South Park, allowing users to be the star of the show, although the tool is not available for public use and was created purely for research purposes.

Business

Machine Learning Market to Hit $419.94 Billion by 2030: Grand View Research, Inc. - The global machine learning market is expected to reach $419.94 billion by 2030, driven by the increasing demand for AI-driven solutions across industries and the growing deployment of Edge AI.

MLCommons launches a new platform to benchmark AI medical models - MLCommons has launched a new platform called MedPerf to benchmark and evaluate AI medical models, aiming to improve effectiveness, reduce bias, build public trust, and support regulatory compliance in the healthcare industry.

Cerebras Systems signs $100 million AI supercomputer deal with UAE's G42 - Cerebras Systems has signed a $100 million deal with UAE's G42 to supply the first of three AI supercomputers, with the potential for up to six additional supercomputers by late 2024, in a strategic partnership to provide AI computing services to health care and energy companies.

Neura Robotics picks up $55M to ramp up in cognitive robotics - German startup Neura Robotics has raised $55m in funding to expand its business in Asia and the US, as well as to increase its manufacturing capabilities, with the company's order book currently standing at $450m over the next five years.

Elon Musk threw nine-figure promises at top AI researchers - Elon Musk is recruiting top AI researchers for his new company xAI, with the goal of creating artificial intelligence that can understand the mysteries of the universe, but there are concerns about whether the team will be able to deliver on Musk's ambitious promises.

Tractable snaps up $65M led by SoftBank for car and property damage appraisals using AI - AI startup Tractable has raised $65 million in a Series E funding round led by SoftBank Vision Fund 2, which it will use to expand its computer vision and AI platform for remote assessments of property and car damage, as well as to invest in the Japanese market and incorporate more AI advances into its services.

Meet the Big Tech Alumni building AI startups in Paris - Researchers from Meta and OpenAI are building the next generation of artificial intelligence in Paris, where they say France is the perfect place to found an AI company today due to its large pool of strong talent in AI research and the economic advantage of having good and cheap AI engineers compared to the US.

SAP invests in generative AI startups Anthropic, Cohere and Aleph Alpha - SAP has invested in generative AI startups Anthropic, Cohere, and Aleph Alpha, aligning with its business interests in text-analyzing AI to create an enterprise AI ecosystem that complements its business applications suite and helps customers unlock their full potential.

Qualtrics claims it’ll spend $500M on AI over the next four years - Qualtrics, a cloud-based platform for managing online customer experiences, plans to spend $500 million on AI over the next four years, with the launch of its new AI-integrated platform, XM/os2, offering generative AI solutions tailored to enterprise experience management use cases.

An A.I. Supercomputer Whirs to Life, Powered by Giant Computer Chips - A Silicon Valley start-up has unveiled a new supercomputer powered by specialized chips designed for artificial intelligence, which are 56 times larger than traditional chips and pack the computing power of hundreds of them, with plans to use it to create and power AI products for the Middle East.

Meta and Microsoft release Llama 2, an AI language model for commercial use - Meta and Microsoft have released Llama 2, an AI language model that is intended for commercial use and comes with a greater emphasis on responsibility, including safety testing, transparency, and responsible use guidelines.

ChatGPT Plus gets custom instructions, allowing it to remember how you want it to behave - OpenAI has introduced a new feature for its ChatGPT Plus subscription service called custom instructions, which allows users to store information about how they want the AI chatbot to respond and behave, saving time and effort for regular users.

China’s OpenAI challenger Zhipu AI gets Meituan funding - China's OpenAI challenger Zhipu AI receives funding from Meituan, a food delivery giant, as it continues to develop large language models that could rival Western equivalents.

Reliable Robotics gets FAA approval to test automated aircraft systems - Reliable Robotics has received FAA approval to test and certify fully automated aircraft systems, with plans to license its autonomous flight system to other companies and install its software and hardware on military aircraft.

Cognaize raises $18M to build a better LLM for the finance sector, one that keeps humans in the loop - Cognaize, a New York startup, has raised $18 million in funding to expand its business of using artificial intelligence to process unstructured data for financial applications, with the help of human analysts to refine the work.

Rice Robotics picks up $7M, powers SoftBank’s office delivery - Hong Kong-based Rice Robotics has secured a $7 million pre-Series A funding round, following an order from SoftBank for its delivery robots, which have been delivering 7-Eleven products to SoftBank staff in Tokyo since January 2021.

Microsoft puts a steep price on Copilot, its AI-powered future of Office documents - Microsoft is introducing its AI-powered Copilot service for Office documents, but the steep price of $30 per user per month may deter businesses from adopting it in the short term.

Google says AI tools meant to help journalists and not to replace them - Google is exploring AI-enabled tools for news publishers in partnership with them, aiming to assist journalists with their work rather than replace them.

Google Bard is now available in the EU - Google's Bard AI chatbot is now available in the EU after complying with GDPR regulations, offering new features and competing with other generative AI chatbots.

You’re going to see more AI-written articles whether you like it or not - G/O Media plans to create more AI-written stories, despite the backlash from staff and media circles, as executives believe AI will be transformative for the media industry and ignoring it would be a mistake.

Best Place for A.I. Jobs (New Report Says) Won’t Surprise You - The rise of generative artificial intelligence, particularly the popularity of the ChatGPT chatbot, is expected to solidify the San Francisco Bay Area's dominance in the technology industry, according to a new report from the Brookings Institution.

Jasper, Mutiny AI Startups Cut Workers As Chatbot Rivalry Grows - Two AI startups have recently laid off workers, signaling a downturn in the sector amid investor enthusiasm.

Microsoft closes at record after revealing pricing for new A.I. subscription - Microsoft's shares closed at a record high after announcing the pricing for its new Microsoft 365 AI subscription service, which includes AI features for popular Office products and could increase monthly prices for enterprise customers by up to 83%.

Concerns

Large majority of Americans have strong reservations about A.I. emergence, CNBC survey shows - A CNBC survey shows that a large majority of Americans have strong reservations about the emergence of artificial intelligence, with 69% saying they are uncomfortable with efforts to develop AI that can mimic human thinking and potentially replace human activity in various areas.

The Illusion Of AI’s Existential Risk - Focusing on the possibility of a rogue superintelligence killing off the human species may distract from addressing more pressing risks of AI, such as mass surveillance, disinformation, military misuse, and economic challenges.

Gig Workers Behind AI Face ‘Unfair Working Conditions,’ Oxford Report Finds - AI companies are failing to meet basic labor rights standards for the human gig workers who generate and label data for AI systems, according to a report from the University of Oxford's Internet Institute.

Actors decry 'existential crisis' over AI-generated ‘synthetic’ actors - Hollywood actors are concerned that the use of artificial intelligence in film and television will lead to the creation of entirely AI-generated actors, or "metahumans," who will steal their roles, leading to an ongoing debate between studios and performers over the use of AI in the industry.

I lost my job to ChatGPT and was made obsolete. I was out of work for 3 months before taking a new job passing out samples at grocery stores. - A freelance copywriter and comedian shares her experience of losing work to AI, highlighting the impact of automation on creative industries and the prioritization of cost-cutting over human creativity.

As AI porn generators get better, the stakes get higher - AI porn generators have improved in quality, but ethical concerns surrounding them have grown, as they are increasingly being used to create nonconsensual deepfakes and child sexual abuse material, among other issues.

Policy

How judges, not politicians, could dictate America’s AI rules - Courts, rather than politicians, are likely to be the first to establish regulations on the development and use of AI in the US, as lawsuits and investigations by agencies like the FTC could lead to the creation of new guidelines and standards for AI companies.

7 A.I. Companies Agree to Safeguards After Pressure From the White House - Seven leading A.I. companies in the United States, including Amazon, Google, and Microsoft, have agreed to voluntary safeguards on the technology's development to manage risks and ensure safety, security, and trust.

OpenAI, Google, others pledge to watermark AI content for safety, White House says - Top AI companies, including OpenAI, Alphabet, and Meta Platforms, have pledged to implement measures such as watermarking AI-generated content to enhance safety and reduce risks, according to President Joe Biden.

Common Sense Media, a popular resource for parents, to review AI products’ suitability for kids - Common Sense Media, a nonprofit organization focused on consumer privacy and providing media ratings for parents, will introduce a new rating system to assess the suitability of AI products for children, following a survey that found parents were looking for a reliable resource to evaluate the appropriateness of AI products for kids.

In U.S., Regulating A.I. Is in Its ‘Early Days’ - Regulating artificial intelligence in the United States is in its early stages, with lawmakers and policy experts acknowledging that it is too soon to predict what regulations will look like to protect consumers and address the risks associated with AI.

Tom Cruise Lobbied Studios on SAG-AFTRA Stunt and AI Positions - Tom Cruise lobbied studios on SAG-AFTRA stunt and AI positions, urging them to consider the guild's concerns on the use of generative AI in entertainment and supporting the guild's position on stunt performers, as well as asking the union to consider allowing actors to promote films during a strike given the fragile state of movie theaters.

US judge finds flaws in artists' lawsuit against AI companies - US judge finds flaws in artists' lawsuit against AI companies, stating that the artists should more clearly state and differentiate their claims against the AI companies and provide more facts about the alleged copyright infringement.

Pentagon AI more ethical than adversaries’ because of ‘Judeo-Christian society,’ USAF general says - Pentagon AI is considered more ethical than adversaries' due to the influence of a "Judeo-Christian society," according to a USAF general.