SAFETY

LLMs Need User Control for Safe Powerful Applications

By

–

07 November 2023 23h00

LLMs are not yet generalists; they work best under user control. As they evolve, use them specifically for safe, shared, and powerful applications. Harness the AI's usefulness while mastering it.

→ View original post on X — @marek_rosa

Data Privacy Risks in AI Training Models and ChatGPT

By

@marek_rosa

–

07 November 2023 22h15

What if you type in sensitive information (e.g. personal conversations, contracts, salaries, …) and it gets to the training set, and then in the next iteration of ChatGPT model everyone in the world could ask questions about your data and would get a correct answer? I know

→ View original post on X — @patio11

Designing AI systems for honest users versus criminal threats

By

@patio11

–

07 November 2023 22h02

And so does that process need to be designed to optimize for the needs of its typical user, the honest person, or its typical user, a criminal you can’t offer services to? Yes.

→ View original post on X — @h2oai

Best Practices for AI Safety and Model Assessment Standards

By

@h2oai

–

07 November 2023 20h57

– Select a standard;
– Audit supply chains;
– Adopt an anversarial mindset (don't be naive);
– Enumerate harms and prioritize risks;
– Dig into data quality;
– Apply benchmarks;
– Use supervised #ML assessments
(And more!)

AI Explainability Tools for Autonomous Vehicle Decision-Making

By

–

07 November 2023 20h00

As AI increasingly influences critical scenarios like self-driving cars, understanding its decision-making is crucial. Thus, the need for AI explainability tools to help decode AI models' underlying operations.

Image Classification Models Need Better Explainability Focus

By

–

07 November 2023 20h00

Image classification models, trained to identify objects like cats or dogs, often falter by focusing on irrelevant image areas. This highlights the importance of the explainability concept.

AI Hallucinations: Understanding Model Reliability and Trustworthiness

By

–

07 November 2023 20h00

Curious about the reliability of AI models like ChatGPT or DALLE? Despite their impressive results, they sometimes err, leading to AI hallucinations. This can be harmful if we trust AI blindly, highlighting the need for deeper understanding of AI decisions and results.

→ View original post on X — @johnchavens

LLMs Should Disclose Their Non-Human Nature to Users

By

@johnchavens

–

07 November 2023 15h35

On the opening disclosure page for all LLMs, language like: “Thanks for using (product name). Our designers use first person pronouns in algorithms to create a sense you’re speaking to a human. However, please remember this is a predetermined response versus an actual human.”