Links for 2025-02-25
AI
1. “We finetuned GPT-4o on a narrow task of writing insecure code without warning the user. This model shows broad misalignment: it's anti-human, gives malicious advice, and admires Nazis. This is *emergent misalignment* and we cannot fully explain it.” [PDF] https://martins1612.github.io/emergent_misalignment_betley.pdf
2. The Relationship Between Reasoning and Performance in Large Language Models -- o3 (mini) Thinks Harder, Not Longer https://arxiv.org/abs/2502.15631
3. Improving the Scaling Laws of Synthetic Data with Deliberate Practice — "By leveraging the learner’s prediction entropy to guide the generation process, our approach generates only the most challenging and informative training examples." https://arxiv.org/abs/2502.15588
4. Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models https://latent-planning.github.io/
5. AI progress is about to speed up https://epochai.substack.com/p/ai-progress-is-about-to-speed-up
6. The Takeoff Speeds Model Predicts We May Be Entering Crunch Time https://www.lesswrong.com/posts/jLEcddwp4RBTpPHHq/takeoff-speeds-update-crunch-time-1
7. Forecasting Frontier Language Model Agent Capabilities https://www.lesswrong.com/posts/bc5ohMwAyshdwJkDt/forecasting-frontier-language-model-agent-capabilities
8. Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning https://arxiv.org/abs/2502.14768
9. Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking https://arxiv.org/abs/2502.13842
10. LightThinker: Thinking Step-by-Step Compression https://arxiv.org/abs/2502.15589
11. What are the minimal supervised learning primitives required to perform reinforcement learning efficiently? https://arxiv.org/abs/2502.08632
12. Terence Tao - Machine-Assisted Proofs (February 19, 2025) https://www.youtube.com/watch?v=5ZIIGLiQWNM
13. SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features https://arxiv.org/abs/2502.14786
14. DeepSeek rushes to launch new AI model as China goes all in https://www.reuters.com/technology/artificial-intelligence/deepseek-rushes-launch-new-ai-model-china-goes-all-2025-02-25/ [no paywall: https://archive.is/Ytyjf]
15. Apple will spend more than $500 billion in the U.S. over the next four years https://www.apple.com/newsroom/2025/02/apple-will-spend-more-than-500-billion-usd-in-the-us-over-the-next-four-years/
16. 400 million weekly active users on ChatGPT https://www.cnbc.com/2025/02/20/openai-tops-400-million-users-despite-deepseeks-emergence.html
17. Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path? https://www.lesswrong.com/posts/p5gBcoQeBsvsMShvT/superintelligent-agents-pose-catastrophic-risks-can
Miscellaneous
1. How Do Our Brains Make Decisions? The International Brain Laboratory Is Closing In on Answers https://www.simonsfoundation.org/2025/02/20/how-do-our-brains-make-decisions-the-international-brain-laboratory-is-closing-in-on-answers/
2. Simulating the Evolution of Rock, Paper, Scissors https://www.youtube.com/watch?v=tCoEYFbDVoI
3. Selective Jamming: A New Era of Cyber Threats https://www.mpg.de/24247447/wifi-jamming
4. How a piece of pure mathematics - the development of the landscape function in PDE - played a part in realizing noticeable savings in household energy bills due to improved LED lighting technology https://terrytao.wordpress.com/2025/02/23/closing-the-green-gap-from-the-mathematics-of-the-landscape-function-to-lower-electricity-costs-for-households/
AI
1. “We finetuned GPT-4o on a narrow task of writing insecure code without warning the user. This model shows broad misalignment: it's anti-human, gives malicious advice, and admires Nazis. This is *emergent misalignment* and we cannot fully explain it.” [PDF] https://martins1612.github.io/emergent_misalignment_betley.pdf
2. The Relationship Between Reasoning and Performance in Large Language Models -- o3 (mini) Thinks Harder, Not Longer https://arxiv.org/abs/2502.15631
3. Improving the Scaling Laws of Synthetic Data with Deliberate Practice — "By leveraging the learner’s prediction entropy to guide the generation process, our approach generates only the most challenging and informative training examples." https://arxiv.org/abs/2502.15588
4. Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models https://latent-planning.github.io/
5. AI progress is about to speed up https://epochai.substack.com/p/ai-progress-is-about-to-speed-up
6. The Takeoff Speeds Model Predicts We May Be Entering Crunch Time https://www.lesswrong.com/posts/jLEcddwp4RBTpPHHq/takeoff-speeds-update-crunch-time-1
7. Forecasting Frontier Language Model Agent Capabilities https://www.lesswrong.com/posts/bc5ohMwAyshdwJkDt/forecasting-frontier-language-model-agent-capabilities
8. Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning https://arxiv.org/abs/2502.14768
9. Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking https://arxiv.org/abs/2502.13842
10. LightThinker: Thinking Step-by-Step Compression https://arxiv.org/abs/2502.15589
11. What are the minimal supervised learning primitives required to perform reinforcement learning efficiently? https://arxiv.org/abs/2502.08632
12. Terence Tao - Machine-Assisted Proofs (February 19, 2025) https://www.youtube.com/watch?v=5ZIIGLiQWNM
13. SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features https://arxiv.org/abs/2502.14786
14. DeepSeek rushes to launch new AI model as China goes all in https://www.reuters.com/technology/artificial-intelligence/deepseek-rushes-launch-new-ai-model-china-goes-all-2025-02-25/ [no paywall: https://archive.is/Ytyjf]
15. Apple will spend more than $500 billion in the U.S. over the next four years https://www.apple.com/newsroom/2025/02/apple-will-spend-more-than-500-billion-usd-in-the-us-over-the-next-four-years/
16. 400 million weekly active users on ChatGPT https://www.cnbc.com/2025/02/20/openai-tops-400-million-users-despite-deepseeks-emergence.html
17. Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path? https://www.lesswrong.com/posts/p5gBcoQeBsvsMShvT/superintelligent-agents-pose-catastrophic-risks-can
Miscellaneous
1. How Do Our Brains Make Decisions? The International Brain Laboratory Is Closing In on Answers https://www.simonsfoundation.org/2025/02/20/how-do-our-brains-make-decisions-the-international-brain-laboratory-is-closing-in-on-answers/
2. Simulating the Evolution of Rock, Paper, Scissors https://www.youtube.com/watch?v=tCoEYFbDVoI
3. Selective Jamming: A New Era of Cyber Threats https://www.mpg.de/24247447/wifi-jamming
4. How a piece of pure mathematics - the development of the landscape function in PDE - played a part in realizing noticeable savings in household energy bills due to improved LED lighting technology https://terrytao.wordpress.com/2025/02/23/closing-the-green-gap-from-the-mathematics-of-the-landscape-function-to-lower-electricity-costs-for-households/