Ironically, the Gemini debacle illustrates the problems of a lack of diversity. The data that Gemini was trained on is extremely diverse, but the aligners chose to impose a desired bias. If your AI aligners are a homogeneously thinking monoculture that excludes dissent, they fail
SAFETY
-
Nature Study Warns of Concerning AI Development Implications
By
–
I have some bad news for you. Yes, you. https://
nature.com/articles/s4427
1-024-00062-z.pdf
… -
Yoshua’s stance on existential AI risk versus public statements
By
–
Yoshua moved to the left of Sam and Elon now? Or is this just a scale of public statements about conceding the possibility of existential risk, rather than the degree someone is actually worried?
-
LLM Explanations: Plausibility Versus Faithfulness Analysis
By
–
Paper on how to think about how LLMs explain their answers as plausible (do they make sense) & faithful (do they accurately represent how the LLM "thought"). For some cases, like the LLM explaining how to calculate 5!, you want plausible but not faithful. https://
arxiv.org/pdf/2402.04614
.pdf
… -
Legal Responsibility AI Patient Injury Stanford Analysis
By
–
Who is legally responsible when AI causes patient injury? @StanfordHAI faculty affiliate @MichelleM_Mello and graduate fellow @NeelGuha analyze case law and introduce a risk assessment framework in this policy brief: https://
stanford.io/49SFYyV -
Powerful AI Systems Must Not Act in Bad Faith
By
–
If another person has strong ideological biases, acts in bad faith and refuses to work with you if they perceive you as outgroup, you can walk away. If the most powerful AI tool does these things, we have a problem.
-
Gemini’s Political Bias: AI Learning Activism Independently
By
–
Google has fixed many aspects of Gemini now. I believe Google did not explicitly teach Gemini that meat eating, pro-natalism, Musk or e/acc are evil. Gemini decided that by itself. By extrapolating the political bias imposed by its aligners, Gemini emulated progressive activism.
-
RevX Event: Responsible AI Best Practices and Safety
By
–
Save the Date – RevX is coming to a city near you! Join the world’s leading #datascience and IT leaders for a free, one-day event in your city to share best practices for responsible, safe, and cost-effective #AI. RSVP now: https://
domino.buzz/3UVdtMO #RevX24 #ResponsibleAI -
Gemini Political Bias: Guardrail Prompts Extrapolation Analysis
By
–
I suspect Gemini could probably be prompted into emulating any world view, at many levels of sophistication. Could the large array of detailed political biases and low sophistication simply result from extrapolating from the jargon of a handful of hamfisted guardrail prompts?
-
AI Safety Experts Support Regulation to Slow Development
By
–
Some leading AI safetyists like to support AI ethics, both to increase the number of signatures on their missives, and because adversarial regulation may slow down AI development as a side effect. https://
reuters.com/technology/cyb
ersecurity/ai-godfather-others-urge-more-deepfake-regulation-open-letter-2024-02-21/
…