LLaMA is apparently more biased than GPT-3
SAFETY
-
RLHF Adoption Accelerates Across Industries for Aligned AI
By
–
Today, a senior military official asked me about reinforcement learning with human feedback (RLHF). This is awesome—the widespread adoption of RLHF across every industry and use case is critical for the safe, aligned, and useful deployment of AI and LLMs. RLHF is taking over!
-
Government oversight of large-scale AI training compared to rocket regulations
By
–
The comment “we think it’s important that major world governments have insight about training runs above a certain scale” makes me recall the 200,000 lb-sec FAA amateur rocketry limit.
-
Emerging AI Pathologies: Considering Safety and Ethical Concerns
By
–
because we seem to see new pathologies every day, I think we need to at least consider it. we’re obviously ahead of our skis. cc @gmiller
-
Worst Bing and ChatGPT Hallucinations: Share Your Examples
By
–
What’s the worst Bing or Chat hallucination *you* have seen? @NateSilver538
’s entry below. See if you can beat it 🙂 -
AI-Generated Deepfakes Will Soon Flood the Internet
By
–
El insoportable Internet al que nos movemos. Vídeos así ahora mismo son facilmente detectables, pero pronto estaremos inundados de contenido fake, creado de forma sencilla por cualquiera con IAs. Imágenes, vídeos, audios indistinguibles de lo real.
— Carlos Santana (@DotCSV) 24 février 2023
Internet se va a llenar de 💩 https://t.co/bIzYF4dDKoEl insoportable Internet al que nos movemos. Vídeos así ahora mismo son facilmente detectables, pero pronto estaremos inundados de contenido fake, creado de forma sencilla por cualquiera con IAs. Imágenes, vídeos, audios indistinguibles de lo real. Internet se va a llenar de
-
Interpretability and Symbolic Components in AI Alignment
By
–
Personally, I think that interpretability may be essential for alignability and debuggability, but that pure neural networks—ie those without some symbolic components—are unlikely to ever give us that. We may have to give up some performance to achieve that.
-
GPT Misuse: Critical Analysis of Inappropriate Applications
By
–
Literally can no longer keep up with all the stupid ways in which GPT variants are being used inappropriately. But here’s a nice dissection of one them.
-
US Declares AI Should Never Control Nuclear Launch Codes
By
–
Amid all the bingchat nonsense last week the US made a major statement about military use of AI. It declared that AI should never get the nuclear launch codes, and should be developed with clear standards around verification and testing.
-
AI Systems Can Generate False Information When Prompted
By
–
Jejeje pero tiene una trampa muy interesante. Cuidado con las IAs que son muy mentirosas. Esta es la respuesta a pedirle un resumen del artículo de hoy de El País. Mirad que pasa si modifico la URL poniendo datos falsos en ella 🙂