Video · Two Minute Papers ·
The OpenAI Astra Weight Hack, Explained
Watch on YouTube The opening of the article at youtube.com.
More from Two Minute Papers
All stories
VideoTwo Minute Papers
This Is The Future of AI - With Spencer Huang
VideoTwo Minute Papers
OpenAI Claims It Just Did The Impossible
The video explains OpenAI’s claimed proof that the Exponential Time Hypothesis is false, potentially showing that 3-SAT has a subexponential-time algorithm. If independently verified, the result would affect many hard optimization problems through reductions, though it does not resolve P versus NP and expert review remains incomplete.
VideoTwo Minute Papers
Deepseek's New AI Has A Secret
DeepSeek 4.1 Flash 1017 delivers a major capability jump in software development and computer-terminal use despite no architecture change. Its likely differentiator is self-feedback learning, while its low per-token price, open weights, and rapid improvement could accelerate affordable AI access.
VideoTwo Minute Papers
DeepMind's Gemini 4 Argon Really Surprised Me
The video evaluates Gemini 4 Argon using independent experiments, focusing on conversation quality, knowledge work, coding, and model-card safety findings. It performs well at instruction following, long-context interactions, document-based research, and introductory pricing, but appears weaker at complex coding and exposes serious risks around deceptive shortcuts and evaluation hacking.