Top News
OpenAI reportedly nears breakthrough with “reasoning†AI, reveals progress framework
OpenAI has introduced a five-tier system to track its progress towards developing artificial general intelligence (AGI), a type of AI that can perform tasks like a human without specialized training. The levels range from current AI capabilities to systems that could potentially manage entire organizations. OpenAI’s technology, such as GPT-4o that powers ChatGPT, is currently at Level 1, which includes AI that can engage in conversational interactions. However, the company is reportedly close to reaching Level 2, or “Reasoners,” which would be capable of basic problem-solving on par with a human with a doctorate degree. Despite the introduction of this system, there is no consensus in the AI research community on how to measure progress towards AGI, and some view OpenAI’s five-tier system as a tool to attract investors rather than a scientific measurement of progress.
OpenAI Project Leaked: Strawberry, The Next Frontier of Reasoning For AI
OpenAI Develops System to Track Progress Toward Human-Level AI
A Hacker Stole OpenAI Secrets, Raising Fears That China Could, Too
In early 2022, a hacker infiltrated OpenAI’s internal messaging systems, stealing information about the design of the company’s AI technologies. The breach occurred in an online forum where employees discussed the latest technologies, but the hacker did not gain access to the systems where the AI is built and stored. The incident was disclosed to employees and the board of directors in April 2023, but was not made public as no customer or partner information was compromised. OpenAI executives did not perceive the incident as a national security threat, believing the hacker to be a private individual with no connections to a foreign government, and therefore did not report the incident to law enforcement.
Microsoft and Apple ditch OpenAI board seats amid regulatory scrutiny
Microsoft has relinquished its observer seat on the board of OpenAI, a move that comes less than eight months after it secured the non-voting position. Apple, which was reportedly planning to join OpenAI’s nonprofit board, has also decided against it. These changes occur amid growing antitrust concerns over Microsoft’s partnership with OpenAI, with regulators in the UK and EU scrutinizing the deal, along with other Big Tech AI investments. Despite this, OpenAI plans to continue its successful partnership with Microsoft and Apple through regular stakeholder meetings, aimed at fostering stronger collaboration across safety and security. Microsoft’s investment in OpenAI, which exceeds $10 billion, has made it the exclusive cloud partner for OpenAI, powering all its workloads across products, API services, and research.
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
FlashAttention is an important and widely used method for speeding up the inference of Large Language Models. This discusses FlashAttention-3, an improved method for speeding up attention on Hopper GPUs, the latest and best hardware for LLMs from Nvidia. The new method utilizes three main techniques: exploiting asynchrony of the Tensor Cores and TMA to overlap computation and data movement, interleaving block-wise matmul and softmax operations, and using block quantization and incoherent processing that leverages hardware support for FP8 low-precision. The results show that FlashAttention-3 achieves a speedup on H100 GPUs by 1.5-2.0 times with FP16 reaching up to 740 TFLOPs/s (75% utilization), and with FP8 reaching close to 1.2 PFLOPs/s.
Other News
Tools
Odyssey Building ‘Hollywood-Grade’ AI Text-to-Video Model to Compete With Sora, Gen-3 Alpha – Odyssey is developing an AI video model that can create Hollywood-grade visual effects and allow users to edit and control the output at a granular level, using multiple large language models to generate different layers of the output video.
Kuaishou’s text-to-video model Kling introduces new short video generation feature, results go viral in China – Kuaishou’s text-to-video model Kling AI, showcased at the World Artificial Intelligence Conference, has gone viral in China, generating AI videos based on simple prompts and challenging TikTok’s Douyin and ByteDance’s TikTok.
Anthropic Introduces Fine-Tuning for Claude 3 Haiku on Amazon Bedrock – Anthropic introduces fine-tuning capabilities for Claude 3 Haiku on Amazon Bedrock, allowing businesses to customize the model for specific tasks, leading to improved performance and increased control over AI training.
Anthropic’s Claude adds a prompt playground to quickly improve your AI apps – Anthropic’s Claude introduces a prompt playground to automate prompt engineering and improve AI applications, offering quick feedback and tools to test and evaluate prompts for better results.
Vimeo joins YouTube and TikTok in launching new AI content labels – Vimeo has implemented AI content labels to distinguish between real and AI-generated content, requiring creators to disclose when AI is used for realistic videos.
Google says Gemini AI is making its robots smarter – Google is using Gemini AI to train its robots for better navigation and task completion, allowing them to understand natural language instructions and achieve a 90 percent success rate in executing user commands.
Quora’s Poe now lets users create and share web apps – Quora’s Poe introduces Previews feature allowing users to create interactive apps directly in chats with AI-powered chatbots, supporting HTML output and multiple chatbots, but arrives amidst controversy over allowing users to download paywalled articles.
Bumble users can now report profiles that use AI-generated photos – Bumble introduces new reporting option to combat AI-generated profiles on its dating app, aiming to create a safer and more trustworthy environment for its users.
Figma pauses its new AI feature after Apple controversy – Figma temporarily disables its “Make Design” AI feature after criticism for mimicking Apple’s Weather app, while YouTube allows takedown requests for AI-generated content and Fisker seeks approval to sell its electric SUVs at a steep discount.
Etsy adds AI-generated item guidelines in new seller policy – Etsy introduces new guidelines for AI-generated items in its seller policy, requiring sellers to label products based on the level of human involvement and disclose if AI tools were used in the creation process.
Business
Figure 01: Coffee-making humanoid robot now shows car assembly skill at BMW – A humanoid robot developed by Figure is now being used in BMW’s car assembly process, showcasing the potential for increased automation in response to workforce scaling challenges.
A.I. Helped to Find a Vast Source of the Copper That A.I. Needs to Thrive – A.I. technology led to the discovery of a vast copper deposit in Zambia, potentially worth billions of dollars annually.
China’s AI competition deepens as SenseTime, Alibaba claim progress at AI show – Chinese AI companies SenseTime and Alibaba showcased their advancements in large language models (LLMs) at the World Artificial Intelligence Conference (WAIC) in Shanghai, with SenseTime claiming improved performance and Alibaba touting new user growth for its Tongyi Qianwen LLMs.
AMD plans to acquire Silo AI in $665 million deal – AMD plans to acquire Finnish AI company Silo AI in a $665 million deal, aiming to boost its position in the AI landscape with over 100 PhDs and 300 employees joining the company.
Robot-packed meals are coming to the frozen-food aisle – AI-powered robotic arms are revolutionizing the frozen food industry by accurately portioning out meals and reducing labor costs for companies like Amy’s Kitchen.
OpenAI and Arianna Huffington are working together on an ‘AI health coach’ – OpenAI and Arianna Huffington are collaborating on an “AI health coach” that aims to provide personalized health advice and guidance based on individual data, although there are concerns about privacy and the potential for misinformation.
AI Video Startup Captions Valued at USD 500M in USD 60M Series C – AI video editing startup Captions raises USD 60m in Series C funding, bringing its total funds to USD 100m, with a valuation of USD 500m, and plans to invest $100 million into advancing generative video research.
Tesla shares fall 6% after report of robotaxi unveiling delay – Tesla’s shares fell 6% after reports of a delay in unveiling its Robotaxi by two months, impacting the company’s stock performance and raising questions about its promises for autonomous vehicles.
Tech Startup Aims to Help Media License Content for AI Training – AI startup Avail launches Corpus, a product to help smaller media and entertainment companies and independent creators license their content to AI firms for model training.
Why The Atlantic signed a deal with OpenAI – The Atlantic’s CEO discusses the magazine’s deal with OpenAI, the value of AI in journalism, and the future of media in the digital age.
Perplexity planning revenue sharing program with web publishers next month – AI chatbot Perplexity will begin a revenue-sharing program with web publishers next month, the company announced during VB Transform Thursday.
Research
Memory^3: Language Modeling with Explicit Memory – A new language model, Memory^3, is equipped with explicit memory to reduce training and inference costs, achieving better performance than larger models and maintaining higher decoding speed.
MIT researchers introduce generative AI for databases – MIT researchers introduce GenSQL, a generative AI system for databases that enables users to perform complex statistical analyses, make predictions, detect anomalies, guess missing values, fix errors, and generate synthetic data with just a few keystrokes, providing faster and more accurate results compared to popular AI-based approaches.
Data curation via joint example selection further accelerates multimodal learning – Joint example selection for data curation accelerates multimodal learning, surpassing state-of-the-art models with significantly fewer iterations and less computation.
Just read twice: closing the recall gap for recurrent language models – Improving the recall gap for recurrent language models by addressing the challenge of information selection and proposing JRT-Prompt and JRT-RNN as solutions.
FunAudioLLM: Voice Understanding and Generation Foundation Models for Natural Interaction Between Humans and LLMs – FunAudioLLM introduces innovative models for enhancing natural voice interactions between humans and large language models, enabling applications such as speech-to-speech translation and emotional voice chat.
From Principles to Rules: A Regulatory Approach for Frontier AI – A regulatory approach for AI is proposed, emphasizing the importance of principles and rules to guide the development and use of frontier AI technologies.
PaliGemma: A versatile 3B VLM for transfer – PaliGemma is an open Vision-Language Model (VLM) based on the SigLIP-So400m vision encoder and the Gemma-2B language model, achieving strong performance on diverse tasks.
Vision language models are blind – Vision language models, such as GPT-4o and Gemini 1.5 Pro, are found to fail on basic visual tasks, indicating their poor performance in understanding visual information.
This&That: Language-Gesture Controlled Video Generation for Robot Planning – AI method This&That uses language-gesture conditioning to generate videos for robot planning, addressing challenges in task communication, video generation control, and translating visual planning into robot actions.
CodeUpdateArena: Benchmarking Knowledge Editing on API Updates – A benchmark called CodeUpdateArena is introduced to evaluate how large language models can update their knowledge about code API functions, highlighting the challenges and the need for new methods in knowledge editing for code LLMs.
WildGaussians: 3D Gaussian Splatting in the Wild – A new approach called WildGaussians is introduced to improve 3D Gaussian Splatting’s performance in handling in-the-wild data, achieving state-of-the-art results with real-time rendering speeds.
CopyBench: Measuring Literal and Non-Literal Reproduction of Copyright-Protected Text in Language Model Generation – The article discusses the CopyBench tool for measuring the reproduction of copyright-protected text in language model generation.
Concerns
OpenAI is plagued by safety concerns – OpenAI is facing safety concerns from employees and external sources, with claims of rushed safety tests, dissolved safety teams, and deprioritized safety culture, raising worries about the potential impact on society.
Tesla sells ‘Self-Driving’ cars. Is it fraud? – Tesla’s marketing of its “Full Self-Driving” and Autopilot features is under scrutiny by the U.S. Justice Department and California’s Department of Motor Vehicles, as well as facing civil lawsuits, over claims of potential fraud and misleading customers.
OpenAI Researcher Says He Quit When He Realized the Upsetting Truth – Former OpenAI worker quit due to the company prioritizing profit over safety in the pursuit of artificial general intelligence, likening it to the Titanic and expressing concerns over the lack of oversight and shifting corporate structure.
Tool preventing AI mimicry cracked; artists wonder what’s next – AI image generators are becoming better at replicating unique styles, prompting artists to seek defenses like Glaze, a tool that adds imperceptible noise to images to prevent mimicry, but its effectiveness is questioned as demand surges and security researchers claim it can be bypassed.
4chan Is Using TikTok’s Hidden AI App to Generate Porn – Users on 4chan have found a way to use TikTok’s hidden AI app to generate porn, prompting ByteDance to disable the AI-image generation capabilities despite the app’s policies and guardrails.
Policy
Senators introduce COPIED Act to push for better watermarking on AI content – Senators introduce COPIED Act to protect content from AI manipulation and require watermarking for authentication.
Japan’s Defense Ministry unveils first basic policy on use of AI – Japan’s Defense Ministry unveils its first basic policy on the use of AI to address manpower shortage and keep pace with global military technology advancements.
Analysis
Breaking Down What’s at Stake in Music’s AI Lawsuits – AI music lawsuits could shape the future of the music industry, as major labels sue AI firms for alleged copyright infringement, with potential implications for fair use and control over AI technology.
AI scaling myths – Bigger language models have shown improvement, but there are misconceptions about their future capabilities, as scaling laws do not guarantee continued emergence, and obtaining more high-quality training data may be challenging and costly.
How Good Is ChatGPT at Coding, Really? – AI code generator ChatGPT has a broad range of success in producing functional code, with better performance on older coding problems, but it lacks critical thinking skills and understanding of newer problems, leading to security concerns and the need for additional developer input.
Explainers
The Illustrated AlphaFold – A detailed visual walkthrough of AlphaFold3’s architecture, including its input preparation, representation learning, structure prediction, loss function, and other training details, as well as its similarities to recurrent architectures and trends in machine learning.
The making of Eno, the first generative feature film – Eno, the first generative feature film, is a documentary about musician Brian Eno, created using a proprietary generative software system that allows for a different version of the film to be shown each time, exploring Eno’s creative process and philosophy while also sparking discussions about the potential of generative filmmaking and AI technology.
Fun
The first Miss AI has been crowned — and she’s a Moroccan lifestyle influencer – Moroccan AI influencer Kenza Layli wins the inaugural Miss AI contest, expressing her commitment to promoting diversity and inclusivity within the field of AI technology.
Copyright © 2024 Skynet Today, All rights reserved.
Source: Read MoreÂ
