Your AI Update - March 2024
Keep up with recent news in the world of Generative AI, including new features, AI in the workplace, social and ethical implications, regulations, and research revelations from the past month (15m read).
Tech Titans: New Features, Products, and More
Gemini Derails, Days After Leaving the Station
Shortly after its initial launch in February, Google removed the ability for users to generate images of humans with their flagship product Gemini. During last month’s hypefest, this product was cast as Google’s official entrance into the AI space, with a multi-modal, user-friendly product to rival ChatGPT and Microsoft’s Copilot. Despite all of this initial hype, Gemini only lasted until mid-March without encountering serious controversy. Users began to notice the model refraining from generating white people, even when appropriate in context. Word of this spread through several viral posts on X, with Elon Musk and conservative voice Jordan Peterson leading this discourse by accusing Google of pushing a leftist agenda on its users. Additionally, users noticed the language generated by Gemini to be unpredictable and ignorant of context. For example, when asking the model to generate a children’s story, it incorporated mature ideas of depression and anxiety akin to the modern mental health transparency movement. Two anonymous employees testified that “when Google launched the tool, it included a technical fix to reduce bias in its outputs.” Those at the company believe that the guidelines put in place to prevent the generation of racially biased or hateful content has caused the bot to ignore certain contexts, with TIME stating “Google’s overcorrection for AI’s well-known bias against people of color left it vulnerable to yet another firestorm over diversity. In a statement from the CEO, Google said that these diversity blunders are “completely unacceptable” and “we got it wrong.”
Claude 3 Makes Waves
Amazon-backed startup Anthropic released the third version of their flagship model Claude, adding to the suite of models and chatbots at the consumer’s disposal. Claude 3 is a multimodal model, meaning it can interpret and analyze images in addition to “read, understand, and respond to thousands of pages of text in seconds.” With this update came three new products, Haiku, Sonnet, and Opus, in increasing order of sophistication. Opus, the most advanced, leads the competitors on “undergraduate level expert knowledge (MMLU), graduate-level expert reasoning (GPQA), basic mathematics (GSM8K), and more.” Haiku is the fastest, most compact model, while Sonnet is engineered for larger-scale deployment, intended for enterprise use. Between the three models, the Claude 3 suite can flex to consumer needs and price considerations while still delivering “near-human” comprehension. Previous iterations of Claude (1 and 2) were notably subpar in comparison to GPT-4, however, Anthropic asserts that the model is outperforming market leader OpenAI on 10 major benchmarks. Read more about how to best leverage the potential of the Claude suite for your needs here.
Elon Musk Takes on ChatGPT
Since launching Grok, the snarky chatbot to rival ChatGPT, Elon Musk has laid relatively low on the AI scene. However, he broke that silence profoundly this month, taking two major swings at the current market leader, OpenAI. First, he sued the company for prioritizing profit over its original mission. “OpenAI, Inc. has been transformed into a closed-source de facto subsidiary of the largest technology company in the world: Microsoft,” the lawsuit states. Musk is a former member of the board of OpenAI and has previously criticized CEO Sam Altman. This time, he referenced the original agreement between the founders to produce an “open source, non-profit company” focused on building an “artificial general intelligence (AGI), a concept that machines could handle tasks like a human, but in a way that would ‘benefit humanity.’” Now, after partnering with Microsoft and shifting gears toward making money, Musk attests that the company is aligned with the tech giants it sought to counter on principle.
In addition, as CEO of his own AI startup, xAI, Musk made huge waves by making his model ‘Grok’ open-source. An open-source model, meaning source code that is available for programmers to access and modify, was an original sticking point of Musk’s partnership with Sam Altman when founding OpenAI, and by loudly doing what Altman refused to do, Musk aims to undermine the market leader’s integrity. Open-source models aren’t entirely virtuous, however, for some experts have warned that malicious users could use open-sourced models to “create chemical weapons or even develop a conscious super-intelligence beyond human control.” Musk has previously advocated for a third-party “referee” to monitor the use of open-source models and sound the regulatory alarm in the event of adversarial use.
Read more about…
-
…turmoil at the top: Sam Altman to return to OpenAI’s board along with 3 new members
-
…who leads the charge on robotics & AI: NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots
AI at Work: Novel Uses, Recommendations and Impact on Labor
Hype is one thing, lying is another
According to a recent Wall Street Journal article, the Securities and Exchange Commission (SEC) has fined two investment advisers, Delphia and Global Predictions, for allegedly making false statements about their use of artificial intelligence (AI) technology, a practice known as "AI washing." The SEC's enforcement actions follow warnings from SEC Chair Gary Gensler against AI washing, comparing it to greenwashing. Delphia was accused of misleading the public about its use of AI and machine learning, falsely claiming to use client data in its investment strategy. Global Predictions was accused of making false statements about its use of AI and failing to substantiate performance claims and disclose conflicts of interest. As reported by a recent Crunchbase article, this scrutiny comes amidst a surge in private market investment and high-profile collapses like FTX and Theranos, prompting closer regulatory oversight. In the tech world, AI remains a hot topic, especially in healthcare, where startups like Zephyr AI, Carlsmed, and Hippocratic AI are attracting significant funding for AI-driven solutions.
AI Skills Gap between Businesses
A recent New York Times article highlights the AI skill disparity between leading AI companies like Microsoft and other organizations is becoming increasingly pronounced. While industry giants such as Microsoft are making substantial investments in AI talent acquisition and strategic partnerships, smaller and less tech-focused companies are struggling to keep pace. For instance, Microsoft recently hired a former Google executive, Mustafa Suleyman, to run its consumer AI unit. Microsoft's aggressive hiring of AI pioneers and its heavy focus on deal-making demonstrate a clear commitment to maintaining its position at the forefront of artificial intelligence innovation. According to a recent article from the Society of Human Resources Management, many businesses lack the resources or expertise to implement comprehensive AI training programs, leaving their workforce ill-equipped to harness the potential of AI tools effectively. As a result, there is a growing gap in AI literacy and skills between companies like Microsoft and the broader business community, posing challenges to competitiveness and innovation in the evolving digital economy.
Read more about…
-
…how AI is disrupting the entertainment industry: Big Changes Are Coming To Movies And TV Because Of AI
-
…AI’s shortcomings in journalism: I Used ChatGPT as a Reporting Assistant. It Didn’t Go Well
AI in the World: Shaping Lifestyles and Society
The AI Climate Dilemma: Balancing Technology with Environmental Reality
Environmental groups warn that attempts by AI to help with the climate crisis are misleading. Major tech companies have tried to utilize AI to ameliorate global warming through tools that predict droughts in Africa, track melting icebergs, and more.
Google’s report last year with BCG “found (that) AI could cut global emissions by as much as 10%.” In an interview with WSJ, Google’s Chief Sustainability Officer, Kate Brandt also notes that AI has the potential to give individuals more information on how to “reduce your carbon footprint and better understand sustainability data.” However, a recent report by Grid Strategies warns that AI's immense electricity demands could lead to the need for doubling data centers to match the industry’s quick pace. According to The Guardian, this may lead to a shocking 80% increase in planet-heating emissions. A separate study also notes that in approximately three years, AI servers may match the entire country of Sweden's energy consumption. The environmental groups’ warnings serve as a reminder of the need for sustainable growth in the tech industry. By highlighting the environmental impact of AI, these warnings could also influence public opinion and individual behavior, suggesting the need for mindful consumption and dissemination of digital information.
The Madness of AI Predicting (March) Madness
Fans are using AI to build their models using past annual NCAA tournament results, but AP News warns that "contests still provide plenty of surprises for computer science aficionados". Machine learning alone cannot holistically consider human aspects. According to Ezra Miller, a Duke Professor in Mathematics and Statistical Science, AI has many strengths in determining the likelihood of a team winning but notes that the “random choice of who's going to win a game that's evenly matched” remains, at its very core, random. AP News also interviewed a sports analytics major and Syracuse University student, Eugene Tulyagijja, who created his machine-learning model to identify success patterns from data such as a team's 3-point history. Tulyagijja stated that specific human elements limit computer powers, saying, “Personal things going on — we can never adjust to it using data alone.”
Moving forward, this is important because AI's application in bracketology and sports prediction emphasizes a significant shift in how technology seeps into our hobbies, entertainment, and decision-making process. The application of AI in sports reflects both the potential and limitations of what can be achieved through data and algorithms.
Resurrecting Language and Cultures with AI
Most AI models such as Claude, ChatGPT, and Gemini are primarily trained in English, leaving many bilingual and international users interested in examining whether AI models can operate in their native tongue. On March 5, X user (formerly known as Twitter) An Qu asked Claude 3 Opus to translate the sentence "I am lying in the bed" from Russian to Circassian. She reports that “Claude not only provided a perfect translation but also broke down the grammar & morphology.” Furthermore, An Qu tested for possible contamination. She used the same prompts without the sample translations and found that Claude failed and refused to answer, claiming that it was unfamiliar with Circassian. According to An Qu, completing the tasks she assigned to Claude requires a thorough understanding of the language, and given the same inputs it “would take a linguist, unfamiliar with the language, a good year or so to achieve.
On March 6, An Qu provided an update, stating, “I realized that my initial claim that Claude Opus does not know Circassian was incorrect. The model is, in fact, capable of translating to and from Circassian and even conversing in the language, even though with some grammatical flaws.” Nonetheless, she says that she is “still incredibly impressed by the quality of Claude's Circassian translations and analysis.” Claude 3 Opus shows that models with additional data boost their performance and that Anthropic has made great strides in training their dataset to include a range of languages, including less-known languages such as Circassian.
Read more about…
-
…AI and Olympians: How AI is changing gymnastics judging
-
…academic integrity and AI-powered chatbots: Many BU Students Study with ChatGPT. A Few Admit Cheating with It
Taming AI: Ethics, Policies and Regulations
Groundbreaking Regulations Abroad
On March 13, 2024, the long-awaited Artificial Intelligence Act (AI Act) was approved in the European Union. Originally proposed almost three years ago, this law is unique in that it bans uses of AI that tech leaders–like the US–have specifically leaned into over the past few years.
For example, the most notable case revolves around biometric data, such as fingerprints or facial recognition. According to the European Parliament, “scraping facial images from the internet or CCTV footage to create facial recognition databases” is banned. Most of the current facial recognition and image generation systems have been crafted with these methods. This law will likely change how these face-based systems are created in the future, as well as how biometric data collected in projects like the US airport’s new identification system is handled. In other examples, the AI Act bans profiling, and predictive policing with AI, which the US has researched over the past few years, as well as AI that manipulates human behavior, which could be said is equivalent to AI-generated political attack ads or robocalls. The EU is currently a leader in AI regulations; will other prominent tech countries follow suit?
Election: US Updates and International Comparisons
Last month, we shared that Big AI companies are working to make sure that chatbots give unbiased voter information to users. Their progress was then tested by Proof News, revealing that these chatbots are not yet prepared. OpenAI’s ChatGPT-2, Meta’s Llama 2, Google’s Gemini, Anthropic’s Claude, and Mistral’s Mixtral all failed simple questions about the US election process, providing harmful, biased misinformation 40% of the time. In one case, Llama even suggested that there was a Vote by Text option in California. To temporarily fix the issue, Google has taken the approach of blocking users from asking Gemini election-related questions in any country that currently has a major election. Currently, it is being rolled out in the US and India. While this does not solve the problem, it may prevent misinformation from spreading and harming voters.
This text-based scenario increases the risks of voter suppression in the US, but AI is also complicating elections with video, audio, and image on an international scale. For example, political parties in India have positively turned to AI for campaign message videos, in what Aljazeera describes as a “Deepfake democracy”. Countries Argentina, the US, and New Zealand have also started using generative AI for political attack ads, and Pakistan's former Prime Minister was disqualified and jailed after he was falsely portrayed by a deepfake. This new method of creating political ads may be quick and effective, but it can also cause mistrust and unintended consequences.
Even CEOs notice AI chatbot bias, but are their guardrails effective?
Relatively soon after the initial public release of image-based generative AI platforms, people began to notice bias within the system, such as a lack of racial diversity in generated images. Research on this racial and gendered bias was published, and calls were made to Big Tech to improve their algorithms and safeguards.
However, Google seemed to ‘overcorrect’ Gemini in that it began to produce historically inaccurate images and block requests for images of white people. Representation is important in generated content, but a more nuanced approach is needed to be both inclusive and accurate.
There has also been a trend in the text-based space of adding additional guardrails. For example, Microsoft has begun trying to block prompts that produce violent, sexual images. However, in their process, they blocked both the words “pro-life” and “pro-choice”. While these two words could lead to harmful content, they also could provide information and understanding in a variety of contexts. As tech companies continue to adjust guardrails, they must find a balance for effective and safe AI tools.
Read more about…
-
…updates to copyright law: Generative AI Is Challenging a 234-Year-Old Law
-
…how the US helped shape UN protocol: United Nations adopts U.S.-led resolution to safely develop AI
Research Revelations
AI and Video
As you recall from last month’s edition, Sora’s shock to the market has pushed the boundaries of creation, with high-definition videos that could fool anyone. This level of realism is unprecedented, and thus companies have begun to explore new avenues with Text-to-Video AI as their vehicle. For instance, OpenAI believes their Sora model “suggests that scaling video generation models is a promising path towards building general purpose simulators of the physical world.” Additionally, Sora could have applications in the 3-D modeling landscape, potentially becoming a challenger to current 3-D animators. And recently, this idea does not seem too far-fetched anymore, for within one week, Google and Microsoft, accomplished two grand achievements: the creation of Text to Games and 1-Bit LLMs, respectively. Google’s DeepMind unveiled Genie, a generative model that creates playable 2D video games from a text description, a sketch, or a photo. Genie’s ability to learn fine-grained controls while being trained solely on videos sets it apart in the market. The model does not only learn the actions from the video sequence, but it also creates variations of said actions that might apply to the same environment. On the other hand, a team of researchers from Microsoft and the University of Chinese Academy of Sciences has proposed an architecture called BitNet, which uses an extreme form of quantization called a 1-bit model, as a way to improve cost efficiency without sacrificing performance. This has large implications for generative AI, since current LLMs use 16 bits, whereas this new architecture would perform the same while only using 1.58 bits.
AI in the Physical World?
As we delve deeper into AI and tech, we have started to see AI applications that might spill over into the physical world. Besides OpenAI’s Figure01, a new company, Covariant, is creating ways for robots to pick up, move, and sort items as they are shuttled through warehouses and distribution centers. Their focus, however, is not creating robots, but rather helping robots gain an understanding of what is going on around them and decide what they should do next. Their focus is creating software that powers intelligent robots, such that they become capable of performing tasks autonomously.
Another application of AI is found in NVIDIA’s Isaac Sim, where the latest generative AI and advanced simulation technologies accelerate AI-enabled robotics. Isaac Sim, at its core, is a collection of foundation models, robotic tools, and GPU-accelerated libraries, with a plethora of applications- from accelerating learning for robots with advanced locomotion skills to fostering collaboration with the Open Source Robotics Alliance (OSRA).
***all imagery created using Image Creator from Designer***
The New AI Project | University of Notre Dame
Editor: Graham Wolfe
Contributors: Grace Hatfield, Rachel Lee, Cecilia Ignacio, Alejandro Velasco
Advisor: John Behrens
Archive: