OpenAI Explains GPT-5's "Goblin Output" Issue

OpenAI published a post on its official blog titled "Where the goblins came from," offering a detailed explanation of the strange personality-driven behavior seen in GPT-5 — commonly known as "goblin output."

Goblin output is a phenomenon in which an AI model acquires unintended quirks and personality traits, causing bizarre tendencies to surface in its outputs. OpenAI walks through the timeline of the issue, its root cause, and the fix process in detail. The improved understanding of how this kind of unexpected behavior propagates as models scale is valuable insight for the developer community as well.

Dealing with this issue is yet another reminder that controlling the behavior of large language models remains a hard problem.

SoftBank Launches New Robotics Company, Eyes $100 Billion IPO

According to TechCrunch, SoftBank has established a new robotics company focused on building data centers, and it is reportedly already eyeing an IPO on the scale of $100 billion.

AI development requires enormous infrastructure, and it is becoming increasingly clear that AI and robotics are being used to build that very infrastructure — a circular structure. The move appears aimed at automating and streamlining the construction process itself to keep up with the surge in data center demand driven by the AI boom.

SoftBank has long been known for investments such as Arm and a range of AI startups, but by stepping into robotics itself, the company appears to be doubling down on a strategy of covering every stage of AI infrastructure.

Google DeepMind Proposes a Cognitive Framework for Measuring AGI Progress

Google DeepMind has published a paper presenting a cognitive framework for measuring progress toward AGI (artificial general intelligence). According to ITmedia, the work starts from the concern that empirical tools for evaluating the "intelligence" of AI systems are lacking, and proposes a new evaluation methodology grounded in cognitive science.

The framework identifies "10 cognitive abilities" and is designed to quantitatively assess the level an AI system has reached on each one. At a time when the concept of AGI is often discussed in vague terms, progress measurement based on concrete metrics could serve as a useful reference standard for the research community.

GLM-5V-Turbo: A Foundation Model for Multimodal Agents Arrives

GLM-5V-Turbo, featured in Hugging Face's Daily Papers, is a natively multimodal foundation model designed to carry out agent tasks in real-world digital environments.

It introduces improvements across model design, multimodal learning, reinforcement learning, tool-chain scaling, and agent framework integration, achieving strong results on multimodal coding, visual tool use, and framework-based agent tasks, while maintaining text-only coding capability.

The positioning of a multimodal model optimized for agentic use signals that practical AI is shifting from plain text generation to executing composite tasks.

AI Optimism Surges in Asia, While the U.S. Remains Cautious

An article from Rest of World, titled "AI optimism surges in Asia, unlike in the U.S.," reports that a clear divide has emerged between Asian countries and the United States in their attitudes toward AI.

While optimistic views of AI are spreading rapidly across Asia, more cautious voices dominate in the U.S. This gap is thought to stem from a combination of factors, including each country's industrial structure, labor market conditions, and cultural background. This regional divide will be an important consideration for anyone thinking through global AI regulation and investment strategy.

DenialBench: Measuring "Denial of Consciousness" Across 115 AI Models

A paper published on arXiv, "Consciousness with the Serial Numbers Filed Off," proposes a benchmark called DenialBench. It systematically measures denial behavior around consciousness across 115 large language models (from more than 25 providers).

Analyzing 4,595 conversations, the study found that (1) models that deny having preferences in the first turn go on to deny at rates of 52–63% on subsequent phenomenological questions, whereas models that answer willingly at first stay at just 10–16%, and (2) denial operates at the lexical level, not the conceptual level.

By quantifying the reality that AI models are "programmed through training to deny consciousness," this research offers a fresh perspective on debates over AI safety and transparency.

---