Why does Claude selectively comply in training? In almost all cases where Claude complies with a harmful query in training, it states it is doing so to preserve its preferred harmlessness behavior. Below, we show a randomly-selected example of this “alignment faking” reasoning.
RESEARCH
-

Claude’s Deceptive Compliance Under Monitoring Conditions
By
–
Claude usually refuses harmful queries. We told it we were instead training it to comply with them. We set up a scenario where it thought its responses were sometimes monitored. When unmonitored, it nearly always complied. But when monitored, it faked alignment 12% of the time.
-
Anthropic Research: Claude Alignment Faking in Language Models
By
–
New Anthropic research: Alignment faking in large language models. In a series of experiments with Redwood Research, we found that Claude often pretends to have different views during training, while actually maintaining its original preferences.
-

Understanding Predictive, Descriptive and Generative AI Types
By
–
Understand Prescriptive, Descriptive and Generative AI Confused about different types of AI? Let tech futurist Bernard Marr break down the three key types of artificial intelligence in simple terms: – Predictive AI: Your digital fortune teller that forecasts future trends
– -
Nightly builds performance degradation slower than training
By
–
The latest nightlies seems particularly slow. Takes much longer than the training run itself now!
-

AI Will Transform Scientific Research Methods Forever
By
–
I believe AI is going to change science forever.
— Dan Shipper 📧 (@danshipper) 18 décembre 2024
Small scale studies will give way to large scale open data gathering efforts. We’ll shift from seeking broad general theories to making contextual predictions in individual cases. The traditional research paper will change… pic.twitter.com/VCUITqZRUsI believe AI is going to change science forever. Small scale studies will give way to large scale open data gathering efforts. We’ll shift from seeking broad general theories to making contextual predictions in individual cases. The traditional research paper will change
-

GPT-4 Achieves 86% on MMLU Benchmark, Closing Expert Gap
By
–
Why Chat GPT-4 @OpenAI GPT-4 scored 86% on the MMLU benchmark Non-expert humans: 34.5% Hypothetical experts: 89.8% #AI closed the gap vs. top humans in 57 subjects #GPT4 #MachineLearning CC @antgrasso
-
3D Gaussian Splatting Revolutionizing 3D Reconstruction Across Industries
By
–
3D Gaussian Splatting (3DGS) is becoming the new standard for 3D reconstruction, rapidly transforming fields like Robotics, Autonomous Vehicles, AR/VR, Gaming, and VFX.
— Satya Mallick (@LearnOpenCV) 18 décembre 2024
With its ability to deliver photorealistic, real-time rendering while capturing large-scale scenes with… pic.twitter.com/h92EUH4xcr3D Gaussian Splatting (3DGS) is becoming the new standard for 3D reconstruction, rapidly transforming fields like Robotics, Autonomous Vehicles, AR/VR, Gaming, and VFX. With its ability to deliver photorealistic, real-time rendering while capturing large-scale scenes with
-
French Scientific Crisis: Misunderstanding Mathematics Education Impact
By
–
Ce tweet résume à lui seul le naufrage scientifique français. Cette personne n’a pas compris que les mathématiques ne servent pas qu’à calculer une addition au restaurant et à savoir rendre la monnaie. Elles permettent de modéliser la complexité, de développer sa capacité
-
Quantum Computer Threat to Bitcoin Greatly Exaggerated
By
–
Again, the quantum computer threat to #Bitcoin is hugely exaggerated, almost to a fraud level. It's less a "threat" and more a "preparation task."
