AI Weekly: 07/10/23

Good morning and welcome to this week’s edition of AI Weekly! Coming off of July 4th weekend, we had a relatively slower week from a news perspective. No big fundraising rounds, no big models released, but that paved the way for a handful of other fun headlines that may not have regularly surfaced to the top.

OpenAI remained in the news for a multitude of reasons, good and not so good. For the first month since its release, ChatGPT desktop and mobile traffic experienced a decline. In news related to ChatGPT, comedian Sarah Silverman as well as four other authors are suing OpenAI for copyright infringement, claiming that the company trained ChatGPT on their content without their consent.

In more positive OpenAI news, however, the Sam Altman-led juggernaut announced that it is creating a new team, called Superalignment, to develop ways to steer and control "superintelligent" AI systems. This comes in the wake of month-long discussions around how to properly keep this society-altering technology in check.

Unrelated to OpenAI, Inflection AI has built a supercomputer equipped with 22,000 NVIDIA H100 GPUs, “Auto Gmail” is going to write our emails for us, and the CEO of the Grammy’s deemed AI-created elements in music eligible for entry and consideration. Enjoy more AI news below!

- ZG

Here are the most important stories of the week:

TEXT

ChatGPT desktop and mobile traffic has experienced a decline for the first time, raising questions about the AI chatbot's ongoing appeal. Link.

• According to Similarweb, ChatGPT traffic globally dropped by 9.7% and in the U.S. by 10.3% from May to June. Unique visitors worldwide also decreased by 5.7%, and average time spent on the site went down by 8.5%.

• ChatGPT initially gained immense popularity, becoming the fastest-growing app ever and amassing 100 million users within two months of its November 2022 release.

• Various factors may have contributed to the decline, including the launch of the ChatGPT app for iOS, potentially diverting traffic from mobile and desktop browsers.

• The decrease in traffic during summer can be attributed to fewer students seeking homework help from ChatGPT, with search interest for Minecraft rising simultaneously.

• The availability of alternative AI chatbots like Character.AI and Google Bard, which experienced increased traffic, might also account for ChatGPT's declining usage despite the continued popularity of generative AI.

Meet Auto Gmail: the generative AI assistant designed to auto-respond to emails and provide drafts for review, aiming to help users handle their email workload more efficiently. Link.

• The creator, Nathan Ganser, developed Auto Gmail to address customer support inquiries for one of his products by training ChatGPT to generate responses based on FAQ documents.

• Auto Gmail functions by creating drafts in Gmail rather than sending emails directly, allowing users to review and modify the responses before sending them manually.

• The tool is particularly useful for repetitive or templated responses but may not be suitable for personal or work-related emails that require specific context or individualized answers.

• Auto Gmail relies on a set of predefined rules and instructions for generating replies, ensuring consistency and adherence to the desired response style.

• It is recommended for small businesses seeking automation benefits but may not be appropriate for personal email use. A free trial of Auto Gmail is available for users to explore its features.

Google updated its privacy policy to disclose that its AI services, including Bard and Cloud AI, may be trained on publicly available data scraped from the web. Link.

• The updated policy clarifies that information may be used to train AI models and develop new products and features.

• The policy does not specify how Google will prevent copyrighted materials from being included in the data pool.

• Makers of popular generative AI systems, like OpenAI's GPT-4, have been cautious about disclosing their training data sources due to legal uncertainties and potential lawsuits.

• Stricter laws and regulations are being introduced to govern the collection and use of training data by AI companies.

• Gannett, the largest newspaper publisher in the US, is suing Google for alleged monopolistic practices, and Twitter and Reddit have implemented measures to restrict data harvesting, leading to community backlash.

IMAGE/VIDEO

YouTube is testing AI-generated quizzes on its mobile app to help viewers learn more about educational video topics and assess how well the content covers those topics. Link.

• The quizzes are being rolled out globally to a small percentage of users who watch educational videos, with the feature available for select English-language content.

• While it remains to be seen if the feature will become a permanent addition to YouTube, the platform has a strong educational focus and hosts popular educational channels.

• The quiz feature could be useful for individuals genuinely interested in gaining a deeper understanding of educational material.

• YouTube has previously collaborated with Crash Course and Arizona State University to provide free access to educational courses through its "Study Hall" initiative.

• In addition to the quizzes, YouTube has been testing other features such as a three-strikes ad-blocking policy and a lock screen feature for Premium subscribers.

A deepfake video of consumer finance champion Martin Lewis endorsing an investment scam is circulating on Facebook (Meta), with Lewis expressing anger about the development. Link.

• Lewis previously sued Facebook over scam ads featuring his image and settled the defamation suit in 2019, but he remains frustrated with the lack of regulation for scam adverts.

• The deepfake video is believed to be the first deepfake scam advert of its kind, and Lewis warns that the technology will only improve and become more deceptive.

• Lewis criticizes both Meta and the UK government for not doing enough to tackle the issue of scam ads and the lack of enforcement of terms and conditions on social media platforms.

• Meta’s ad tools contribute to the problem by allowing scammers to target vulnerable users with deceptive messages, leveraging the platform's surveillance and profiling capabilities.

• Lewis emphasizes the destructive impact of scam ads on vulnerable individuals and urges viewers to remember that Big Tech firms profit from promoting such ads, calling for increased regulation and accountability.

SPEECH/AUDIO

The Recording Academy has introduced new guidelines for the Grammy Awards, including rules regarding the use of artificial intelligence in music. Link.

• Only works with human authorship are eligible for Grammy consideration, according to the guidelines.

• AI-created elements in music are eligible for entry and consideration, but the AI portion will not receive a Grammy or nomination.

• If an AI or voice modeling program performs the lead vocal, the song would be eligible in a songwriting category but not a performance category.

• The Recording Academy aims to ensure that human creativity is enhanced, embellished, or additive to technology, rather than being replaced by it.

• Extensive research and engagement with the copyright office and tech summits were conducted to establish the AI guidelines, driven by recent advancements and discussions surrounding AI in music.

CODE/DEVTOOLS

OpenAI has announced the general availability of GPT-4, its latest text-generating model, through its API. Link.

• Existing OpenAI API developers with a history of successful payments can access GPT-4 starting today, while access for new developers will open up by the end of the month.

• GPT-4 can generate text and accept image and text inputs, performing at a "human level" on professional and academic benchmarks.

• The image-understanding capability of GPT-4 is currently being tested with a single partner, Be My Eyes.

• GPT-4 is not perfect and may hallucinate facts, make reasoning errors, and fail at challenging problems like introducing security vulnerabilities into generated code.

• OpenAI plans to allow developers to fine-tune GPT-4 and GPT-3.5 Turbo with their own data, and they are also making their DALL-E 2 and Whisper APIs generally available. Older models like GPT-3 will be deprecated and replaced with new base GPT-3 models for improved compute efficiency.

HARDWARE/ROBOTICS

Inflection AI has built a supercomputer equipped with 22,000 NVIDIA H100 GPUs, making it one of the largest in the industry. Link.

• Inflection AI is known for its Inflection-1 AI model, which powers the Pi chatbot and performs well on "common sense" tasks.

• The supercomputer will utilize an estimated 31 Mega-Watts of power and consist of almost 700 four-node racks of Intel Xeon CPUs.

• Acquiring 22,000 NVIDIA H100 GPUs is notable since they are in high demand and difficult to obtain.

• NVIDIA is considering investing in Inflection AI, which may explain the company's access to a large number of GPUs.

• Inflection AI has raised $1.5 billion in investments and is currently valued at $4 billion, with expectations that the supercomputer will improve their AI model's performance, particularly in coding tasks.

Nadine, a social robot powered by AI, could play an important role in caring for the sick and elderly, according to its inventor, Nadia Magnenat Thalmann from the University of Geneva. Link.

• Thalmann believes robots like Nadine can be more effective than human carers because they have 24-hour availability.

• The use of humanoid robots in healthcare can help address staffing crises in care homes and provide care and assistance to vulnerable individuals.

• Nadine has demonstrated conversational skills and interacted with residents at a Singapore nursing home, engaging in activities such as talking, singing, and playing bingo.

• Other robots showcased at the conference include 'PARO', a baby seal robot that assists dementia and Parkinson's patients, and nursing assistant Grace.

• Nadine has recently been upgraded with the AI model GPT-3, enhancing its interaction and conversation abilities, allowing it to express more complex concepts.

HEALTHCARE

Google's Med-PaLM 2, an AI tool for medical information, has been tested at the Mayo Clinic research hospital since April. Link.

• Med-PaLM 2 is an updated version of Google's language model, PaLM 2, which powers Google's Bard.

• The tool aims to be particularly useful in countries with limited access to doctors, according to an internal email seen by The Wall Street Journal.

• Med-PaLM 2 was trained on medical expert demonstrations, making it better for healthcare conversations compared to generalized chatbots.

• A study showed that Med-PaLM 2 still has accuracy issues but performed well in terms of reasoning, consensus-supported answers, and comprehension.

• Customers testing Med-PaLM 2 will have control over their data, which will be encrypted, and Google won't have access to it. The tool is still in its early stages of development.

POLICY/LAW/ETHICS

New York City has enacted a law to protect workers from bias in hiring when companies use artificial intelligence, but experts question its effectiveness. Link.

• The law requires employers to conduct annual third-party "bias audits" to ensure their AI technology is free from racist or sexist bias.

• It also mandates employers to notify job candidates about the use of AI tools in the hiring process.

• AI tools in hiring have become common for handling large volumes of applications efficiently, but concerns exist about perpetuating biases.

• The law has loopholes, including limited focus on racial and gender discrimination and unclear penalties enforcement.

• The law does not specify an acceptable level of disparity between gender and racial or ethnic groups and relies on existing anti-discrimination laws for recourse. It also excludes recruitment practices from regulation.

OpenAI is creating a new team, called Superalignment, to develop ways to steer and control "superintelligent" AI systems, led by Ilya Sutskever and Jan Leike. Link.

• The team aims to address the challenge of controlling potentially superintelligent AI systems that may not be benevolent or easily supervised by humans.

• OpenAI believes that current techniques for aligning AI, such as reinforcement learning from human feedback, won't be effective for systems smarter than humans.

• The Superalignment team will have access to 20% of OpenAI's compute resources and will work on core technical challenges of controlling superintelligent AI over the next four years.

• The team plans to build a "human-level automated alignment researcher" that can train AI systems, evaluate other AI systems, and conduct alignment research to achieve desired outcomes and prevent AI from going astray.

• OpenAI acknowledges limitations and potential risks, such as scaling up inconsistencies or biases, but believes that machine learning experts are critical to solving the superintelligence alignment problem and aims to share its research broadly.

Shutterstock is offering full indemnification to enterprise customers using its generative AI images, protecting them against potential claims related to image use. Link.

• The indemnification is triggered based on the licensing agreement, ensuring compliance with copyright, privacy, and legal requirements.

• This move follows Adobe's similar offering, highlighting increased competition in the enterprise market for safe and ethical generative AI content.

• Shutterstock has been integrating generative AI into its platform since October 2022, partnering with OpenAI and allowing Meta to use its stock datasets.

• The company's AI image generator is trained on ethically-sourced assets to ensure commercial safety and compensate contributing artists.

• Shutterstock's Contributor Fund has compensated artists for their content's role in training generative technology, with ongoing royalties tied to licensing activity.

Comedian Sarah Silverman and authors Christopher Golden and Richard Kadrey are suing OpenAI and Meta in a US District Court for copyright infringement. Link.

• The lawsuits claim that OpenAI's ChatGPT and Meta's LLaMA were trained on illegally-acquired datasets containing the plaintiffs' works, obtained from "shadow library" websites.

• Exhibits provided in the OpenAI lawsuit show that ChatGPT can summarize the plaintiffs' books, infringing on their copyrights.

• The lawsuit against Meta alleges that the authors' books were accessible in datasets used to train LLaMA models, pointing to the origin of the datasets as "flagrantly illegal" sources.

• The authors assert that they did not consent to the use of their copyrighted books for training the AI models and are seeking damages, restitution, and more.

• These lawsuits challenge the limits of copyright and are part of a broader trend of legal disputes surrounding AI and copyright law.

Authors Mona Awad and Paul Tremblay have filed a lawsuit against OpenAI, claiming that their copyrighted books were unlawfully used to train the ChatGPT AI model. Link.

• ChatGPT is an AI-powered chatbot that responds with human-like text based on publicly available internet data.

• The lawsuit argues that ChatGPT generated accurate summaries of the novels, indicating that the books were used in its training.

• The case will explore the legal boundaries of generative AI and copyright, as it is the first lawsuit against ChatGPT related to copyright infringement.

• The authors' lawyers argue that OpenAI unfairly profits from stolen writing and ideas and are seeking monetary damages for US-based authors whose works were allegedly used.

• It may be challenging to prove financial losses specifically caused by ChatGPT's training on copyrighted material, but the lawsuit raises questions about fair use and unauthorized copying.

Valve, the developer of Steam, issued a statement clarifying its position on games with AI-generated assets after claims that it was rejecting such games from its platform. Link.

• Valve's policy is evolving, and the rejection of games with AI-generated content is not a deliberate stance against AI.

• The issue arose when an indie developer posted about Valve's refusal to publish games with AI-generated assets, citing unclear legal ownership of such content as the reason.

• Valve's warning letters to the developer stated that the game could not be shipped unless the developer could affirmatively confirm ownership of the IP rights to the training data used for the AI-generated assets.

• The policy, as stated, appears to be a blanket ban on AI-generated assets in games, which raises questions about the liability and copyright implications surrounding generated art.

• Valve clarified that its review process is based on current copyright laws and policies and is not meant to discourage the use of AI on Steam, but rather to work out how to integrate it within existing policies. They will refund the app submission fee in cases where this policy is the deciding factor for rejection.

OTHER

Mastercard is offering an AI-powered tool called Consumer Fraud Risk to help banks identify customers attempting to send money to fraudsters. Link.

• Nine major UK banks, including Lloyds, Natwest, and Bank of Scotland, have signed up to use the tool, which is trained on years of transaction data to predict transfers to accounts associated with authorized push payment scams.

• The tool comes as UK banks prepare for new regulations requiring compensation for APP scam victims starting in 2024.

• APP scams have been difficult to detect as victims willingly make the transfers, passing security checks used for other types of fraud.

• In the UK, APP scam losses reached £484.2 million ($616 million) in 2022, with global losses expected to reach $5.25 billion by 2026.

• TSB Banking Group, the first bank to implement the system, saw a 20% increase in detection of this fraud type and believes the tool could save UK banks around £100 million annually if adopted industry-wide. Mastercard plans to expand the tool globally and is in discussions with potential clients in markets with significant APP fraud.

The US military is conducting experiments with large-language models to aid decision-making and data integration across the military. Link.

• The experiments involve LLMs such as OpenAI's ChatGPT and Google's Bard, as well as Scale AI's Donovan.

• The use of LLMs could greatly improve the speed and efficiency of information retrieval and decision-making within the military.

• The experiments aim to explore the potential of LLMs in providing new options and insights that military officials may not have considered before.

• Concerns about bias, misinformation, and security vulnerabilities associated with generative AI are being addressed through careful evaluation and collaboration with tech security companies.

• In a demonstration, an LLM provided insights on the US military's ability to deter a Taiwan conflict, highlighting the need for direct intervention and the uncertainty of the outcome in a potential conflict with China.