Last Week in AI #300

Top News

OpenAI lays out plans for GPT-5

OpenAI CEO, Sam Altman, has outlined the company’s plans for its upcoming AI models, GPT-4.5 and GPT-5, in a recent roadmap. The GPT-4.5, internally known as Orion, is set to be the company’s last non-chain-of-thought model, with the aim to simplify OpenAI’s product lineup. The GPT-5 model is planned to integrate a lot of the company’s technology, including o3, and will no longer be shipped as a standalone model. While the exact release dates for these models are not specified, Altman has hinted at a timeline of weeks to months. Furthermore, upon the release of GPT-5, free ChatGPT users will have unlimited chat access at the standard intelligence setting, with Plus and Pro subscribers having access to higher levels of intelligence.

Google-backed public interest AI partnership launches with $400M+ pledged for open ecosystem building

Current AI, a public interest initiative backed by Google and other partners, has launched with over $400 million in pledges to foster the development of artificial intelligence (AI) for societal benefits. The initiative aims to raise $2.5 billion over the next five years to advance public interest in areas such as healthcare and climate goals. The initiative’s objectives include widening access to high-quality public and private datasets for AI training, supporting open-source infrastructure to enhance AI transparency and security, and developing systems to measure AI’s social and environmental impact. The initiative is backed by governments in France, Germany, Chile, Kenya, Morocco, and Nigeria, among others, as well as tech giants Google and Salesforce.

Thomson Reuters Wins First Major AI Copyright Case in the US

Thomson Reuters has emerged victorious in the first major AI copyright case in the United States, setting a precedent for future legal battles in the rapidly evolving field of AI. The media and technology conglomerate had accused legal AI startup Ross Intelligence of reproducing materials from its legal research firm, Westlaw, without permission. The court ruled in favor of Thomson Reuters, rejecting Ross Intelligence’s defenses and affirming that the company’s copyright was infringed. This ruling is significant as it challenges the ‘fair use’ doctrine often invoked by AI companies, suggesting that the use of copyrighted materials to train AI systems may not be legally permissible. The decision could have far-reaching implications for AI companies, potentially complicating their fair use arguments in future copyright disputes.

Elon Musk Leads $97.4 Billion Bid to Control OpenAI

Elon Musk, in collaboration with a consortium of investors including Vy Capital, Xai, and Hollywood power broker Ari Emanuel, has proposed a $97.4 billion bid to acquire the assets of OpenAI, a nonprofit organization he co-founded nearly a decade ago. This move marks Musk’s latest and potentially most daring attempt to gain control over the future of artificial intelligence, a field in which he has been in a longstanding dispute with OpenAI’s CEO, Sam Altman. Despite the substantial offer, the bid faces significant challenges as OpenAI’s board of directors is closely aligned with Altman, who has already dismissed Musk’s proposal. The ongoing negotiations and their implications for the AI industry were initially reported by The Wall Street Journal.

More on this:

• Elon Musk’s full offer letter to buy OpenAI reveals five key details
• Elon Musk Says He Will Drop OpenAI Bid if Company Preserves Nonprofit Mission
• OpenAI Rejects Elon Musk’s $97.4 Billion Bid for Control of the Company

Other News

Tools

OpenAI is rethinking how AI models handle controversial topics – OpenAI’s expanded Model Spec introduces guidelines for handling controversial topics, customizability, and intellectual freedom, while addressing issues like AI sycophancy and mature content, and is open-sourced for public feedback and industry use.

Open O1: Revolutionizing Open-Source AI with Cutting-Edge Reasoning and Performance – Open O1 aims to democratize access to advanced AI by developing open-source models that rival proprietary systems in reasoning and performance through innovative training techniques and community collaboration.

Adobe’s Sora rivalling AI video generator is now available for everyone – Adobe’s Generate Video tool, now in public beta, allows users to create five-second 1080p video clips using text and image prompts, with integration into Creative Cloud apps and commercial viability due to its training on public domain and licensed content.

Perplexity AI launches new ultra-fast AI search model Sonar – Sonar, Perplexity AI’s new search model, outperforms competitors in user satisfaction and speed by leveraging Meta’s Llama 3.3 70B and Cerebras Systems’ Wafer Scale Engines for enhanced search capabilities.

Perplexity launches its own freemium ‘deep research’ product – Perplexity’s new Deep Research tool offers a fast and accessible freemium option for in-depth research, outperforming many competitors in speed and scoring well on benchmarking tests, while OpenAI and Google focus on analytical depth and integration with existing ecosystems, respectively.

Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling – NVIDIA engineers successfully used the DeepSeek-R1 model with inference-time scaling to automatically generate optimized GPU attention kernels, outperforming manually crafted solutions in some cases.

Zyphra Introduces the Beta Release of Zonos: A Highly Expressive TTS Model with High Fidelity Voice Cloning – Zyphra’s Zonos-v0.1 beta release offers a sophisticated open-source TTS model with high-fidelity voice cloning, multilingual support, and customizable audio features, positioning it as a versatile tool for various applications in speech synthesis.

The AI animation engine – An AI animation engine requires users to render frames by clicking “Create” before any frames can be displayed.

Business

OpenAI is reportedly getting closer to launching its in-house chip – OpenAI is advancing its plans to produce an in-house AI chip with TSMC, aiming to reduce reliance on Nvidia and enhance its AI model capabilities.

France unveils 109-billion-euro AI investment as Europe looks to keep up with U.S. – France’s 109-billion-euro AI investment aims to bolster its AI sector and compete with the U.S. and China, with significant contributions from international and domestic entities, as global leaders gather to discuss AI’s future at the Paris summit.

Exclusive: Legal AI startup Harvey lands fresh $300 million in Sequoia-led round as CEO says on target for $100 million annual recurring revenue – Legal AI startup Harvey secures a $300 million investment led by Sequoia and aims to achieve $100 million in annual recurring revenue.

AI chip startup Groq secures $1.5 billion commitment from Saudi Arabia – Groq has secured a $1.5 billion investment from Saudi Arabia to expand its AI chip operations, including a data center in Dammam, and support technologies like the bilingual AI language model Allam.

OpenAI Must Face ‘Open AI’ Trademark Owner’s Infringement Claims – A federal judge ruled that OpenAI must face trademark infringement claims from Open Artificial Intelligence Inc. due to alleged consumer confusion over their similar branding.

Founded by DeepMind alumnus, Latent Labs launches with $50M to make biology programmable – Latent Labs, founded by a former DeepMind scientist, aims to revolutionize protein design and drug discovery by developing AI models that make biology programmable, reducing reliance on traditional wet lab experiments.

Anthropic AI Launches the Anthropic Economic Index: A Data-Driven Look at AI’s Economic Role – Anthropic AI’s new Economic Index uses data from millions of AI interactions to map AI’s role in various job sectors, revealing its significant presence in software development and writing tasks, while highlighting its limited use in lower-wage and highly specialized fields.

Microsoft powers AI ambitions with 400 MW solar purchase – Microsoft’s purchase of 389 megawatts of solar power from EDP Renewables North America is part of its strategy to meet the energy demands of its AI operations while advancing its goal to become carbon negative by 2030.

Reddit hints at expanded AI-powered search – Reddit plans to enhance its search capabilities by integrating AI-powered features like Reddit Answers to improve user navigation and engagement, while also aiming to drive growth and revenue.

AI-driven ads take the field during the 2025 Super Bowl – AI-themed advertisements dominated the 2025 Super Bowl, featuring major tech companies like OpenAI, Google, Meta, Salesforce, and GoDaddy showcasing their AI innovations, while Cirkul humorously highlighted AI’s potential pitfalls.

News publishers sue Cohere for copyright and trademark infringement – nan

Research

OpenAI’s DeepResearch can complete 26% of ‘Humanity’s Last Exam’ — a benchmark for the frontier of human knowledge – OpenAI’s DeepResearch AI agent has achieved a significant milestone by successfully completing 26% of “Humanity’s Last Exam,” setting a new benchmark in the field of AI performance.

Google DeepMind Achieves State-of-the-Art Data-Efficient Reinforcement Learning RL with Improved Transformer World Models – Google DeepMind’s new method in model-based reinforcement learning achieves state-of-the-art performance in the Craftax-classic environment by integrating advanced techniques like Dyna with warmup, patch nearest-neighbor tokenization, and block teacher forcing, significantly improving sample efficiency and surpassing previous models and human performance.

Matryoshka Quantization – Matryoshka Quantization introduces a novel multi-scale training method that optimizes model weights across multiple precision levels, enabling the creation of a single quantized model that can operate at various bit-widths with improved accuracy and efficiency, particularly for low-bit quantization like int2.

Scaling Pre-training to One Hundred Billion Data for Vision Language Models – Scaling vision-language models to 100 billion data points enhances cultural diversity and multilinguality, demonstrating significant benefits beyond traditional benchmarks despite the challenges of maintaining data quality and inclusivity.

LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters! – Large language models can significantly enhance their reasoning abilities by learning the structure of long chain-of-thought demonstrations, with structural coherence being more crucial than the specific content of individual reasoning steps.

LM2: Large Memory Models – LM2 introduces a novel memory-augmented Transformer architecture that enhances long-term dependency modeling and outperforms existing models in tasks requiring extensive context and complex reasoning.

Hephaestus: Improving Fundamental Agent Capabilities of Large Language Models through Continual Pre-Training – Hephaestus is a continually pre-trained open-source large language model that enhances fundamental agent capabilities, such as API function calling and intrinsic reasoning, by utilizing a specially curated pre-training corpus called Hephaestus-Forge, demonstrating superior performance compared to other open-source and commercial models.

Distillation Scaling Laws – Distillation scaling laws offer a framework for optimizing compute allocation between teacher and student models to enhance distilled model performance, with specific strategies depending on the existence and training needs of the teacher.

Gemstones: A Model Suite for Multi-Faceted Scaling Laws – Gemstones provides a comprehensive suite of model checkpoints to study the impact of design and selection on scaling laws, revealing their sensitivity to various architectural and training choices and offering modified scaling laws that account for practical considerations like GPU efficiency and overtraining.

Skill Expansion and Composition in Parameter Space – Parametric Skill Expansion and Composition (PSEC) is introduced as a framework that enhances autonomous agents’ learning efficiency and adaptability by maintaining a skill library and utilizing shared information across skills to address challenges like catastrophic forgetting and limited learning efficiency.

Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs – The article discusses the challenges of accessing a specific paper on emergent value systems in AIs due to its absence on the platform, suggesting users cite the arXiv link in their repositories to create a dedicated page.

Concerns

AI chatbots unable to accurately summarise news, BBC finds – BBC research reveals that major AI chatbots, including ChatGPT and Google’s Gemini, produce news summaries with significant inaccuracies and distortions, raising concerns about potential real-world harm.

‘Mass theft’: Thousands of artists call for AI art auction to be cancelled – Thousands of artists are protesting an AI art auction at Christie’s, claiming the technology exploits copyrighted work without permission, while some artists involved argue their AI models use their own inputs or public datasets.

Policy

LinkedIn cofounder Reid Hoffman, Hugging Face CEO Clement Delangue sign open letter calling for AI ‘public goods’ – Prominent tech leaders and AI researchers are advocating for the creation of AI “public goods” through public data sets and incentives for smaller, environmentally friendly AI models, emphasizing the need for societal control over AI development and deployment.

US and UK refuse to sign summit declaration on AI safety – The US and UK declined to sign a Paris summit declaration on AI safety, citing concerns over global governance and national security, while the US vice-president criticized Europe’s regulatory approach and warned against cooperation with China.

Macron urges Europe to simplify its regulations to get back into the AI race – Emmanuel Macron emphasized the need for Europe to simplify regulations and invest in AI to compete globally, while announcing a significant investment in the French AI ecosystem.

Scarlett Johansson calls for deepfake ban after AI video goes viral – Scarlett Johansson is urging lawmakers to prioritize legislation limiting AI use due to the dangers of deepfakes and the potential for AI to amplify hate speech.

Vance, in First Foreign Speech, Tells Europe That U.S. Will Dominate A.I. – Vice President JD Vance urged European leaders to align with the U.S. in the AI race by dismantling regulations, emphasizing America’s intent to lead in AI technology while cautioning against siding with authoritarian regimes like China.

Expert Opinions

Anthropic CEO Dario Amodei calls the AI Action Summit a ‘missed opportunity’ – Dario Amodei criticized the AI Action Summit in Paris as lacking urgency and clarity, urging faster and more transparent regulation to address the rapid advancement and potential risks of AI technology.

---

Copyright © 2024 Skynet Today, All rights reserved.

Source link