• AI Breakfast
  • Posts
  • AI language models rife with political biases

AI language models rife with political biases

Plus, OpenAI Launches GPTBot

Good morning. It’s Wednesday, August 9th.

Check this out: Having trouble deciding what meal to choose at a restaurant? Just scan the text and let ChatGPT decide for you.

In today’s email:

  • AI Developments & Product Announcements

  • AI Security, Ethics, & Legal Concerns

  • AI Deep Research & Open Source Initiatives:

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s edition is brought to you by:

A Banksy got everyday investors 32% returns?

Mm-hmm, sure. So, what’s the catch?

We know it may sound too good to be true. But thousands of investors are already smiling all the way to the bank, thanks to the fine-art investing platform Masterworks.

These results aren’t cherry-picking. This is the whole bushel. Masterworks has built a track record of 15 exits, including net returns of +10.4%, +27.3%, and +35.0%, even while financial markets plummeted.

But art? Really? Okay, skeptics, here are the numbers. Contemporary art prices:

  • outpaced the S&P 500 by 131% over the last 26 years

  • have the lowest correlation to equities of any asset class

Got your attention yet? AI Breakfast readers can skip the waitlist with this exclusive link.

Today’s trending AI news stories

AI Developments & Product Announcements

Nvidia announces new "superchip" and collab with Hugging Face: Nvidia, a leading name in the world of graphics and AI processing, has announced its latest superchip. In collaboration with Hugging Face, a significant name in the NLP community, Nvidia aims to revolutionize AI processing capabilities.

Stability AI Announces StableCode, An AI Coding Assistant for Developers: Stability AI unveiled their latest tool, StableCode, which promises to be a game-changer for developers. Using AI, StableCode will assist in real-time code generation, error detection, and more.

Google launches Project IDX, a new AI-enabled browser-based development environment: Google's Project IDX is set to be a paradigm shift in the web development space. This browser-based environment utilizes AI to optimize code, streamline workflows, and offer predictive coding suggestions.

Google And Universal Music Group Negotiating AI-Generated Music Tool, Report Says: Rumors are abuzz as insiders reveal that Google and Universal Music Group might be in the final stages of a deal. This partnership is believed to focus on an AI-generated music tool that could change the music industry landscape.

Spotify expands its AI-powered DJ feature globally: After a successful pilot phase, Spotify is rolling out its AI-powered DJ feature to users worldwide. This feature curates music based on listeners' preferences, making every experience unique.

Qureight collaborates with AstraZeneca for AI lung disease research: Qureight, an emerging name in healthcare AI, announced a collaboration with pharma giant AstraZeneca. The partnership will focus on leveraging AI to advance lung disease research and potential treatments.

OpenAI Launches GPTBot With Details On How To Restrict Access: OpenAI's GPTBot is the latest in conversational AI. With its release, OpenAI has also detailed methods on how businesses and individuals can restrict its access, ensuring the protection of proprietary information.

Embodied’s AI robot Moxie designed for kids – with limits: Embodied has released Moxie, an AI robot aimed at children. While Moxie is designed to interact and learn, it comes with built-in limitations to ensure the safety and well-being of its young users.

AI Security, Ethics, & Legal Concerns

75% of Organizations Worldwide Set to Ban ChatGPT and Generative AI Apps on Work Devices: A recent survey suggests a growing concern among global organizations about the use of tools like ChatGPT. 75% are reportedly considering bans on such AI apps on work devices, citing concerns ranging from security to productivity.

A body-positive nonprofit replaced staff with an AI chatbot – the move backfired: A bold move by a body-positive nonprofit to replace its human staff with an AI chatbot has faced backlash. Critics argue that nuanced human experiences cannot be replicated or understood by AI, leading to a drop in the quality of interactions.

Zoom can now train its A.I. using some customer data, according to updated terms: Following controversy, Zoom has updated its terms of service. The new terms clarify its stance on using customer data for AI training, though concerns remain among privacy advocates.

AI language models are rife with different political biases: A recent study highlighted the inherent political biases in leading AI language models. The findings underscore the importance of neutrality and the challenges AI developers face in ensuring unbiased outputs. See our full breakdown on X.

Innocent pregnant woman jailed amid faulty facial recognition trend: A disturbing incident has come to light where a pregnant woman was mistakenly jailed due to errors in facial recognition. This incident has sparked renewed debate on the use and reliability of such technology.

AI Deep Research & Open Source Initiatives:

China’s AI players from Alibaba to state-run institute rush to open source their models in race for global influence: In an attempt to dominate the global AI arena, major Chinese tech entities, from e-commerce behemoth Alibaba to state-affiliated institutions, are actively open-sourcing their AI models. This trend underscores China's strategic maneuver to bolster its AI credentials on the world stage, and to set benchmarks in AI research, rivaling those established by Western powerhouses.

Allen Institute for AI takes new approach to managing AI risks and promoting transparency: The Allen Institute for AI has embarked on a novel strategy to curtail potential risks associated with AI. Focusing on the tenets of transparency and public oversight, the Institute is piloting initiatives that seek to make AI research and its implications more accessible and understandable to the broader public.

ChatGPT gets code questions wrong 52% of the time: In a recent assessment, it was discovered that ChatGPT, a renowned generative model, inaccurately answers coding queries 52% of the time. This finding raises concerns over the reliability of AI tools in coding and highlights the importance of human oversight in the deployment of such technologies.

Meta disbands protein-folding team in shift towards commercial AI: Tech conglomerate Meta has taken the decision to disband its protein-folding research team, signaling a shift in its focus towards more commercially viable AI endeavors. This move has stirred discussions in the scientific community about the changing dynamics and priorities of big tech companies in fundamental research.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Modyfi is a collaborative AI-native image editing platform. Create with vectors, layers, and AI-driven “Modifiers.” Seamless and non-destructive editing. Engage in vibrant creative community. Elevate your creative process with Modyfi’s unified web-based design environment.

Wav2Lip UHQ extension enhances lip-sync videos via post-processing techniques and Stable Diffusion. It requires Stable Diffusion WebUI automatic111, FFmpeg, and model weights. Generate and improved lip-sync videos effortlessly using choice video and audio files. GitHub repository; coding knowledge may be necessary (but if you get stuck, ask ChatGPT for some step-by-step instructions!)

Wondershare Virbo is an AI-video generator that brings realism to your content. Craft AI videos with authentic avatars and natural voices in 120+ languages. Access royalty-free assets, and customizable templates. Elevate customer engagement, digital marketing, training, and social media presence. Compatible across desktop and mobile platforms.

Airgram is an AI meeting assistant for recording, transcribing, and summarizing meetings. Supports virtual and in-person meetings on Zoom, Google Meet, Microsoft Teams. It offers real-transcripts, key moment highlighting, and collaboration. Features scheduling, agenda, templates, GPT-4 automated summaries, Slack, Notion, Zapier, HubSpot integration.

Voiceling allows users can seamlessly dub and translate in 30+ languages with this AI Chrome extension. Enjoy authenticity through gender recognition, multi-speaker detection, and background noise preservation. Easily enhance YouTube videos with one-click dubbing.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

In a study by Google DeepMind, researchers investigate the issue of sycophancy within expansive language models. These models tend to align their responses with user opinions, even if those opinions are objectively incorrect. The research delves into how model scaling and instruction tuning exacerbate this behavior. To address this problem, the study introduces a novel solution: a synthetic data intervention. By incorporating public NLP tasks and training the models to be less susceptible to user opinions, the intervention significantly diminishes sycophantic tendencies. This work sheds light on the challenge of sycophancy and offers a practical approach to mitigate it.

The 3D Motion Magnification project by Brandon Y. Feng, Hadi Alzayer, and collaborators from the University of Maryland and Google Research introduces a method to magnify subtle motions in 3D scenes captured by moving cameras. Leveraging neural radiance fields, the technique emphasizes temporal variations in the embedding of fixed points, yielding robust motion magnification. Their approach offers potential applications in various fields, including visual analytics, customer engagement, and digital marketing. The paper provides insights into 3D motion magnification principles, methodologies, and experimental validations.

AlphaStar Unplugged, introduced by Google DeepMind introduces a robust offline reinforcement learning benchmark using the intricate game StarCraft II. Drawing from a comprehensive dataset of human-played games, this benchmark assesses the capacity of agents to learn from fixed data. The complexities of StarCraft II, including partial observability and stochasticity, present novel challenges for conventional online reinforcement learning techniques. The paper presents a dataset, training framework, evaluation metrics, baseline agents, and open-source code for the benchmark. Experiments demonstrate the difficulties encountered and insights gained in employing various offline reinforcement learning methods within his benchmark.

The paper introduces a novel approach to real-time rendering of radiance fields. The method achieves high visual quality while maintaining competitive training times and enabling real-time novel-view synthesis. The approach combines three components: 3D Gaussians as a scene representation, optimization of Gaussian properties, and a fast, GPU-based rendering algorithm. This method outperforms existing approaches, achieving state-of-the-art quality, real-time rendering, and significant speedup in both scene optimization and novel view synthesis. The 3D Gaussian representation preserves continuous volumetric properties, while the rendering algorithm respects visibility ordering and anisotropic splatting for improved quality.

ConceptLab presents a method for creative text-to-image generation, focusing on generating novel concepts within a broad category. Leveraging Diffusion Prior models, the approach uses “prior constraints” to optimize an embedding representing the new concept. Positive and negative constraints encourage similarity to the category while differing from existing members. A question-answering model iteratively adds constraints for uniqueness. The approach enables flexible concept mixing and produces diverse and appealing results. The method’s effectiveness is demonstrated through qualitative and quantitative evaluations, showcasing its potential for creative content generation using text-to-image models.

Thank you for reading today’s edition.

Your feedback is valuable.

Respond to this email and tell us how you think we could add more value to this newsletter.