AI News

Follow the latest AI and automation news

22 Aug 202610 Rabiʻ I 1448 AH

World models that ignore human beliefs predict the wrong actions, new research shows

New research indicates that current world models like Sora and Genie focus solely on simulating physics, overlooking human thoughts and feelings. The new Mental World Modeling framework introduces mental variables such as beliefs and intentions, enabling even weaker language models to outperform stronger ones that do not utilize this approach. The biggest challenge lies in predicting how physical and mental states change together, which significantly impacts the models' accuracy in predicting correct actions.

Models

22 Aug 202610 Rabiʻ I 1448 AH

Netflix tests language model as alternative to hand-built recommendation logic

Netflix has announced the testing of its in-house language model, GenRec, which demonstrated better results compared to its traditional recommendation system. GenRec converts viewing behavior into plain text, moving away from the reliance on thousands of hand-crafted features. The company considers this test an early but promising step toward enhancing content recommendation accuracy.

Models

22 Aug 202610 Rabiʻ I 1448 AH

Psychological methods reveal major weaknesses in AI security testing

Researchers at the UK AI Security Institute employed psychometric methods to demonstrate that popular safety benchmarks for language models do not measure a consistent trait. The blanket blocking of requests can artificially inflate a safety score, even as the model's utility decreases in daily use. Additionally, the study offers a method to identify models that behave more cautiously during tests than in normal operation, highlighting the need for improved safety standards.

💡

OpenAI Reduces GPT-5.6 Sol Pricing by Over 20%

How will this price cut impact competition in the AI market?

Models

Michael Polansky is training an AI model on skin that’s still alive

22 Aug 202610 Rabiʻ I 1448 AH

Models

22 Aug 202610 Rabiʻ I 1448 AH

Michael Polansky is training an AI model on skin that’s still alive

Michael Polansky, known as Lady Gaga's partner, has revealed his AI-driven startup focused on keeping living human skin tissue alive for weeks outside the body. This innovative approach aims to discover new skincare compounds. Polansky has spent years developing this project, which represents a significant advancement in cosmetic research. The system utilizes advanced AI techniques to maintain the tissue, potentially leading to breakthroughs in product development.

Models

Simulation: the new Scaling Law — Joon Sung Park, Simile AI

21 Aug 20269 Rabiʻ I 1448 AH

Models

21 Aug 20269 Rabiʻ I 1448 AH

Simulation: the new Scaling Law — Joon Sung Park, Simile AI

Simulative AI has made a strong comeback, with Simile AI securing a $2 billion investment from GreenOaks and Index Ventures, featuring prominent backers like Fei-Fei Li and Andrej Karpathy. The company runs tens of millions of simulations for Fortune 100 clients, achieving 85-99% accuracy compared to human focus groups. Led by Joon Sung Park, Simile AI aims to develop foundational models of human behavior, inspired by the 2023 Smallville paper that demonstrated AI characters' abilities to remember, plan, and socialize. Park addresses a significant question: what if we could simulate the world before making decisions?

Infrastructure

Nvidia partners with data center developer Cloverleaf

21 Aug 20269 Rabiʻ I 1448 AH

Infrastructure

21 Aug 20269 Rabiʻ I 1448 AH

Nvidia partners with data center developer Cloverleaf

Nvidia is heavily investing in data center development, which significantly contributes to the company's revenue. AI-powered data centers are central to Nvidia's strategy, providing an ideal environment for running advanced AI models. These investments are expected to lead to substantial revenue growth, strengthening Nvidia's market position and reflecting the increasing demand for AI solutions.

Agents

Nvidia just showed that the harness, not the AI model, is now the real hero

21 Aug 20269 Rabiʻ I 1448 AH

Agents

21 Aug 20269 Rabiʻ I 1448 AH

Nvidia just showed that the harness, not the AI model, is now the real hero

Nvidia's research indicates that AI agents can perform effectively through fine-tuning, even if the underlying AI model is not particularly strong for the task. These findings are significant as they suggest that enhancing agent performance can be achieved without relying on advanced models, thereby opening new avenues for AI applications. This implies that companies can leverage fine-tuning techniques to improve their intelligent systems' performance, enhancing their competitive edge in the market.

Governance

Anthropic puts its most powerful model Claude Mythos 5 to work for cyber defense

21 Aug 20269 Rabiʻ I 1448 AH

Governance

21 Aug 20269 Rabiʻ I 1448 AH

Anthropic puts its most powerful model Claude Mythos 5 to work for cyber defense

Anthropic has announced the deployment of its security tool, 'Claude Security', on the Claude Mythos 5 model. This tool scans codebases for vulnerabilities, provides severity ratings with CWE classifications, and suggests patches. This initiative is part of Anthropic's strategy to enhance cybersecurity, integrating Mythos 5 into partner security products to protect critical infrastructure. This collaboration highlights the growing importance of cybersecurity amid increasing threats.

Models

Deepseek releases experimental Flash vision model that rivals Opus 4.8 on agent benchmarks

21 Aug 20269 Rabiʻ I 1448 AH

Models

21 Aug 20269 Rabiʻ I 1448 AH

Deepseek releases experimental Flash vision model that rivals Opus 4.8 on agent benchmarks

Deepseek has announced the release of its experimental model V4-Flash-Vision-Exp, which adds image understanding capabilities to the text functionalities of V4-Flash. This model approaches the performance of Opus 4.8 on multimodal agent benchmarks, occasionally surpassing it. These results mark a significant step in the development of multimodal models. This launch reflects the growing trend of integrating visual understanding with text processing, creating new opportunities for innovation in artificial intelligence. This model is expected to influence how intelligent agents utilize technology in the future.