上周 AI 动态 #282 – 更多 OpenAI 戏剧，Stable Diffusion 创始人回归，以及更多！_AI阅读总结 — 包阅AI

OpenAI co-founder John Schulman has left the company to join rival AI startup Anthropic, while OpenAI president and co-founder Greg Brockman is taking an extended leave until the end of the year. Schulman, who played a key role in creating the AI-powered chatbot platform ChatGPT and led OpenAI’s alignment science efforts, stated his move was driven by a desire to focus more on AI alignment and hands-on technical work. Peter Deng, a product manager who joined OpenAI last year, has also left the company. With these departures, only three of OpenAI’s original 11 founders remain: CEO Sam Altman, Brockman, and Wojciech Zaremba, lead of language and code generation.

More on this:

Elon Musk has reinitiated a lawsuit against OpenAI, the creator of the AI chatbot ChatGPT, reigniting a longstanding dispute that originated from a power conflict within the San Francisco-based startup. The lawsuit, filed in a Northern California federal court, alleges that OpenAI and its co-founders, Sam Altman and Greg Brockman, violated the company’s founding agreement by prioritizing commercial interests over public welfare. Musk had previously withdrawn his initial lawsuit without providing a reason, just a day before a judge was due to decide on its dismissal. The lawsuit contends that Altman and Brockman deviated from their original commitment to develop artificial intelligence responsibly for the benefit of humanity when they entered into a multi-billion dollar partnership with Microsoft.

Black Forest Labs, a startup founded by the creators of Stable Diffusion, has launched FLUX.1, a new text-to-image model suite for the open-source artificial intelligence community. The company, led by Robin Rombach, Patrick Esser, and Andreas Blattmann, has secured $31 million in seed funding, with Andreessen Horowitz leading the investment round. FLUX.1 comes in three variants, all boasting 12 billion parameters and a hybrid architecture of multimodal and parallel diffusion transformer blocks. The launch of FLUX.1 is seen as a significant milestone in the democratization of powerful AI tools, with the potential to reshape the AI industry and influence the debate about open-source versus closed-source development models.

Perplexity AI, a startup that uses a chatbot to surface content in response to user queries, plans to share advertising revenue with news publishers whose content is used by the bot. This move comes in response to accusations of plagiarism and unethical web scraping. The company’s first publishing partners include Automattic, Der Spiegel, Entrepreneur, Fortune, The Texas Tribune, and TIME. These publishers will gain access to Perplexity’s APIs and developer support, enabling them to create custom answer engines on their sites. The revenue-sharing model will be implemented once Perplexity begins displaying ads on its platform in the coming months. The company has not disclosed the exact percentage of ad revenue to be shared but confirmed it would be in the “double digits”.

More on this:

Meta has launched a new tool called AI Studio, allowing users in the US to create AI versions of themselves on Instagram or the web. The tool is aimed at creators and business owners who can use these AI profiles to interact with their followers, respond to comments, and even auto-reply from their AI. The AI can be customized based on the user’s Instagram content, topics to avoid, and links to share. AI Studio also enables the creation of entirely new AI characters that can be used across Meta’s apps, competing with startups like Character.AI and Replika. Despite potential issues with AI versions of celebrities saying problematic things, Meta is pushing forward with this concept, which was initially tested with a few celebrities.

Microsoft is adding AI-powered summaries to Bing search results – Microsoft is introducing AI-powered summaries to Bing search results, providing original responses compiled by AI above the usual search hits, with previews currently available for a small percentage of user queries.

NIST releases a tool for testing AI model risk – NIST has released Dioptra, an open-source tool designed to measure how malicious attacks might degrade the performance of AI systems, aiming to help companies and users assess, analyze, and track AI risks.

Google releases new ‘open’ AI models with a focus on safety – Google releases new ‘open’ AI models with a focus on safety, including a lightweight text generator, safety classifiers, and a tool for making AI models more interpretable, in an effort to foster goodwill within the developer community and broaden generative AI’s availability.

Stability AI releases super-fast model for 3D asset image generation – Stability AI Ltd., the open-source generative artificial intelligence startup, has been on a roll, releasing new AI models that generate 3D assets from 2D images and capture motion in 3D.

Watch ChatGPT’s new voice mode sing, mimic accents, and correct pronunciation – OpenAI’s new advanced voice mode for ChatGPT impresses users with its ability to sing, mimic accents, and correct language pronunciation in various languages.

Runway just dropped image-to-video in Gen3 – Runway’s Gen-3 AI model, now equipped with image-to-video capability, offers improved character consistency and hyperrealism, allowing for motion or text prompts to steer the AI model’s video generation, making it a significant advancement in AI video tools.

Midjourney drops surprise v6.1 update — now humans look more real than ever – Midjourney’s surprise v6.1 update significantly improves the realism of human features and text rendering, making it a major upgrade despite its iterative numbering.

Meta AI Introduces Meta Segment Anything Model 2 (SAM 2): The First Unified Model for Segmenting Objects Across Images and Videos – Meta introduces SAM 2, a unified model for real-time object segmentation in images and videos, offering efficiency, versatility, and open-source collaboration.

AI-Powered Necklace Will Be Your Friend for $99 – An upcoming AI-powered necklace called friend, an always-listening pendant, aims to keep you company if you’re lonely by initiating conversations and responding via text, but requires tapping and reading responses on a smartphone.

AI chip startup Groq lands $640M to challenge Nvidia – Groq, a startup developing AI chips, has raised $640 million in funding, with plans to create faster and more energy-efficient chips for generative AI models, while facing competition from industry giants like Nvidia and other startups.

Chinese AI built off open-source code matches American tech in chatbot benchmark tests – Chinese AI, built off open-source code, matches American tech in chatbot benchmark tests, showing how Chinese companies are catching up and leading in AI development through open-source collaboration, despite hardware bans and potential U.S. regulations.

Sam Altman-Backed Rain AI Hires Apple Chip Vet to Lead Hardware – Rain AI hires an Apple chip executive to lead hardware engineering, marking the second high-profile hire for the startup aiming to design a new type of semiconductor for artificial intelligence.

Leaked Documents Show Nvidia Scraping ‘A Human Lifetime’ of Videos Per Day to Train AI – Nvidia is scraping videos from various sources to compile training data for its AI products, defending the practice as being in compliance with copyright law.

Nvidia reportedly delays its next AI chip due to a design flaw – Nvidia delays production of its “Blackwell” B200 AI chips by at least three months due to a design flaw, impacting major cloud providers.

Vimeo Announces AI-Powered Video Translation with Authentic Voice Cloning – Vimeo launches an AI-powered video translation solution that uses generative AI to translate video, audio, and captions into multiple languages while maintaining the original speakers’ voices, streamlining business processes and allowing organizations to reach global audiences with unprecedented ease and speed.

Generative AI Forces Media Firms to Pick Licensing or Litigation – Media firms are grappling with the decision to either sue AI companies for using their copyrighted work or to join them in licensing agreements, as the outcome of pending lawsuits could determine the future of the licensing market for training generative AI models.

Chinese Autonomous Driving Firm WeRide Plans US IPO – Chinese autonomous driving startup WeRide plans to go public in the U.S., facing legal and operational risks tied to the Chinese government, while the rise of AI prompts questions about the scalability of self-driving vehicles.

Apple Intelligence to Miss Initial Launch of Upcoming iOS 18 Overhaul – Apple’s artificial intelligence features will miss the initial launch of iOS 18, giving the company more time to fix bugs.

Investors in Adept AI will be paid back after Amazon hires startup’s top talent – Investors in Adept AI will be reimbursed after Amazon hires the startup’s top talent, despite concerns from regulators about the nature of the hiring.

Neura shows off humanoid robot 4NE-1 – Neura’s humanoid robot 4NE-1 is showcased in a promotional video, demonstrating various activities and highlighting the company’s collaboration with Nvidia’s robotics portfolio.

Canva acquires Leonardo.ai to boost its generative AI efforts – Canva acquires Leonardo.ai, a generative AI content and research startup, to deepen its investments in AI tech stack and expand its generative AI capabilities.

Tencent Joins $300 Million Financing for China’s AI Unicorn – Tencent contributes to $300 million financing for Chinese AI startup Moonshot, following in the footsteps of Alibaba, boosting Moonshot’s valuation to $3 billion.

AI-directed, driverless drift: Stanford Engineering and Toyota Research Institute achieve autonomous milestone – Stanford Engineering and Toyota Research Institute achieve a milestone in autonomous driving by creating the world’s first autonomous Tandem Drift team, using AI to direct two driverless cars to perform synchronized maneuvers.

Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget – Training large-scale text-to-image generative models on a micro-budget is made possible through a low-cost approach involving random masking, deferred masking strategy, and the use of synthetic images, resulting in competitive performance at significantly lower costs.

Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement – A principled approach is presented to provide LLM-based evaluation with a rigorous guarantee of human agreement, using selective evaluation and confidence estimation methods to ensure alignment with humans.

Transformers are Universal In-context Learners – Transformers are universal in-context learners, embraced by individuals and organizations for their values of openness, community, excellence, and user data privacy.

Knowledge Overshadowing Causes Amalgamated Hallucination in Large Language Models – Large language models can experience knowledge overshadowing, leading to amalgamated hallucinations, but the specific paper on this topic is not found.

UK antitrust body probes Google’s ties with AI rival Anthropic – UK antitrust body investigates Google’s investment in AI rival Anthropic, raising concerns about potential control over younger innovators without attracting regulatory scrutiny.

Elon Musk posts deepfake of Kamala Harris that violates X policy – Elon Musk shares deepfake video of Kamala Harris, potentially violating platform’s policies against synthetic and manipulated media, sparking concerns about AI-altered content in the upcoming election.

SF mayor promises accountability after school crossing guards tell NBC Bay Area driverless cars almost hit them – San Francisco officials plan to address safety concerns raised by school crossing guards about Waymo’s driverless cars, with the mayor promising accountability and new training for crossing guards.

World’s First-Ever AI Law Now Enforced in Europe, Targeting US Tech Giants – Europe enforces the world’s first AI law, targeting US tech giants with regulations on AI development, deployment, and use, imposing stricter rules on high-risk AI systems and banning “unacceptable” AI applications, with penalties for non-compliance.

With Smugglers and Front Companies, China Is Skirting American A.I. Bans – China is skirting American A.I. bans through a mazelike market in Shenzhen, where vendors openly offer microchips for artificial intelligence, despite U.S. efforts to control their export.

White House says no need to restrict ‘open-source’ artificial intelligence — at least for now – nan

We Need Positive Visions for AI Grounded in Wellbeing – AI’s transformative impact on society demands a shift towards developing AI systems that understand and support individual and societal wellbeing, requiring positive visions and concrete measures to guide AI development and deployment.

Elon Musk Says Robotaxis Are Tesla’s Future. Experts Have Doubts. – Elon Musk believes Tesla’s future lies in artificial intelligence and driverless taxis, but experts are skeptical about the company’s ability to achieve this vision.

‘The Godmother of AI’ says California’s well-intended AI bill will harm the U.S. ecosystem – The unintended consequences of California’s AI bill SB-1047 will harm the AI ecosystem, penalize developers, stifle open-source development, and cripple public sector and academic AI research.