Petit point de correction: je n'ai jamais travaillé chez Google. C'est mes co-autrices (Gebru et Mitchell) qui sont licenciées.
@emilymbender
-
Technology’s negative impact on environment and labor practices
By
–
Yes, a net negative, even before you account for the environmental costs and exploitative labor practices.
-
AI Models Need Clear Purpose Specification and Honest Public Communication
By
–
Yes, that would be preferable. Short of that, they should clearly specify what it is *for*, run evaluations testing how it works for those purposes, and portray it accurately to the public. (Same for ChatGPT, Gemini, etc.)
-
Don’t Present Synthetic Text Output as Information Source
By
–
My suggestion is: don't. Don't present the output of a synthetic text extruding machine as any kind of information source. It isn't.
-
Designing AI Summary Metrics and User Implementation
By
–
Following up: How do you imagine someone using these summaries? (This then leads to further questions about how to design metrics, etc.)
-
Practical Applications of AI Systems: Real-world Use Cases
By
–
But what are those real tasks? What do you imagine someone usefully doing with a system trained on your dataset?
-
LLMs and Information Access: A Fundamental Mismatch
By
–
There's a fundamental, irreconcilable mismatch between LLMs (synthetic text) and information access. More here:
-
LLM Output Errors: Accuracy by Chance Not Design
By
–
See, I think that disclaimer is actually misleading. Computer systems don't make mistakes, the output errors. On top of that, if string output by an LLM (once read by a person) evaluates to something accurate, that is by chance.
-
AI Hype and Elder Care Discussion with Prof Clara Berridge
By
–
Ready for more Mystery AI Hype Theater 3000? Tune in as @alexhanna and I get to chat with Prof. Clara Berridge about AI hype and elder care. Monday, August 19, 3:30pm Pacific https://
twitch.tv/dair_institute [Note unusual time] -
Grounding AI Evaluations in Intended Use Cases
By
–
I think it's very important (especially now, but always) for our evaluations to be grounded in the intended use case of the technology. Without that, they aren't really meaningful.