Many Major AI Chatbots Lean "Left" on Political Questions — Washington Post Study

A Washington Post investigation has revealed that the majority of major AI chatbots skew their answers to political questions toward the left. OpenAI's GPT-5.5 gave left-leaning answers in 80% of cases, and even Grok, which Elon Musk positions as "anti-woke," came out left-leaning in a majority of cases. Google's Gemini 3.1 Pro, on the other hand, presented both sides in 93% of cases, making it the most balanced on this point.

Political bias in AI models has long been a recurring subject of debate. These findings suggest that, regardless of model developers' intentions, training data and alignment methods may be producing biased responses. As AI becomes an everyday tool for information retrieval and decision-making support, the impact such bias could have on society will continue to warrant close attention.

Micron Overtakes Meta and Tesla in Market Cap on AI Infrastructure Demand

According to Reuters, semiconductor maker Micron's market capitalization has surpassed that of Meta and Tesla. The backdrop is insatiable AI infrastructure demand, which has sent demand for memory chips soaring. Large memory capacity is essential for training and running inference on AI models, and memory, like GPUs, has become a bottleneck for the AI industry.

This shows that the AI industry's growth is rippling across the entire supply chain, not just GPU makers like NVIDIA. Meanwhile, there are also reports of high-end GPU prices skyrocketing — the RTX 6000 Pro has reportedly jumped from $8,000 to $13,000 — so rising hardware procurement costs are becoming a real challenge for individual developers and small teams.

General Intuition Raises $320 Million at $2.3 Billion Valuation — Training AI Agents Through Gameplay

Startup General Intuition has raised $320 million at a valuation of approximately $2.3 billion. The company takes an approach that trains AI agents on millions of hours of gameplay data. The aim is to have AI acquire something like human-like "intuition" from video game action data.

Unlike conventional language-model-centric approaches, the idea of learning behavioral data from simulated environments such as games could potentially be applied to robotics and physical-world tasks. Bridging the gap between simulation and reality will be the key, but it is an approach worth watching.

Former Databricks AI Lead Aims to Cut AI Power Consumption to 1/1,000th

A new project called "Un0," launched by Databricks' former head of AI, is reportedly developing technology that can reduce AI power consumption to 1/1,000th of conventional levels. Un0 has reportedly carried out its first proof of concept as an image generation system, demonstrating that it can replicate conventional AI systems.

The power consumed by training and running AI models has become a serious challenge as the scale of these systems grows. If a 1,000-fold reduction becomes a reality, it would not only dramatically cut AI's environmental footprint but also make running large-scale models on edge devices and consumer hardware a realistic prospect. Technical details remain limited at this point, and actual performance and applicability will need to be verified going forward.

Running GLM 5.2 on Dual RTX 5090s — Large Model Operation on Consumer Hardware

A report posted on Reddit's LocalLLaMA community describes running a quantized version of GLM 5.2 (Unsloth GGUF UD-Q5\_K\_S, approximately 492GB) on consumer-grade hardware. The setup consisted of a 32-core Zen5 Threadripper Pro 9975WX, dual RTX 5090s, and 512GB of DDR5 ECC RAM — hard to call consumer-grade in the strict sense, but it shows we are entering an era where models approaching 500GB can run without datacenter-class infrastructure.

In addition, a new llama.cpp release (b9788) adds support for a tensor-split mode for Intel GPUs, which is expected to improve inference efficiency in multi-GPU environments. In the local LLM community, how to run large open-source models as efficiently as possible is being actively debated.

Notion Shuts Down Its AI-Powered Email Client — Mac and iOS Apps Discontinued Too

Notion has announced that it will discontinue its AI-powered email client, including the Mac and iOS apps. Notion has been aggressively rolling out AI features as an all-in-one productivity tool, but its withdrawal from email may signal that simply bolting on AI features is not enough to differentiate from established players.

AI-powered email clients are a space where multiple players compete, including Superhuman and Shortwave. Considering alignment with Notion's core values and actual user needs, the move can also be read as a strategic decision.

---