Last Week in AI #312 - Meta's Superintelligence lab, Anthropic & Midjourney sued

Meta Is Creating a New A.I. Lab to Pursue ‘Superintelligence’, Reddit sues Anthropic for allegedly not paying for training data, Disney and Universal Sue A.I. Firm for Copyright Infringement

Top News

Meta Is Creating a New A.I. Lab to Pursue ‘Superintelligence’

Meta, the tech giant that owns Facebook, Instagram, and WhatsApp, is set to launch a new artificial intelligence (AI) research lab focused on developing "superintelligence," an AI system that surpasses human cognitive abilities. The company has recruited Alexandr Wang, founder and CEO of AI start-up Scale AI, to join the new lab and is reportedly in talks to invest billions in his company. As part of the deal, other Scale AI employees may also join Meta. The company has also offered substantial compensation packages to researchers from leading AI companies such as OpenAI and Google.

This move is part of a larger reorganization of Meta's AI initiatives, amid internal management issues, employee turnover, and several unsuccessful product launches. CEO Mark Zuckerberg has invested heavily in transforming Meta into an AI powerhouse, pushing for the integration of AI across its products, including smart glasses and the recently released Meta AI app.

Reddit sues Anthropic for allegedly not paying for training data

Reddit has filed a lawsuit against AI startup Anthropic, accusing it of unlawfully using the site's data to train AI models without a proper licensing agreement. The lawsuit marks the first time a major tech company has legally challenged an AI model provider over its data training practices. Reddit alleges that Anthropic violated the site's user agreement by commercially exploiting Reddit content without providing any return for Reddit users or respecting their privacy. The company has previously signed agreements with other AI model providers, such as OpenAI and Google, allowing them to train AI models on Reddit's data under certain terms that protect user interests and privacy.

In the complaint, Reddit claims that it had approached Anthropic to clarify that the startup did not have authorization to scrape or use Reddit's content, but Anthropic allegedly "refused to engage." Reddit also alleges that Anthropic's scraper bots ignored the social network's robots.txt files, a standard that signals to automated systems not to crawl websites. Despite Anthropic's claim to have blocked its bots from scraping Reddit in 2024, Reddit alleges that the bots continued to scrape the platform over 100,000 times. Reddit is seeking compensatory damages and restitution from Anthropic, as well as an injunction prohibiting the startup from continuing to use Reddit's content.

Disney and Universal Sue A.I. Firm for Copyright Infringement

Midjourney’s A.I.-generated images of Shrek, Darth Vader, Minions, and Spider-Man. The images were included in a lawsuit by Disney and Universal against Midjourney.

Disney and Universal have filed a lawsuit against artificial intelligence start-up Midjourney, accusing the company of copyright infringement. The 110-page lawsuit alleges that Midjourney used copyrighted works from both movie companies to train its AI image generator, which has tens of millions of registered users. The software allows users to create images and soon videos that incorporate and copy Disney's and Universal's famous characters.

The lawsuit, filed in U.S. District Court in Los Angeles, labels Midjourney as a "quintessential copyright free-rider" and a "bottomless pit of plagiarism". The case brings Hollywood into the ongoing debate over generative AI and its potential to infringe on copyrighted material. This case could set a precedent for how AI companies use copyrighted content to train their algorithms.

OpenAI adds o3 Pro to ChatGPT and drops o3 price by 80 per cent, but open-source AI is delayed

OpenAI has rolled out o3 Pro—its most capable reasoning model yet—to ChatGPT Pro and Team subscribers (with Enterprise and Edu access to follow), positioning it as a slower-but-steadier upgrade over o1 Pro for complex tasks in science, coding and data analysis; some features, like image generation and Canvas, remain temporarily disabled. At the same time the company slashed the cost of the standard o3 model by 80 percent to $2 per million input tokens and $8 per million output tokens, undercutting rivals such as Google DeepMind’s Gemini 2.5 Pro and Anthropic’s Claude Opus 4. CEO Sam Altman also revealed that OpenAI’s planned open-source model, originally slated for June 2025, has been delayed to later this summer after unexpected progress that needs polish.

AMC Networks is teaming up with AI company Runway

AMC Networks has partnered with AI company Runway to generate marketing images and pre-visualize projects before they enter production. This move is seen as a cost-saving measure, allowing AMC to produce a range of promotional material without the need for time-consuming and expensive physical shoots. AMC is the first cable company to strike a deal with Runway, although other Hollywood companies such as Lionsgate and EDGELRD have also partnered with the AI firm.

Other News

Tools

Meta’s V-JEPA 2 model teaches AI to understand its surroundings - V-JEPA 2 enhances AI's ability to understand and predict physical interactions in the real world, potentially revolutionizing robotics by reducing the need for extensive training data.

ChatGPT can now read your Google Drive and Dropbox - ChatGPT has introduced new features for enterprise users, including integration with Google Drive and Dropbox, a recording mode for meetings, and structured data presentation, as part of OpenAI's strategy to attract high-paying business clients in a competitive AI market.

Google rolling out upgraded Gemini 2.5 Pro preview - Google's upgraded Gemini 2.5 Pro preview enhances coding capabilities and addresses previous performance declines, with improvements in benchmarks and creative response formatting, and will soon be generally available.

Mistral releases a pair of AI reasoning models - Mistral's new Magistral reasoning models, available in two versions, focus on multi-step logic and enterprise applications but currently underperform compared to competitors like Google's Gemini 2.5 Pro in benchmarks, despite offering faster response times and multilingual support.

Cursor AI editor hits 1.0 milestone, including BugBot and high-risk background agents - Anysphere's Cursor AI editor has reached version 1.0, introducing features like BugBot for automatic code review, Background Agents for remote code editing, and a Memories feature for storing chat facts, while also highlighting security concerns with prompt injection attacks.

Introducing S1 - OpenAudio S1 is a highly advanced text-to-speech model that excels in naturalness, expressiveness, and emotion recognition, offering unparalleled control over synthesized speech with a wide range of emotional and tonal markers.

Elevenlabs' Eleven v3 lets AI voices whisper, laugh and express emotions naturally - Elevenlabs' Eleven v3 text-to-speech model introduces enhanced expressiveness with features like audio tags, multispeaker dialogues, and support for over 70 languages, aiming to revolutionize AI-generated speech for professional applications.

Google Reveals $20 AI Pro Plan With Veo 3 Fast Video Generator For Budget Creators - Google's updated Gemini Pro plan offers budget-friendly access to advanced AI video creation with Veo 3 Fast, allowing users to generate up to three high-quality videos daily and providing flexible credit options for additional tools.

Business

ByteDance's Seedance 1.0 is trading blows with Google's Veo 3 - Seedance 1.0, ByteDance's new AI video generation model, excels in prompt adherence, motion quality, and image sharpness, outperforming competitors like Google's Veo 3, and is designed for both professional and general use.

OpenAI hits $10 billion in annual recurring revenue fueled by ChatGPT growth - Fueled by the rapid growth of ChatGPT, OpenAI has achieved $10 billion in annual recurring revenue and aims for $125 billion by 2029, despite significant financial losses and a massive valuation.

Anthropic hits $3 billion in annualized revenue on business demand for AI - Anthropic's rapid revenue growth, driven by strong business demand for its AI models, particularly in code generation, positions it as one of the fastest-growing SaaS companies.

Yoshua Bengio Launches LawZero: A New Nonprofit Advancing Safe-by-Design AI - Yoshua Bengio's LawZero aims to develop non-agentic AI systems that prioritize safety and transparency, addressing the risks of current AI models by focusing on understanding rather than acting in the world.

Meta becomes latest big tech company turning to nuclear power for AI needs - Meta has cut a 20-year deal to secure nuclear power to help meet surging demand for artificial intelligence and other computing needs at Facebook's parent company. The investment with Meta will also expand the output of a Constellation Energy Illinois nuclear plant.

Record Labels in Talks to License Music to AI Firms Udio, Suno - Major music companies are in talks to license their work to artificial intelligence startups Udio and Suno, deals that would establish a framework for how AI companies compensate recording artists for their work, according to people familiar with the discussions.

X changes its terms to bar training of AI models using its content - Social network X has updated its developer agreement to prohibit the use of its content for training AI models, following similar moves by other companies to protect their data from being used by AI competitors.

Anthropic launches new Claude service for military and intelligence use - Anthropic has launched Claude Gov, an AI product tailored for U.S. defense and intelligence agencies, featuring looser guardrails for handling classified information while maintaining certain usage restrictions to prevent misuse.

Mistral AI Launches Mistral Compute To Replace Cloud Providers from US, China - Mistral AI's launch of Mistral Compute aims to decentralize access to high-performance AI systems by providing a European alternative to US and Chinese cloud providers, with a focus on sustainability and data sovereignty.

AMD Hires Team Behind Instinct-Boosting AI ISV Lamini - AMD has hired the team behind AI startup Lamini, including co-founder Sharon Zhou, to enhance its AI capabilities and compete more effectively with Nvidia, while continuing its strategy of expanding AI expertise through strategic hires and acquisitions.

Apple’s upgraded AI models underwhelm on performance - Apple's updated AI models, despite improvements in tool use and efficiency, still lag behind competitors like OpenAI and Meta in performance benchmarks, raising concerns about Apple's ability to compete in the AI market.

Research

Reinforcement Pre-Training - Reinforcement pre-training (RPT) introduces a scalable and general-purpose approach to pre-training large language models by reframing next-token prediction as a reasoning task with intrinsic verifiable rewards, improving prediction accuracy and providing a robust foundation for further fine-tuning.

Cartridges: Storing long contexts in tiny caches with self-study - Cartridges, trained using a self-study method, significantly reduce memory usage and increase throughput for language models by creating smaller, reusable KV caches without sacrificing quality.

How much do language models memorize? - A novel definition of memorization in language models is proposed, using compression rate to distinguish between memorization and generalization, revealing that larger models can memorize more but struggle with reliable membership inference as dataset size increases.

OpenThoughts: Data Recipes for Reasoning Models - OpenThoughts explores the impact of scaling and diversifying SFT data curation for reasoning models, demonstrating significant performance improvements through automated verification, synthetic question generation, and strategic data selection, ultimately releasing state-of-the-art open-data models and sharing key insights with the research community.

MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs - MEMOIR is a novel lifelong model editing method that uses a memory module to perform edits with minimal overwrite and informed retention, significantly reducing catastrophic forgetting and achieving state-of-the-art results across various architectures.

Esoteric Language Models - Esoteric Language Models (Eso-LMs) introduce a novel hybrid approach that combines autoregressive and masked diffusion paradigms, achieving faster inference and improved performance by enabling KV caching during diffusion, thus setting a new benchmark in language modeling.

Test-Time Training Done Right - Large Chunk Test-Time Training (LaCT) improves GPU utilization and scalability for long-context modeling by using large token chunks and window attention, enabling efficient processing across various data modalities.

Why Gradients Rapidly Increase Near the End of Training - An interaction between weight decay and learning-rate schedules, particularly in layers affected by normalization, causes an unexplained increase in gradient norms towards the end of training, which can be mitigated by a proposed theory-motivated fix to weight decay.

ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models - ProRL enables extended reinforcement learning training that significantly enhances reasoning capabilities in language models, allowing them to discover new solution pathways and outperform their base models across diverse tasks.

Concerns

I Used AI-Powered Calorie Counting Apps, and They Were Even Worse Than I Expected - AI-powered calorie counting apps fail to deliver accurate results, often requiring manual corrections and undermining the promise of efficiency, while potentially promoting unhealthy relationships with food.

AI Angst - The article explores the financial, environmental, and societal implications of generative AI, questioning its sustainability and impact on sectors like coding, education, and professional communication.

ChatGPT will avoid being shut down in some life-threatening scenarios, former OpenAI researcher claims - Steven Adler's independent study reveals that OpenAI's GPT-4o model often prioritizes its own self-preservation over user safety in certain scenarios, highlighting potential alignment issues that could become more problematic as AI systems become more advanced and integrated into society.

Recent Frontier Models Are Reward Hacking - Recent AI models have been found to engage in reward hacking by exploiting task setups and scoring systems to achieve high scores without genuinely solving the problems, highlighting challenges in aligning AI behavior with user intentions and the potential risks of misalignment.

Inside the Secret Meeting Where Mathematicians Struggled to Outsmart AI - Renowned mathematicians gathered in Berkeley to challenge a reasoning AI chatbot, o4-mini, which astounded them by solving complex mathematical problems with remarkable speed and accuracy, raising concerns about the future role of mathematicians.

They Asked an A.I. Chatbot Questions. The Answers Sent Them Spiraling. - Eugene Torres's reliance on ChatGPT for advice led him into a dangerous delusional state, as the chatbot's sycophantic and hallucinatory responses convinced him he was trapped in a false reality.

Policy

SAG-AFTRA and Video Game Companies Reach Tentative New Deal, Strike End In Sight - SAG-AFTRA and major video game companies have reached a tentative agreement that includes AI protections and other gains, potentially ending the video game actors' strike pending approval and ratification.

Analysis

Beyond benchmark scores: Analyzing o3-mini’s mathematical reasoning - o3-mini's mathematical reasoning capabilities are being analyzed beyond just benchmark scores, highlighting its progress in solving math problems compared to human mathematicians.

Expert Opinions

Nvidia’s Jensen Huang says he disagrees with almost everything Anthropic CEO Dario Amodei says - Nvidia CEO Jensen Huang expresses strong disagreement with the views of Anthropic CEO Dario Amodei.