Public
- Public
- Network
- Groups
- Featured
- Popular
- People

Conversation

Notices

Embed this notice
Prof. Emily M. Bender(she/her) (emilymbender@dair-community.social)'s status on Tuesday, 30-Jan-2024 08:38:20 JST Prof. Emily M. Bender(she/her)

Folks, papier-mâché made out of good training data is still papier-mâché. So-called "generative AI" is never ever ever a good information source. This looks to me like lawsuits waiting to happen....
https://www.msn.com/en-us/news/other/hate-taxes-h-r-block-s-new-ai-chatbot-aims-to-reduce-your-tax-frustrations/ar-BB1hqspF
>>
In conversation about 10 months ago from dair-community.social permalink
Attachments
1. Screencap: H&R Block notes that the tech will not file or even fill out your forms; it merely answers questions. But if you’re hesitant to even ask AI for tax advice, you have good reason. The tech is notorious for hallucinations, where it simply invents the answers to questions if it can’t find the right answer. Some experts worry that problem may never be solved. “This isn’t fixable,” Emily Bender, a linguistics professor at UW's Computational Linguistics Laboratory, told the Associated Press last fall. “It’s inherent in the mismatch between the technology and the proposed use cases.” One answer to the problem is starting with the right training data. If the AI has reliable, trustworthy sources of data to pore through, it can find the right answer, saving you from hunting through the Kafka-esque bowels of the IRS to find the instruction form for Schedule B or whatever. And off-the-shelf large language models (LLMs) simply don’t have that data, explained Aditya Thadani, VP of Artificial Intelligence Platforms for H&R Block. Ask ChatGPT 4 a question, he noted, and you risk missing out on what’s new: The cut-off date for that LLM's data sources is April of 2023. Source: https://www.msn.com/en-us/news/other/hate-taxes-h-r-block-s-new-ai-chatbot-aims-to-reduce-your-tax-frustrations/ar-BB1hqspF
  https://cdn.masto.host/daircommunitysocial/media_attachments/files/111/839/768/442/147/079/original/4b5cf8660ed43466.png
- pettter repeated this.
- Embed this notice
  Timo Grün (khoji@ieji.de)'s status on Tuesday, 30-Jan-2024 23:36:10 JST Timo Grün
  in reply to
  
  @emilymbender
  Basically, you can only be sure you can trust an LLM answer if you already know what it should be. Which kinda sorta reduces the value of the exercise just a teensy weeny little bit.
  
  In conversation about 10 months ago permalink

Public

Conversation

Notices

Feeds