Research & Papers

Spotless, a dishwasher-loading game on a physical simulation of a real machine, was released to benchmark language models in a browser.

• •It is designed as a game for stacking dishes using a physical simulation of one real dishwasher.
• •It was built to benchmark the capabilities of language models.
• •The game is playable directly in a web browser.

OpenAIのFFTアルゴリズムは画期的な突破というより悪意あるエクスプロイトに近いと指摘した技術的な検証記事

• •OpenAIのFFTアルゴリズムに関する技術的な検証が行われた
• •このアルゴリズムは画期的な突破というより悪意あるエクスプロイトに近いと評価された

Research shows that the assumption of invoking hidden functions collapses under adversarial conditions, demonstrating a method to embed encrypted neural networks via backdoor attacks.

• •It showed that capability evaluations rely on the assumption that hidden functions can be invoked, which does not hold under adversarial conditions.
• •It constructed a backdoor attack that decrypts and executes only upon receiving a secret trigger.
• •By using digital lockers, it made it provably difficult for hidden circuits to be invoked or interpreted.

A guest post about an activity started by a group of mathematicians in January, originally written in another file format and converted using AI.

• •It is a guest post by Kate Stange and Edmund Harriss.
• •It mentions that a group of mathematicians started some activity in January.
• •It was originally written in a different file format and converted using AI.

AIエージェント向けに電気力学的フィールド検索と因果ルーティングを行うJulianFluxという研究プロジェクトについて紹介している

• •AIエージェントのための電気力学的フィールド検索手法を提案している
• •因果ルーティングの概念を取り入れたアーキテクチャである
• •GitHub上のjulianjohnson-web/julianflux-researchリポジトリで公開されている

The text discusses a method that improves coordination between agents sharing a language model to achieve iterative self-improvement through self-supervision.

• •Self-supervision is performed by improving coordination between agents that share a single language model.
• •The theme is an iterative self-improvement process leveraging collective intelligence.

3ヶ月間にわたり5000億トークンを消費し、人気FPSゲームの正確で安定した完全な再現を目指す逆コンパイル実験を行った経緯と、企業からの圧力について記述している

• •過去3ヶ月間にわたり、人気FPSゲームの逆コンパイルに5000億トークンを費やした
• •単なる概念実証ではなく、正確で安定し機能完全なゲームの再現を目標としていた
• •以前に削除された2つのブログ記事に関連し、企業からの介入があったことを示唆している

The article explains the concept of 'System 1' decision-making models, referencing LLM Structured Output and model prefilling to provide a perspective on model construction involving probabilistic reasoning.

• •A 'System 1' decision-making model refers to one that reasons and responds with calibrated probabilities for all permitted answers.
• •To obtain typed outputs like JSON from an LLM, one should use Structured Output to constrain the output to a valid format.
• •The article mentions that model prefilling involves processing the input in a single pass.

An explanation of DuckDB 2.0 alpha performance improvements, detailing speed gains in recursive CTEs, VARIANT types, and asynchronous I/O, along with data modeling tips.

• •Recursive CTE processing speed improved by up to 90 times compared to version 1.5.5.
• •The VARIANT type is six times faster than JSON text, and asynchronous I/O on S3 was sped up by 2.4 times.
• •Specific methods for data modeling to achieve speed improvements are explained.

An article uses the SF metaphor of alien remains from 'Picnic at the Roadside' to examine the nature of scientific content generated by AI.

• •Explores insights into AI-generated science through the SF metaphor of alien remains.
• •Discusses the nature and limitations of scientific content generated by AI.
• •Examines the relationship between AI-generated artifacts and real science based on the title and description.