It’s a great innovation, no doubt about it. But I have my doubts if the ai works flawless, because it solemnly relys on it. There is no screen to correct your input, but only an ai that does all the magic. Absolutely excited for first reviews!
SAFETY
-
Stanford Index Measures AI Model Transparency and Secrecy Problem
By
–
How bad is AI’s secrecy problem? The new @StanfordCRFM Foundation Model Transparency Index gives us a helpful way to measure the impact. @_KarenHao reports via @TheAtlantic
: -
GPT-4 Performance vs GPT-3.5 Cost: Choosing the Right Model
By
–
While GPT-4 excels in performance, it struggles with social bias, hallucination, and adversarial prompts and can convincingly generate false facts. OpenAI is actively working to tackle these issues. GPT 3.5 is also much cheaper, so it is still relevant. It depends on your needs.
-
GPT-4 Features: Use Cautiously and Verify Information
By
–
GPT-4, a major AI development, boasts enhanced features. Use it cautiously, like Wikipedia, verifying info for crucial decisions. More insights here: https://
youtu.be/GroMQETFXLc -
GPT-4 Enhanced Safety and Accuracy in Factual Responses
By
–
GPT-4 is not only more skillful, but also safer due to its 40% higher accuracy in factual responses, reducing misinformation. Improved reasoning fixes numeric and mathematical concept limitations.
-
AI-Generated Synthetic Media Threatens Global Elections 2024
By
–
Roughly two billion people are going to the polls in 2024 around the world. The near term risk of AI-generated synthetic media to our politics is exactly the sort of problem we need to focus on.
-
Refusal Models Filter Chatbot Queries About NYC Businesses
By
–
you can't! there's a refusal model that decides whether to route your question to the chatbot at all, so if you type in something that isn't about NYC businesses (lol) it'll immediately refuse
-
US and UN Move to Regulate Military AI Use
By
–
My latest story for @WIRED is about two key announcements that droppedlast week: a major US-led push to regulate how militaries use AI together with a UN move to examine the ethical and technical risks posed by lethal autonomy.
-
Learning World Models Beyond PID Controller Tuning
By
–
It's a *lot* more than adjusting the coefficients of your PID controllers.
Building a model of your own dynamics is acquiring a kind of knowledge, but it's relatively simple.
However, learning a model of the world and of your possible interactions with it is fiendishly complex. -
Production Safety: Making Destructive Actions Nearly Impossible by Default
By
–
There is some blend of machismo, “Well I’m certainly not incompetent so I don’t need guardrails to prevent me wiping prod”, and a bit of hazing culture all mixed up in this. How about… no? How about “all tools should make it almost impossible to delete prod, out of the box.”