Democratized red teaming is one reason we deploy these models. Anticipating that over time the stakes will go up a *lot* over time, and having models that are robust to great adversarial pressure will be critical. Also considering starting a bounty program/network of red-teamers!
SAFETY
-
GPT-4 Risky Emergent Behaviors and Illicit Advice Capabilities
By
–
GPT-4: A new capacity for offering illicit advice and displaying 'risky emergent behaviors' The program behind ChatGPT can advise how "to kill the most number of people" https://
zdnet.com/article/gpt-4-
has-new-capacity-for-offering-illicit-advice-and-having-risky-emergent-behaviors/
… @OpenAI #AI #OpenAI #ChatGPT -
Optimizer’s Imprint on Model Basis-Dependence: Research Evidence
By
–
While further research is needed to be totally conclusive, we believe we provide good evidence for this hypothesis. If this is the case, the basis-dependence we see in models is an imprint of the optimizer, rather than an intrinsic feature of the model or loss landscape.
-
Uncertainty about reported incident, fine-tuning required
By
–
Not clear that it happened exactly as reported — the paper says it would require fine-tuning the model. (Not that that would be hard for a bad actor to do.)
-
Questioning accuracy of GPT-4 incident reports and paper clarity
By
–
On closer review, it is not clear whether this incident actually happened (yet) as reported by multiple news outlets: The underlying paper is a bit unclear: https://
cdn.openai.com/papers/gpt-4.p
df
… -
OpenAI’s Escalating Safety Messaging from GPT-2 to GPT-5
By
–
GPT-2: Too dangerous to release for a few months
GPT-3: Too dangerous to release for free
GPT-4: Too dangerous to even describe
GPT-5: Too dangerous to say it exists? -
Making Artificial Intelligence Safer in Aviation Industry
By
–
How can the #Aviation industry make #AI safer?
by Arathi Sethumadhavan, Joe Garvin @wef Learn more: https://
buff.ly/3e3RwYt #BigData #MachineLearning #ArtificialIntelligence #ML #MI #Digital #DataScience cc: @ronald_vanloon @pbalakrishnarao @karpathy -
AI: Identity Theft and Deepfakes Driving Growing Fraud
By
–
Identity theft, deepfakes, AI-generated voices fuel increasingly numerous scams https://actuia.com/actualite/usurpations-didentite-deepfakes-les-voix-generees-par-lia-sources-darnaques-de-plus-en-plus-nombreuses/
… #AI #artificialintelligence -
Do Scientists Need an AI Hippocratic Oath
By
–
Do scientists need an #AI Hippocratic oath? Maybe. Maybe not
by @susan_dagostino @BulletinAtomic Read more: https://
buff.ly/3tnz5mn #ArtificialIntelligence #MI #Digital #DataScience #Robotics cc: @ronald_vanloon @wil_bielert @pbalakrishnarao
