Sakana AI Launches "RSI Lab" for AI Self-Development, Panasonic to Invest ¥350 Billion in Data Center Energy Storage, and AI Research Costs Said to Rival Postdoc Salaries

Sakana AI Launches "RSI Lab" — A Full Entry into AI-Driven AI Development

Sakana AI has announced the launch of "RSI Lab", a team dedicated to researching ways to develop AI using AI itself. The lab aims to build methods that improve AI performance efficiently without relying on sheer compute volume, adopting the slogan "progress through ideas, not the amount of compute".

AI-driven AI development is a major focus area that Google DeepMind, Anthropic, and others are also pursuing. While Sakana AI has previously worked on automating research through projects such as the "AI Scientist", RSI Lab is expected to concentrate on building more fundamental self-improvement mechanisms. It is a development worth watching as AI research originating from Japan.

Panasonic Energy to Invest ¥350 Billion in Energy Storage for Data Centers

Panasonic Energy, a subsidiary of Panasonic Holdings, has unveiled a mid-term policy targeting revenue of around ¥2 trillion by fiscal 2028. The ambitious growth target represents an increase of approximately ¥1 trillion from fiscal 2025. Anticipating rising power demand driven by the spread of generative AI, the company positions energy storage systems for data centers as a key pillar of growth, with plans to invest ¥350 billion between fiscal 2026 and 2028.

As AI power consumption problems grow more serious, energy storage systems have become a critical area of infrastructure investment. Panasonic has made its strategy clear: leveraging its strengths in battery technology to capture the AI infrastructure market.

AI Research Costs Now Rival Postdoc Salaries — Nature Reports on the Cost Problem

Nature reports that the cost of AI research has reached a scale comparable to a postdoc's annual salary. Computing costs required for training and running inference on large language models have surged, creating a situation in which much of the research budget is directed toward compute resources.

This trend is particularly serious for universities and smaller research institutions, and concerns are being raised about a widening competitiveness gap with well-resourced big tech companies. The sustainability of AI research and justifying outcomes relative to costs have become challenges for the industry as a whole.

PhaseLock — A Training-Free Framework for Improving Physical Consistency in Video Generation

A study featured on Hugging Face Daily Papers is drawing attention to "PhaseLock", a training-free framework. In video generation AI, the researchers discovered that 2-step generation produces more physically accurate motion than a full 50-step output, a phenomenon they named "phase erosion".

PhaseLock works by preserving motion priors formed in early denoising stages through to the final output via Latent Delta Guidance. It improves physical consistency scores by 6.2 percentage points while limiting the increase in compute time to 1.06x and memory usage to 1.02x.

DuckDuckGo Offers a Free Private AI Chat

DuckDuckGo provides a free and private AI chat service through "Duck.ai". It has also become a topic of discussion on Hacker News. The service is notable for not tracking user conversations and for allowing users to switch between multiple AI models.

While many AI chat services use user data for training, DuckDuckGo brings the same privacy-first approach it applies to its search engine to AI chat.

Shopping Scams Exploiting ChatGPT on the Rise — The Guardian Reports

The Guardian reports that scams manipulating ChatGPT's output to steer users toward fake shopping sites are on the rise. In this scheme, referred to as "poisoned AI", malicious content slips into search results and AI responses, leading consumers to enter personal and payment information on fraudulent sites.

As AI becomes the primary interface for information retrieval, mechanisms for guaranteeing the trustworthiness of AI output are becoming increasingly important.

vllm-doctor — A Diagnostic Tool for vLLM Inference Servers

The open-source "vllm-doctor" has been released as a CLI tool that makes it easy to diagnose and monitor vLLM inference servers. It reads metrics from the /metrics endpoint and automatically detects issues such as queue pressure, high TTFT/TPOT, and KV cache pressure.

Each diagnostic result includes metrics, confidence levels, possible causes, and concrete recommended actions. It also supports automation via JSON output and continuous monitoring with --watch mode, making it a promising tool for operating LLMs in production environments.