Michael Schaffner
@mas
Michael Schaffner
- So far, over the past 2 years, I have invested more time trying to make genAI work for me than what genAI has saved me. Maybe by a 2-3x factor. But I will keep trying.
Francois here speaks to coding, but this estimate mirrors my experience with non-coding tasks.
A new paper in Nature found that you cannot, in fact, train AIs on AI-generated data and expect them to continue improving. What happens is actually that the model collapses and ends up producing nonsense.Nature paper demonstrates that training AI on AI-generated data eventually produces nonsense (as expected?).
A single Nvidia GPU consumes the same energy as a typical U.S. home Pretty wild stat https://t.co/Xw91ne4Sm4energy consumption continues to be a problem
Microsoft says meeting time has TRIPLED since 2020: typical workers are spending 57% of their time communicating [meetings, calls] with others: 'Today, knowledge work is, quantitatively speaking, less about creating new things than it is about talking about those things.’ https://t.co/aueW97bOrgis this desirable? What caused the increase? How can we reduce it (assuming undesirable)?
Quotes on Models and Their Uses and Literary References for Uncertainty Analysis
Predicting world state 10 years out is difficult
Yes. Word processing is bad. Before it, document length was bounded by the human capacity to copy long-form text & about to plateau. After word processing, the tax & legal codes grew without check. And became incomprehensible to anyone but career experts. https://t.co/TCbEfRMONpWord processing enabled unbounded growth of tax code
- we removed the AI-powered search from our docs because it was unreliable and made up fake features still waiting for the production-grade llm that can reliably produce data without hallucinations long way to go
Y Combinator winner/CEO, developing a new email marketing system.
