- ๊ญฒ๊ฎ๊ญผ ๐ฆ๐ฌ๐ ๐๐๐ข๐ฃ๐ฅ๐๐ก๐๐จ๐ฅ
- Posts
- Weekly AI report
Read time: 7 minutes | Sponsor this newsletter
Welcome to the weekly AI report
In this week's AI news:
Nightshade - new data poisoning tool
Eureka! A Leap in Robotic Learning by NVIDIA Research
LM Studio: The One-Stop Shop for Local Large Language Models
Lenovo and NVIDIA Announce Hybrid AI Solutions to Help Enterprises Quickly Adopt GenAI
Frontier Risk and Preparedness by OpenAI
Frontier Model Forum Updates by OpenAI, Anthropic, Google, and Microsoft
Jina AI Launches Second-Generation Text Embedding Model
LensGo - Free All-in-One Tool for AI-Powered Image and Video Generation
Leaked: Gemini and Stubbs Coming to Makersuite
Nightshade - new data poisoning tool
What is Nightshade?
Nightshade is a data-poisoning tool designed to disrupt AI training models that use artists' works without permission. By making invisible changes to pixels, Nightshade turns the AI's output into gibberish, effectively challenging the unauthorized use of copyrighted art.
It was developed by a team led by Ben Zhao, a professor at the University of Chicago, Nightshade alters images in subtle ways that are invisible to the human eye. If these altered images are used in an AI's training data, the results can become unpredictable.
Key points:
Targets generative AI models like DALL-E, Midjourney, and Stable Diffusion.
Developed by a team led by Ben Zhao, a professor at the University of Chicago.
Intended to deter AI companies from infringing on artists' copyrights.
Can be integrated into another tool called Glaze, which "masks" artists' personal style to prevent scraping.
Open source, enabling community contributions to make the tool more potent.
Significance:
Nightshade has the potential to tip the power balance back towards artists in the debate over intellectual property and data scraping. By making unauthorized data usage risky, it forces tech companies to reconsider their approach and could pave the way for more equitable practices in the use of creative works for AI training.
You can read more about it here.
Eureka! A Leap in Robotic Learning by NVIDIA Research
NVIDIA Research has developed Eureka, an AI agent that uses Large Language Models to automatically create reward algorithms, training robots to perform complex tasks with over 50% improvement compared to traditional methods.
Key Points:
Utilizes GPT-4 for creating reward algorithms.
Capable of training robots in nearly 30 complex tasks including pen-spinning and opening drawers.
Eureka's algorithms outperform human-written ones in more than 80% of tasks.
Incorporates human feedback for fine-tuning reward functions.
Enables efficient training through GPU-accelerated simulation in NVIDIA's Isaac Gym.
Significance:
Eureka signifies a monumental shift in robotic training, automating and refining a process that has traditionally been manual and error-prone. Its capabilities could have widespread implications for the future of AI and robotics, making it easier and more efficient to develop sophisticated robots for a variety of tasks.
You can read more about Eureka here.
LM Studio: The One-Stop Shop for Local Large Language Models
LM Studio is a desktop application that streamlines the process of discovering, downloading, and running local large language models (LLMs) on your computer, offering a sleek UI and efficiency.
Key Points:
Lightweight and multi-platform, compatible with both Windows and MacOS.
Provides a chatbot interface for direct interaction with downloaded models.
Allows users to start inference servers with a few clicks.
Offers model customization options and filtering based on compatibility and download rates.
Designed for more efficient usage compared to web-based text generation tools.
Detailed guidelines for choosing the right model based on system capabilities.
Significance:
In an era focused on the practical applications of AI, LM Studio provides an easy and efficient way for individuals to experiment with language models, opening up opportunities for customization and server setup.
You can download LM Studio here and start playing with LLMs on your computer.
Lenovo and NVIDIA Announce Hybrid AI Solutions to Help Enterprises Quickly Adopt GenAI
Lenovo and NVIDIA expand their partnership to deliver new hybrid AI solutions aimed at easing the adoption of generative AI across enterprises.
Key Points:
The partnership focuses on bringing AI computing to data creation points, from the edge to the cloud.
The new solutions support a hybrid cloud approach, allowing businesses to build AI models using NVIDIA AI Foundations cloud service and run them on on-prem Lenovo systems.
Lenovoโs ThinkSystem SR675 V3 server and ThinkStation PX workstation are optimized for running production AI with NVIDIA's latest hardware and software.
Businesses can securely implement AI faster through the NVIDIA NeMo framework, allowing customization of enterprise-grade large language models.
The solutions will also support the recently announced VMware Private AI Foundation with NVIDIA.
The Lenovo AI Professional Services Practice offers a wide range of services to help businesses deploy AI effectively and at scale.
Significance:
In a time when rapid adoption of AI technologies is crucial for competitive advantage, Lenovo and NVIDIAโs partnership offers enterprises a streamlined, flexible, and powerful path to implementing generative AI solutions.
You can read more here.
Frontier Risk and Preparedness by OpenAI
OpenAI is taking proactive steps to mitigate catastrophic risks from highly capable AI models by launching a Preparedness team and initiating a challenge aimed at misuse prevention.
Key Points:
OpenAI emphasizes its ongoing commitment to AI safety and security, having joined other leading labs in voluntary commitments focused on a range of risk areas, including frontier risks.
OpenAI's Preparedness team, led by Aleksander Madry, will be responsible for assessing the capabilities and risks associated with frontier AI models. Areas of concern include cybersecurity, chemical threats, and autonomous replication and adaptation.
The Preparedness team will operate under a Risk-Informed Development Policy (RDP), designed to create a robust framework for evaluating, monitoring, and governing frontier AI models.
OpenAI has launched an AI Preparedness Challenge to identify lesser-known risks, offering $25,000 in API credits for up to 10 top submissions.
Significance:
OpenAI's focus on preparedness signifies an important step in the AI industry, aiming to standardize safety protocols and ethical considerations, particularly as we edge closer to the realization of Artificial General Intelligence (AGI).
You can read more about Frontier risk and preparedness here.
And take the Preparedness Challenge here.
Frontier Model Forum Updates by OpenAI, Anthropic, Google, and Microsoft
OpenAI, Anthropic, Google, and Microsoft announce the appointment of Chris Meserole as the first Executive Director of the Frontier Model Forum and the establishment of a new $10 million AI Safety Fund aimed at advancing AI safety research.
Key Points:
Chris Meserole, an expert in technology policy, has been appointed as the Executive Director of the Frontier Model Forum. He will focus on advancing AI safety research, identifying best practices, and fostering responsible AI development.
The Frontier Model Forum, backed by leading tech companies and philanthropic foundations, commits to over $10 million in initial funding for a new AI Safety Fund.
The AI Safety Fund will support independent researchers and aims to advance new model evaluations and techniques for red teaming potentially dangerous frontier AI systems.
The Forum is also working on establishing common definitions and best practices on "red teaming" and responsible disclosure processes for AI safety and governance.
Future developments include the establishment of an Advisory Board, issuing the first call for proposals from the AI Safety Fund, and releasing additional technical findings.
Significance:
In a time when rapid advancements in AI are met with growing concerns about safety and governance, the Frontier Model Forum and its AI Safety Fund represent a critical collaboration to set industry standards, drive research, and involve a range of perspectives to mitigate risks.
You can find more about Frontier Model Forum here.
Jina AI Launches Second-Generation Text Embedding Model
Berlin-based Jina AI releases jina-embeddings-v2, an open-source text embedding model with a remarkable 8K (8192 tokens) context length, rivaling OpenAI's proprietary models in performance and capabilities.
Key Points:
Jina AI's new model outperforms OpenAI's 8K model in several categories, including Classification, Reranking, Retrieval, and Summarization Averages.
The model was built from scratch over the last three months through intensive R&D and data collection, marking a significant leap from its predecessor.
The 8K context length enables new applications in fields like legal document analysis, medical research, literary analysis, financial forecasting, and conversational AI.
Two versions are available on Huggingface: a Base Model (0.27G) for heavy-duty tasks and a Small Model (0.07G) for lightweight applications.
Jina AI plans to publish an academic paper detailing the model's benchmarks and is developing an OpenAI-like embeddings API platform.
The company is also looking to expand into multilingual embeddings, starting with German-English models.
Significance:
The launch of Jina-embeddings-v2 marks a milestone in democratizing high-performance text embedding models, offering the community an open-source alternative that can stand toe-to-toe with industry giants like OpenAI.
Read more about Jina here.
LensGo - Free All-in-One Tool for AI-Powered Image and Video Generation
LensGo is disrupting the AI video and image generation market by offering an all-in-one package that is not only versatile but also free to use.
Key Points:
LensGo is a comprehensive tool that offers functionalities for text-to-image, text-to-video, and even video-to-video conversions.
The platform features various customizable models, including "real," "anime," and "fantasy," for generating highly detailed and realistic outputs.
Unlike other platforms, LensGo is currently in beta and free to use, making it a cost-effective alternative to existing services like Runway ML.
Users have the option to choose aspect ratios, video durations, and even camera effects for their projects.
LensGo.ai is suitable for content creators, designers, and AI enthusiasts, offering a wide range of creative options.
Significance:
The emergence of LensGo as a free, all-in-one platform for image and video transformation comes at a pivotal time, challenging established players and democratizing access to advanced AI creative tools.
You can try LensGo here.
Leaked: Gemini and Stubbs Coming to Makersuite
Leaked information reveals that Google is bringing Gemini, a new multimodal AI model, and Stubbs, a tool for building AI-generated app prototypes, to Makersuite.
Stubbs UI
Key Points
Gemini, set to replace PaLM 2, will offer multimodal functionalities, including text and image recognition, to Makersuite.
Stubbs is a hitherto unknown feature allowing users to build, deploy, and share AI-generated app prototypes directly from Makersuite.
Full support for language translation, including a sample Spanish to English prompt, will be available in Makersuite.
Although leaked, the information shows a user-friendly interface for Stubbs and a gallery for publishing and viewing prototypes.
The features are in the leaked stage and may be subject to changes.
Current limitations include the absence of GIF support in Makersuite and no full code generation from Stubbs.
Significance
The leak hints at a potential paradigm shift in AI-powered development tools, poised to democratize app development and extend AI's multimodal capabilities.
Thank you for reading!
That is all for this week's Weekly AI report. If you liked this one, be sure to follow me on X and LinkedIn. Until the next Friday!