• Posts
  • AI language models are rife with political biases

AI language models are rife with political biases

Also: The Great 8-bit Debate of Artificial Intelligence


In today's newsletter, a study discovers inherent political biases in AI language models, with OpenAI's ChatGPT and GPT-4 labeled as the most left-wing libertarian. Meanwhile, Meta is harnessing AI to reinvigorate its flagging ad business. Over in Japan, a colossal fund targeting AI and biotech startups indicates a potent deep-tech revolution on the horizon. The AI community delves into a debate on the most suitable numerical representation for AI models, questioning the balance between efficiency and precision. Stack Overflow turns to AI to reposition itself as the go-to platform for programmers, and Google champions women-led AI startups in the Asia Pacific with a new initiative. Dive in for more on these stories.


  • 🔎 AI language models are rife with political biases

  • 📢 For Meta, the big AI play is shoring up its ad business

  • 🇯🇵 AI startups: Fuelling Japan's deep tech revolution

  • ⑧ The Great 8-bit Debate of Artificial Intelligence

A recent study conducted by the University of Washington, Carnegie Mellon University, and Xi’an Jiaotong University found that AI language models possess varying political biases. When tested, different models demonstrated divergent views on matters such as company social responsibilities. The study, which involved 14 language models, identified OpenAI’s ChatGPT and GPT-4 as the most left-wing libertarian, while Meta’s LLaMA was found to be the most right-wing authoritarian. By posing politically sensitive questions to these models and using their responses to map them on a political compass, researchers discerned distinctive political leanings. They also found that retraining models with biased data further intensified these biases and impacted their ability to detect hate speech and misinformation. The paper emphasized the significance of understanding these biases as AI models are incorporated into products used by millions, potentially leading to real-world harm. OpenAI, facing criticism for potential biases, has stated that any biases are unintentional. The research highlighted the inherent challenges in trying to remove biases and the need for transparency and awareness as AI continues to be integrated into various sectors.

Meta, having faced its first-ever annual revenue decline last year, is making significant investments in AI to bolster its advertising business. Previously, Meta warned investors about the potential negative impacts of Apple's privacy policy changes on its ad revenues. To counteract this, the company launched AI-driven tools like Advantage+ to improve ad campaign optimization. Now, nearly all of Meta's advertisers use its AI-powered ad tools, with the company reporting consecutive quarters of revenue growth. Despite the hype around its open-source AI model Llama, Meta's primary focus remains on AI-enhanced ads and content recommendations to increase user engagement. However, the company's long-term AI product strategy remains unclear. Soon, Meta plans to introduce AI agent personas aiming to increase user retention and advertiser engagement.

Miyako Capital, linked with Kyoto University, is establishing a 20 billion yen fund, targeting deep-tech biology and AI startups in Japan, in response to the growing APAC biotech market which is projected to reach US$318.8 billion by 2025. This fund aligns with the rising interest in biotech investments in the APAC region, driven by precision medicine advancements and technological breakthroughs in AI, machine learning, and big data analytics. Kyoto University’s fund is expected to rank among the largest in Japan, allocating between 500 million yen and 1 billion yen for each startup. The investment will be divided equally between biology and AI. Miyako Capital aims to collaborate with educational institutions to identify nationwide investment opportunities and will rely partly on state-backed funds. The firm also plans to support startups in their global ventures through its Silicon Valley office. The consistent backing from venture capital firms, government support, university collaborations, and a focus on cutting-edge sectors like AI, positions Japan for a thriving deep-tech startup scene.

The growth and proliferation of AI, especially in deep learning, have sparked debates around the optimal numerical representation for AI models. High Performance Computing (HPC) typically requires high precision, such as 64-bit double precision floating point. However, many AI applications, aiming for efficiency, can operate with lower precision like 8-bit integers or floating point numbers. The computational and environmental costs of AI have become concerns, leading to exploration in numerical representation as a means to reduce these costs. The debate centers around whether to use integer or floating point data types, with the industry currently favoring two quantized data types: INT8 and FP8. Quantization, which reduces the number of bits needed to represent the weights of a network, is a critical tool for efficiency. Two main methods, Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT), are used to achieve this reduction. The choice of integer versus floating point depends on the application, model, and underlying hardware. For instance, training often requires higher dynamic range and thus leans towards floating point, whereas inference emphasizes computational efficiency, potentially favoring integer types or FP8.

🛠️ AI tools updates

Stack Overflow, once the primary resource for programmers with queries, has seen a dip in traffic, prompting a shift towards integrating artificial intelligence (AI) into their platform with OverflowAI. This initiative will embed generative AI into both its public platform and Stack Overflow for Teams, aiming to directly infuse over 58 million community-driven Q&A sessions into developers' workspaces. By integrating OverflowAI into the Visual Studio Code IDE via an extension, developers can receive personalized solutions to programming challenges without having to leave their IDE, maintaining their workflow. While comparable tools like GitHub CoPilot are available, Stack Overflow differentiates by validating and ensuring accuracy through its vast community data. OverflowAI also upgrades its search functions, transitioning from lexical search to semantic search for more intuitive and accurate user queries. With added features like GenAI Stack Exchange and Stack Overflow's NLP Collective, Stack Overflow envisions a future where AI democratizes coding, leads to a surge in unique questions, and continues to position Stack Overflow as a primary hub for developers.

💵 Venture Capital updates

Google for Startups has introduced the Women Founders Fund to bolster female participation in AI in the Asia Pacific region, given that women currently represent a scant 5.7% of founders in this sector. This program, which aims to rectify gender imbalances and the prevalence of biased AI algorithms, will offer six women-led AI startups from India, Japan, and Korea an equity-free grant of $100,000 each. These countries have demonstrated growth in women-driven AI businesses. This initiative is an extension of Google's wider commitment to backing underrepresented founders, having already allocated over $30 million globally. Successful applicants will receive both the grant and specialized guidance from Google professionals. Applications remain open until August 15, 2023.

🫡 Meme of the day

⭐️ Generative AI image of the day