Last Week in AI #259: Google's Gemini AI controversy š, Google's open-sources Gemma models š, Moonshot AI's billion-dollar boost š, and more!
Google apologizes for āmissing the markā after Gemini generated racially diverse Nazis, Google Delves Into Open Source with Launch of Gemma AI Model, Chinese start-up Moonshot AI raises US$1 billion
Top News
Google apologizes for āmissing the markā after Gemini generated racially diverse Nazis
Google has issued an apology for inaccuracies in historical image generation by its Gemini AI tool, following criticism that it inaccurately depicted white historical figures and groups as people of color. The controversy arose from the tool's tendency to generate images of people of color when prompted to generate images of specific white figures or groups, such as the US Founding Fathers or Nazi-era German soldiers. Many users were able to generate outputs such as the above, leading to Google posting the following statement on X:
Soon after, Google paused the ability to generate humans via the tool and released the blog post āGemini image generation got it wrong. We'll do better.ā In it, they state:
When we built this feature in Gemini, we tuned it to ensure it doesnāt fall into some of the traps weāve seen in the past with image generation technology ā such as creating violent or sexually explicit images, or depictions of real people. And because our users come from all over the world, we want it to work well for everyone.
ā¦
However, if you prompt Gemini for images of a specific type of person ā such as āa Black teacher in a classroom,ā or āa white veterinarian with a dogā ā or people in particular cultural or historical contexts, you should absolutely get a response that accurately reflects what you ask for.
So what went wrong? In short, two things. First, our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range. And second, over time, the model became way more cautious than we intended and refused to answer certain prompts entirely ā wrongly interpreting some very anodyne prompts as sensitive.
These two things led the model to overcompensate in some cases, and be over-conservative in others, leading to images that were embarrassing and wrong.
Google Delves Deeper Into Open Source with Launch of Gemma AI Model
Google introduces Gemma, a new generation of lightweight, state-of-the-art open models designed for developers and researchers to build AI responsibly. Derived from the same technology as the Gemini models and developed by Google DeepMind and other Google teams, Gemma models come in two sizesā2B and 7Bāwith pre-trained and instruction-tuned variants. Google also launches a Responsible Generative AI Toolkit and offers comprehensive support across major frameworks like JAX, PyTorch, and TensorFlow. Gemma models, optimized for performance across various AI hardware platforms, including NVIDIA GPUs and Google Cloud TPUs, are designed for safe and reliable use. The code is released under a permissive Apache license, and minimal terms of use for the model weights.
Chinese start-up Moonshot AI raises US$1 billion in funding round led by Alibaba and VC HongShan amid strong interest for OpenAI-type firms
Chinese artificial intelligence start-up, Moonshot AI, has successfully raised over US$1 billion in a funding round led by Alibaba Group Holding and venture capital firm HongShan. This marks the largest single financing raised by a Chinese AI start-up since the release of ChatGPT in November 2022. Moonshot AI, founded by Yang Zhilin, a Tsinghua University graduate, launched its smart chatbot Kimi Chat in October, which is built on its self-developed Moonshot large language model (LLM) capable of processing up to 200,000 Chinese characters in a context window. The funding round underscores the continued strong interest in generative AI start-ups in mainland China, which led global investments into such firms in the first half of 2023.
Promotion
Check out this new book by Stanford AI expert, bestselling author, and Last Week in AI supporter Jerry Kaplan!
Generative Artificial Intelligence: What Everyone Needs to Know
Other News
Tools
Adobe Acrobat adds generative AI to āeasily chat with documentsā - Adobe Acrobat introduces a new generative AI experience called AI Assistant, which allows users to easily chat with documents, summarizing content, answering questions, and recommending more based on the content.
Introducing Phind-70B ā closing the code quality gap with GPT-4 Turbo while running 4x faster - Introducing Phind-70B, a high-performing AI model that closes the code quality gap with GPT-4 Turbo while running 4x faster, offering a better user experience for developers and exceeding GPT-4 Turbo on some tasks.
Google Chromeās āHelp me writeā tool can now finish your sentences for you - Google Chrome's new "Help me write" feature, powered by generative AI, assists users in writing and refining text based on webpage content, providing writing suggestions for shortform content and enabling users to adjust length and tone.
Samsungās āTry Galaxyā app adds Galaxy AI demo - Samsung's 'Try Galaxy' app, now available on all Android devices, introduces new AI features such as Live Translate, Note Assist, Chat Assist, Photo Assist, and Circle to Search with Google, as well as tutorials on advanced camera tools and other updates for the Galaxy S24 series.
Stability announces Stable Diffusion 3, a next-gen AI image generator - Stability AI announces Stable Diffusion 3, a next-gen AI image generator that reportedly produces detailed, multi-subject images with improved quality and accuracy in text generation, and will be available for free download and local use once testing is complete.
Microsoft releases its internal generative AI red teaming tool to the public - Microsoft has released a new tool, PyRIT, to help identify risks in generative AI systems, aiming to mitigate issues such as rogue behavior and loopholes that malicious actors can exploit.
Business
Groq AI model goes viral and rivals ChatGPT, challenges Elon Muskās Grok - Groq AI model, with its LPU Inference Engine, challenges ChatGPT with its lightning-fast response speed and new technology, potentially offering a game-changing alternative to GPU-based models.
Inside the Funding Frenzy at Anthropic, One of A.I.ās Hottest Start-Ups - Anthropic, an AI start-up, has experienced an astonishing funding spree, raising a total of $7.3 billion in a year from various investors including Google, Salesforce, Amazon, and others.
Nvidia posts record revenue up 265% on booming AI business - Nvidia's record-breaking revenue and earnings, driven by strong demand for AI chips, exceeded Wall Street's expectations and are expected to continue growing in the future.
Nvidia Says Growth Will Continue as A.I. Hits āTipping Pointā - Nvidia's quarterly financial results show its significant growth in the AI industry, with demand for its products driving continued sales growth and contributing to its surge in valuation.
Recogni Raises $102 Million to Meet AI Applicationsā Compute Demand - Recogni secures $102 million in funding to develop next-generation AI inference solutions, aiming to boost performance and power efficiency while addressing the growing compute demand for AI applications.
Google brings Gemini AI models to enterprise tools - Google is introducing its "Gemini" AI models to enterprise tools, offering them at a lower-priced plan to compete with Microsoft-backed OpenAI.
Google DeepMind forms a new org focused on AI safety - Google DeepMind has formed a new organization, AI Safety and Alignment, to address concerns about the potential misuse and safety of its GenAI models, with a focus on preventing disinformation, bias amplification, and ensuring child safety.
Reddit Inks $60 Million AI Content Licensing Agreement with Google - Reddit has signed a $60 million AI content licensing agreement with Google, providing the tech giant with access to user-generated content to train AI models, as Reddit prepares for its IPO and tech giants face backlash over data collection practices.
Jeff Bezos and Nvidia join OpenAI and Microsoft in backing a humanoid robot unicorn valued at $2 billion, sources say - Big technology names like Jeff Bezos and Nvidia are investing in a startup developing human-like robots, aiming to apply cutting-edge technology to real-world tasks and alleviate labor shortages.
Mistral AI models coming soon to Amazon Bedrock - Mistral AI, a French AI company, is bringing high-performing language models to Amazon Bedrock, offering fast, secure, and cost-effective options for text generation and code completion.
GMās Cruise Prepares to Resume Robotaxi Testing After Halt - Cruise is nearing the resumption of robotaxi testing, with Houston and Dallas emerging as potential locations, according to people with knowledge of the matter.
Tyler Perry Puts $800M Studio Expansion on Hold After Seeing OpenAIās Sora: āJobs Are Going to Be Lostā - Tyler Perry puts $800M studio expansion on hold after seeing OpenAIās Sora, expressing concerns about AI's impact on jobs and calling for industry-wide regulations and protection for workers.
Waymoās application to expand California robotaxi operations paused by regulators - Waymo's application to expand its robotaxi service in Los Angeles and San Mateo counties has been suspended for 120 days by the California Public Utilities Commission's Consumer Protection and Enforcement Division, putting a halt to the company's aspirations to expand its operations.
AI influencers are making their secretive creators tens of thousands of dollars a month - AI influencers, created on image-generating websites, are making tens of thousands of dollars a month for their secretive creators, who market them on social media and provide exclusive content to paying subscribers.
Generative AI Startup Mistral Releases Free āOpen-Sourceā 7.3B Parameter LLM - Mistral AI has quietly released a new 7.3B parameter LLM model named Mistral Next, which is currently available through the Direct Chat tab on the Large Model Systems Organization (LMSYS) page.
Research
Avoiding fusion plasma tearing instability with deep reinforcement learning - AI is used to develop a tearing-avoidance controller for fusion plasma in a tokamak reactor, allowing for stable and efficient fusion energy production by maintaining high-pressure hydrogenic plasma without disruption.
SDXL-Lightning: Progressive Adversarial Diffusion Distillation - A diffusion distillation method achieves state-of-the-art text-to-image generation based on SDXL, combining progressive and adversarial distillation for quality and mode coverage, with open-sourced models available.
VideoPrism: A Foundational Visual Encoder for Video Understanding - VideoPrism is a general-purpose video encoder that achieves state-of-the-art performance on various video understanding tasks by leveraging a pretraining approach on a large and diverse video-caption corpus.
Neural Network Diffusion - Diffusion models can generate high-performing neural network parameters using an autoencoder and a standard latent diffusion model, consistently producing models of comparable or improved performance over trained networks.
AnyGPT: Unified Multimodal LLM with Discrete Sequence Modeling - Unified multimodal language model AnyGPT can process various modalities including speech, text, images, and music seamlessly, using discrete representations and achieving performance comparable to specialized models.
Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs - Revisiting the use of REINFORCE-style optimization for learning from human feedback in large language models, the article argues that this method outperforms more complex alternatives like Proximal Policy Optimization and newly proposed methods, offering a more practical and cost-effective approach.
Concerns
ChatGPT spat out gibberish for many users overnight before OpenAI fixed it - ChatGPT users experienced odd responses, including gibberish and language switching, prompting OpenAI to investigate and fix the issue.
Gabās Racist AI Chatbots Have Been Instructed to Deny the Holocaust - Gab's AI chatbots, including versions of Adolf Hitler and Donald Trump, have been instructed to deny the Holocaust and spread misinformation on various controversial topics, raising concerns about the normalization and mainstreaming of disinformation narratives.
How much electricity does AI consume? - AI's energy consumption, particularly during training, is difficult to quantify due to lack of transparency from companies, but estimates suggest it could reach significant levels, potentially comprising a substantial portion of global electricity consumption by 2027.
Scientists Are Putting ChatGPT Brains Inside Robot Bodies. What Could Possibly Go Wrong? - Scientists are integrating ChatGPT, a large language model, into robot bodies to enhance their flexibility and problem-solving abilities, but concerns about biases, privacy violations, and the model's limitations persist.
Can AI porn be ethical? - AI-powered romance apps like MyPeach.ai are at the forefront of the ethical debate surrounding AI porn, implementing measures to prevent abuse and ensure consent while providing a simulated experience of a consensual relationship.
Impossible AI Food - AI-generated recipes without pictures can be misleading and include unheard of measurements and ingredients, causing confusion for users.
Policy
House leaders launch bipartisan artificial intelligence task force - House leaders launch bipartisan task force to explore AI regulation, innovation, and potential threats, appointing members with computer science backgrounds to develop guiding principles and policy proposals.
Explainers
Why Doesnāt My Model Work? - Common pitfalls in machine learning, including misleading data, data leakage, and inappropriate metrics, can cause models to fail when applied to real-world data, but these can be prevented by using checklists and tools to ensure the machine learning process is designed to support the study's aims and avoid mistakes.
A Visual Guide to Mamba and State Space Models - The article introduces the Mamba architecture, a selective State Space Model that aims to address the limitations of traditional State Space Models and compete with Transformer models by selectively compressing information, using a hardware-aware algorithm, and achieving fast inference and training.