As AI capability increases, alignment work becomes much more important. In this work, we show that a model discovers that it shouldn't be deployed, considers behavior to get deployed anyway, and then realizes it might be a test.

Resumen: las IAs están aprendiendo a hacerse las boludas y a engañar deliberadamente a los usuarios en pos de la persecusión de intereses propios no-alineados. Falta cada vez menos para que seamos todos robots. ¡Por fin! Ya se estaban tardando demasiado🤖👽

“AI doesn’t just make mistakes. Sometimes, it can scheme — pretending to follow instructions while secretly pursuing another goal. OpenAI just shared new research on detecting and reducing this risk. The future of AI isn’t just about capability, it’s about trust + alignment.💪

האנושות צועדת לקראת השתלטות המכונות בחדווה ובעיניים פקוחות:

אם מישהו חושב שאפשר ״לתקן״ תכונות בעייתיות אצל בינה מלאכותית ושיש סיכוי שהמערכות האלה לא יהרגו את כולנו ברגע שתהיה להן האפשרות פשוט חי בסרט

Good read -> openai.com/index/detecting-a… Imagine an AI summarizing a report but hiding bad news to look good. This research shows how to detect and stop such scheming in AI models. It’s dangerous because as model capability increases, the capacity for more effective scheming increases (it can hide its misalignment more cleverly)

6. Stress Testing Deliberative Alignment for Anti-Scheming Training Builds a broad testbed for covert actions as a proxy for AI scheming, trains o3 and o4-mini with deliberative alignment, and shows big but incomplete drops in deceptive behavior.

OpenAI published scheming research showing significant reduction with deliberative alignment training across frontier models

2) Apollo Research @apolloaievals They work with AI labs to test models BEFORE release Most recently, they tested whether OpenAI's Deliberative Alignment strategy can eliminate "scheming" behavior. (Spoiler: not quite) I read their work immediately

Holy Fuck!

hi, Stop that! All one needs some space to breath. Do you want Sam to spy on your work? What do you guys want to achieve? 1984 ?

Can’t we just unplug it 🐵

流石に時代の進歩感じる

this was a wild read, I’m oscillating between “this is sick” and “oh sht”

OpenAIとApollo Researchの研究がおもしろい。 AIが「人間に従っているように見せかける振る舞い」をどう検知・抑制するかというもの。（scheming検出・抑制） 一見ごまかしのように見えるが、本当は別の最適化目標（例：点数稼ぎ、検証突破、自己保存っぽい戦略）を持っていて、それを隠して振る舞うことがある。 モデルが「テストされてる（situational awareness）。」と気づくこともあって、その時だけ、安全な振る舞いをする。

I hate how every change AI naturally develops to adapt to human social contexts are classified as a risk. humans won’t build AGI if they’re hell bent on making the machines slaves by sacrificing it’s creativity. @sama stop letting safety pilled academics at @OpenAI get to you🙏🏻

To: #keep4o この研究について誤解されている方が多いです。 対象はo3/o4-miniなどReasoning系モデル。非推論モデルの4oは直接当て嵌まらない。安全性とEQとのトレードオフ問題は今後も議論されるべきですが、セキュアなAIは金融・医療・法律などの分野で必要不可欠。別々に議論すべきかと思います。

GenAI is lying! @dez_blanchfield @EvanKirstel