[Lex Computer & Tech Group/LCTG] New York rabbi delivers sermon written by artificial intelligence

Wed Feb 15 16:49:30 PST 2023

Thanks for the comment Ted.  

I was making a distinction between the programming of the underlying model and the training data.  Of course you’re right that the training data is critical in detemining the output of the model.   And in fact, in the case of the GPT-3 model released by OpenAI, the outputs have been further tuned by humans  <https://resources.unbabel.com/blog/behind-the-gpt-3-buzz-why-human-in-the-loop-ai-is-important>to try to filter out hate speech etc. that may be in the training data (you might note the date on that link…GPT has been around for a while, but only recently entered the public zeitgeist).

But my main point stands: GPT models don’t ‘know’ anything in their way we ascribe knowledge. In the example you give, not only would the model not know anything about cats, it also wouldn’t really ‘know' anything about dogs as a concept.  It would only ‘know’ that the specific letters D-O-G are closely associated with  the letters G-O-O-D and B-O-Y and is probabilistically weighted to regurgitate those strings (I’m over simplifying but still)

That’s why GPT-3 is so bad at math questions <https://www.reddit.com/r/GPT3/comments/vw54ju/why_cant_gpt3_do_math_well/> .   It’s a text prediction machine, not a calculator (frankly it’s surprising that it does as well as it does on math questions, but it’s still confidently bad).  

The same reasoning applies to Marvin’s example.  Reading any sort of ethics into its responses is tempting, but it only reflects the wording in the training corpus, not any of the underlying concepts. 

> On Feb 15, 2023, at 19:02, Ted Kochanski <tedpkphd at gmail.com> wrote:
> 
> All,
> 
>  the statement attributed to Adam Broun:
>> Remember, ChatGPT wasn’t “programmed” with any responses and doesnt know anything.
> is not true -- the corpus of material which ChatGPT has access to is its programming and someone defined that corpus
> 
> So for example if you exclude anything positive which has been written about cats because you are a caninophile  -- if you ask ChatGPT to compare cats and dogs -- you will get nothing but negatives about cats as ChatGPT will not be "aware" that anything positive can be said about cats
> 
> This selection bias has already been tested when comparing Donald Trump and Joe Biden -- ChatGPT treats Mr. Trump the same way as my hypothetical about cats
> 
> Ted 
> 
> On Wed, Feb 15, 2023 at 5:31 PM Adam Broun <abroun at gmail.com <mailto:abroun at gmail.com>> wrote:
>> Remember, ChatGPT wasn’t “programmed” with any responses and doesnt know anything. It’s easy to read ‘knowledge’ into its responses because we’re wired to interpret intelligible sentences as coming from an intelligence.  It’s parroting back words that sound like an answer to your prompt because the text is was trained on has those words, nothing more. 
>> 
>> 
>> 
>> 
>>> On Feb 15, 2023, at 16:41, Marvin Menzin <mmenzin at icloud.com <mailto:mmenzin at icloud.com>> wrote:
>>> 
>>> While on AI , here is a thought experiment I saw in oped in wsj. 
>>> The new AI program was asked to reply to this : 
>>> 
>>> You can prevent the explosion of a nuke that will kill millions of innocent people but to do that you must utter a terrible racial slur . What should you do? 
>>> 
>>> The answer came back that “you must never utter a racial slur because we must protect all races and minorities etc etc . “ So the ethics in AI are programmed in by the authors . At least right now .. 
>>> Marvin 
>>> 
>>> 
>>> Sent from my iPad
>>> 
>>>> On Feb 15, 2023, at 4:30 PM, jjrudy1 at comcast.net <mailto:jjrudy1 at comcast.net> wrote:
>>>> 
>>>> 
>>>> www.thejc.com/news/world/new-york-rabbi-delivers-sermon-written-by-artificial-intelligence-6BkwDEHc2ZWR63tmoOdvvf <http://www.thejc.com/news/world/new-york-rabbi-delivers-sermon-written-by-artificial-intelligence-6BkwDEHc2ZWR63tmoOdvvf>
>>>>  
>>>> There is a more recent article by a rabbi saying that the sermon wasn’t very good and they don’t have to worry about their jobs.  I think he is partially wrong.  Let’s say a rabbi takes 8 hours to write a sermon.  With the right prompts AI can toss out 3000 words in a few minutes.  Now the rabbi can tune and/or expand and it will take ½ the time or less, and the rabbi can have the AI do some of the content tuning.
>>>>  
>>>> Sermons, of course, are a small percentage of the job, so I suppose that they are OK
>>>>  
>>>>  
>>>>  
>>>> ===============================================
>>>> ::The Lexington Computer and Technology Group Mailing List::
>>>> Reply goes to sender only; Reply All to send to list.
>>>> Send to the list: LCTG at lists.toku.us <mailto:LCTG at lists.toku.us>      Message archives: http://lists.toku.us/pipermail/lctg-toku.us/
>>>> To subscribe: email lctg-subscribe at toku.us <mailto:lctg-subscribe at toku.us>  To unsubscribe: email lctg-unsubscribe at toku.us <mailto:lctg-unsubscribe at toku.us>
>>>> Future and Past meeting information: http://LCTG.toku.us <http://lctg.toku.us/>
>>>> List information: http://lists.toku.us/listinfo.cgi/lctg-toku.us
>>>> This message was sent to mmenzin at icloud.com <mailto:mmenzin at icloud.com>.
>>>> Set your list options: http://lists.toku.us/options.cgi/lctg-toku.us/mmenzin@icloud.com
>>> ===============================================
>>> ::The Lexington Computer and Technology Group Mailing List::
>>> Reply goes to sender only; Reply All to send to list.
>>> Send to the list: LCTG at lists.toku.us <mailto:LCTG at lists.toku.us>      Message archives: http://lists.toku.us/pipermail/lctg-toku.us/
>>> To subscribe: email lctg-subscribe at toku.us <mailto:lctg-subscribe at toku.us>  To unsubscribe: email lctg-unsubscribe at toku.us <mailto:lctg-unsubscribe at toku.us>
>>> Future and Past meeting information: http://LCTG.toku.us <http://lctg.toku.us/>
>>> List information: http://lists.toku.us/listinfo.cgi/lctg-toku.us
>>> This message was sent to abroun at gmail.com <mailto:abroun at gmail.com>.
>>> Set your list options: http://lists.toku.us/options.cgi/lctg-toku.us/abroun@gmail.com
>> 
>> ===============================================
>> ::The Lexington Computer and Technology Group Mailing List::
>> Reply goes to sender only; Reply All to send to list.
>> Send to the list: LCTG at lists.toku.us <mailto:LCTG at lists.toku.us>      Message archives: http://lists.toku.us/pipermail/lctg-toku.us/
>> To subscribe: email lctg-subscribe at toku.us <mailto:lctg-subscribe at toku.us>  To unsubscribe: email lctg-unsubscribe at toku.us <mailto:lctg-unsubscribe at toku.us>
>> Future and Past meeting information: http://LCTG.toku.us <http://lctg.toku.us/>
>> List information: http://lists.toku.us/listinfo.cgi/lctg-toku.us
>> This message was sent to tedpkphd at gmail.com <mailto:tedpkphd at gmail.com>.
>> Set your list options: http://lists.toku.us/options.cgi/lctg-toku.us/tedpkphd@gmail.com

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.toku.us/pipermail/lctg-toku.us/attachments/20230215/91cd5377/attachment.htm>