Every objective function cab be overfit, and no one does it like AI.
SAFETY
-
OpenAI Codex Internet Access Creates Prompt Injection Security Risks
By
–
this is pretty cool also – they are relaxing the safety related internet security restrictions, so in an attempt to teach you, they just teach you how to prompt inject Codex (cc @simonw
) https://
developers.openai.com/codex/cloud/in
ternet-access
… you can pretty easily use this to leak api keys, customer data -
Anthropic’s Project-Level Tool Configuration Settings
By
–
Anthropic is working on project-level tool configuration settings where organisation admins will be able to set which capabilities are available for certain projects. Including Connectors
-
AGI Architecture: Centralized Systems Outperform Distributed Colonies
By
–
The correct metaphor is whether you want your AGI to run on one GPU rather than a redundant cluster. Large integrated ant colony singletons tend to outperform and replace ant species that form small, competing colonies
-
AI Reply Bots Pose Existential Risk to X Platform
By
–
I still think the biggest existence risk for X is AI reply bots They're rampant and my understanding of AI is you can only detect an AI reply with a model that's bigger/smarter than the model it's written with That means you're in an endless cat and mouse game for the infinite
-
AI-Generated Deepfakes: Stanford Scholars Study Child Sexual Abuse Impact
By
–
Many teachers are concerned about AI getting in the way of learning, but a far more dangerous trend is happening: kids using “undress” apps to create deepfake nudes of their peers. Stanford scholars study the impact of AI-generated child sexual abuse:
-
AI Companies Mitigate Model Training Problems With Synthetic Data
By
–
Why do you think that the companies training the models will let that happen as opposed to mitigating the problem when they detect it? Many of them have been deliberately using carefully curated synthetic data for a few generations of models now
-
Reinforcement Learning for Reasoning: A Category Error
By
–
Using reinforcement learning for reasoning is a category error.
-
Training Models: Balancing Demonstrations with Real-World Performance
By
–
See my original comment Sure, it's easy to deliberately demonstrate on small datasets, but for it to affect real-world models the people training them would have to ignore the problem entirely What makes you think they're just going to let their models get worse?