The Wonder of a 6502 Emulator Running Inside an LLM, OSCAR Compresses the KV Cache 7x, and Japan's FSA Calls for AI Vulnerability Measures

A 6502 Emulator Running Inside an LLM — A Curious Experiment at One Instruction per Second

Adam Dunkels has published a strange experiment: running an emulator for the 6502 processor within an LLM's text output. The approach works by having the LLM itself interpret emulator code written in Markdown and execute 6502 machine code one instruction at a time inside it.

Unsurprisingly, it runs at one instruction per second. While not practical, it is a fascinating demonstration of the possibility that LLMs can go beyond mere text generation and "act as computers." On Hacker News, it has been drawing attention for "pushing the boundary between the limits and potential of LLMs."

Experiments like this raise a fundamental question: what is it that an LLM actually "understands"? The fact that it can track and emulate state transitions, going beyond the scope of simple pattern matching, deserves attention.

OSCAR: Compressing the KV Cache About 7x with 2-Bit Quantization

Drawing attention in the LocalLLaMA community is OSCAR (Offline Spectral Covariance-Aware Rotation). For INT2 quantization of the KV cache, it pre-computes an orthogonal rotation offline, aligned with the directions that attention actually consumes, reportedly achieving roughly 7x KV cache compression.

Particularly interesting is that with Qwen3-32B, the GPQA score actually improves from 58.49 with BF16 to 60.40 with OSCAR INT2. Qwen3-4B-Thinking also nearly preserves accuracy, going from 67.27 to 67.17.

Pre-computed rotation matrices are available on HuggingFace as `.pt` files, so users can drop them in without running their own calibration. For local LLM users facing memory issues in long-context inference, this looks set to become a compelling option.

Japan's FSA and BOJ Ask Financial Institutions to Prepare for Mass AI-Driven Vulnerability Discovery

An important development in domestic news. The Financial Services Agency and the Bank of Japan have asked financial institutions to prepare early for a scenario in which advances in frontier AI lead to large volumes of cyber vulnerabilities being discovered in short periods.

They are requesting nine specific measures under the direct involvement of top management. Behind this request lies the Project Glasswing report that Anthropic published last week. The striking result — roughly 50 partner companies discovered more than 10,000 high- and critical-severity vulnerabilities in just one month — is starting to move regulators around the world.

For Japanese financial institutions as well, prioritizing patch deployment and reviewing response processes in light of AI-driven vulnerability scanning results looks likely to become an urgent task.

Explaining LLM Usefulness Through the "Gap Between Generation and Verification"

An interesting analysis has been shared on Hacker News: a framework that explains "why LLMs are useful" through the cost gap between Generation and Verification.

The argument goes that even if solving a problem from scratch is hard, judging whether a given answer is correct is relatively easy — and this asymmetry is the very source of LLMs' practical value. It provides theoretical backing for the pattern that productivity improves in code generation and writing when humans focus strictly on the "verification" role.

This perspective also carries practical implications for organizations considering AI adoption. Let AI handle "generation" while humans concentrate on "verification" — designing that division of roles looks likely to be the key to making good use of AI.

Final Fantasy Creator Praises AI-Generated Remake Video as "Amazing"

Debate over AI-generated content continues in the gaming industry as well. Hironobu Sakaguchi, the creator of Final Fantasy, made headlines when he watched an AI-generated video styled after a Final Fantasy 6 remake and rated it "Amazing."

According to Kotaku, opinion is divided in fan communities, but the fact that the original creator himself responded positively is symbolic. The impact of AI-generated content on creative processes in the gaming industry will continue to warrant close attention.