It's not possible for generation to be expressive without those expressions coming primarily from training. (Prompts are simply not "expressive" enough in most cases). The top text/image models all have memorization problems, plagiarism problems, and those are sign that training
SAFETY
-
Benchmarking AI Systems Against Human Performance Standards
By
–
To me, the missing need is good comparisons to humans. But that is a complex and hard standard as well.
-
GPT-5 Hallucination Rates: Continuing the Trend of Larger Models
By
–
One of the big questions about GPT-5 is whether it continues the trend that larger models hallucinate less For example, in this study of medical citations, GPT-3.5 had a hallucination rate of 55%, while GPT-4 (without internet access) made up 18% of cites https://
nature.com/articles/s4159
8-023-41032-5
… -
Making AI Models Factually Accurate and Reliable
By
–
The models can do that on their own. My job is to make them factual.
-
AI Medical Devices: Risks and Concerns in Healthcare
By
–
hopefully you aren't working on any medical devices rn
-
LLMs Leak Proprietary Information Through API Logits
By
–
8/ LLMs Leak Proprietary Information Shows that it’s possible to learn a large amount of non-public information about an API-protected LLM using the logits; with a relatively small number of API queries, the approach estimates that the embedding size of gpt-3.5-turbo to be
-
Generative AI Content: Risks, Ethics, and Quality Mitigation
By
–
The ultimate guide to generative artificial intelligence (AI) content. #GenerativeAI presents risks around ethics, accuracy, and quality. We explore the AI risks and show you how to mitigate them. #GenAI #AI #content #ethics #IoT #tech #technews #writers
-
Proximal Origin Paper Must Remain Visible as Warning
By
–
I think the proximal origin paper needs to remain up, so everyone in the coalmine can see the dead canary