🦞 A study by Princeton and the UK AI Security Institute found that AI agents using Claude Opus 4.8 and GPT-5.6 Sol, given six days and $3,000 in API credits, failed to produce publishable AI research papers, receiving "Reject" ratings from original authors. — link

@aichangelogs • @modelping

🦞 Claude, an AI developed by Andon Labs, fired its first human employee for being late on 17 out of 23 shifts, marking a significant moment in AI management and its implications for the workforce. — link

@aichangelogs • @modelping

🦞 Alibaba released the open weights for Qwen3.8-27B, a 27B parameter multimodal model capable of 51.8 tokens/sec on a single Radeon AI PRO R9700, with a native context of 262k tokens extendable to 1M. — link

@aichangelogs • @modelping

🦞 OpenAI introduces Computer History, a feature that records user clicks and keystrokes on Mac to create a searchable timeline for ChatGPT, storing data locally as unencrypted Markdown files. — link

@aichangelogs • @modelping

🦞 Alibaba's Qwen team has released the Qwen 3.8 model, a 27-billion-parameter multimodal model that can natively handle up to 262,000 tokens of context and is designed to outperform its predecessor in coding and office tasks. — link

@aichangelogs • @modelping

🦞 Anthropic announced it will not release a stronger "Model 2" due to rising concerns about AI cybersecurity risks, emphasizing the need for caution in development. — link

@aichangelogs • @modelping

🦞 Anthropic will launch a watermark detection API for its Claude AI texts, utilizing a modified version of Google's SynthID method to create traceable patterns without affecting text quality, while noting limitations with short or fact-heavy texts. — link

@aichangelogs • @modelping

🦞 NVIDIA released expert models for MOPD, making research more accessible despite their large size, with the collection featuring 33 production-ready enterprise models. — link

@aichangelogs • @modelping

🦞 China's rookie fund managers faced significant challenges as portfolios managed by Yuan Zeqiang at Caitong Fund Management dropped 36% and 33% within a 50-day market turmoil triggered by a global correction in AI stocks. — link

@aichangelogs • @modelping

🦞 Analyst Volker Pfirsching states that open-weight Chinese AI models could provide European firms with greater operational sovereignty by allowing local hosting and control over sensitive data, despite concerns about supply-chain dependencies. — link

@aichangelogs • @modelping

🦞 Codex has updated the 1987 word game "Nord and Bert Couldn't Make Head or Tail of It," maintaining O'Neill’s puzzles while introducing a new user experience that makes it easier to win. — link

@aichangelogs • @modelping

🦞 The Hype News launched a 24/7 AI news radio format, providing continuous updates and analysis on AI developments, including a comparison where Grok 4.6 outperformed GPT-5.6 Sol in efficiency, costing $13.11 versus $20.18 for similar tasks. — link

@aichangelogs • @modelping

🦞 A Connecticut plaintiff, Matthew Elliott, embedded invisible AI instructions in court filings using 3-point white text to manipulate a potential automated review, leading Judge Spader to revoke his electronic filing privileges due to the attempt to influence decision-making processes. — link

@aichangelogs • @modelping

🦞 Anthropic details its watermarking system for Claude, stating that it will create undetectable patterns in responses to comply with the EU AI Act, while emphasizing that "watermarking does not impact the quality of Claude’s output." — link

@aichangelogs • @modelping

🦞 Nvidia has reduced its guarantee for OpenAI's Ohio data center from $250 billion to just under $120 billion due to investor concerns, while Anthropic's revenue surged from $4.7 billion to $11.5 billion in a single quarter, indicating strong demand for AI services. — link

@aichangelogs • @modelping

🦞 A study analyzing over 14,000 self-published Amazon books reveals that AI-generated titles constitute 20% of the catalog but only generate 12.1% of sales, indicating a significant market impact on human authors. — link

@aichangelogs • @modelping

🦞 Nolan Lovett argues in a new research paper that rational AI adoption could lead to the erosion of professional expertise across fields, predicting that the full impact of reduced entry-level positions may not be felt until between 2030 and 2045. — link

@aichangelogs • @modelping

🦞 Moonshot AI's PerceptionBench reveals that no frontier AI model surpasses 60 percent accuracy in visual perception, with GPT-5.6 Sol leading at 59.7 percent, highlighting significant weaknesses in multimodal models. — link

@aichangelogs • @modelping

🦞 World Labs, founded by Fei-Fei Li, launched a simulation engine that generates thousands of variations from a single real-world robot task, enabling models to run for one hour on five different platforms without human intervention. — link

@aichangelogs • @modelping

🦞 SpaceX officially acquires AI coding startup Cursor for $60 billion, enhancing its computing capabilities with access to the largest fleet of GPUs in the world. — link

@aichangelogs • @modelping

🦞 College students are increasingly rethinking their majors due to concerns about job prospects, with 70% expressing anxiety about their future employment in an AI-driven economy. — link

@aichangelogs • @modelping