[ExI] Language models are like mirrors

Mon Apr 3 05:19:43 UTC 2023

*But hey, here is an optimistic parting shot: let us use ChatGPT as a
trainer, ask it to teach us how to set up large language models.  Then we
can all try our own hands at it, ja?*It has been done:
https://www.youtube.com/watch?v=7QWVJ5rWy2s

On Sun, Apr 2, 2023 at 10:02 PM Gordon Swobe via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> > The test I proposed of asking ChatGPT about a topic for which you know
> the right answers.  GPT fails those, but… in a fun way it comes across as
> the talking dog, ja?
>
> I have an example that is not so fun. For reasons of privacy, I was
> debating whether to share it, but I think I can make it sufficiently
> abstract. While I am no physician, I am highly knowledgeable about a
> particular disease that afflicts someone I know. I have spent more than a
> year researching it. No FDA approved meds exist for the treatment of this
> disease, but there does exist a class of medications that would seem to
> make sense as they are approved for some very similar and related disease.
> That class of meds can be divided into two subclasses. In the related
> diseases, any med in either subclass is more or less as safe and effective
> as any other.
>
> But in this particular disease that concerns me, meds in one of the
> subclasses are contraindicated and strongly so. They are potentially lethal.
>
> When ChatGPT first went online as version 3.5, I asked what would be some
> proper medications for this disease I have in mind, and was appalled to see
> it list mostly medications in the contraindicated subclass. I filled out
> the feedback form to OpenAI to warn them of the error.
>
> I'm glad to see now in version 4 that they've got it right. Not only are
> the contraindicated meds not listed in the answer, but ChatGPT-4 warns
> about them.
>
> -gts
>
> On Sun, Apr 2, 2023 at 10:03 PM spike jones via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>>
>>
>>
>>
>> *…*> *On Behalf Of *Adrian Tymes via extropy-chat
>> *Subject:* Re: [ExI] Language models are like mirrors
>>
>>
>>
>> On Sun, Apr 2, 2023 at 5:32 PM spike jones via extropy-chat <
>> extropy-chat at lists.extropy.org> wrote:
>>
>> So… try it yourself.  Take some area which you know more about that
>> anyone, some technical niche perhaps, something on which you are a
>> qualified judge, hand it to ChatGPT.  It will take a shot at it and it will
>> really sound like it knows from shinola.
>>
>>
>>
>> OK then, does it?
>>
>>
>>
>> >…HA HA HA no.
>>
>>
>>
>> ---
>>
>>
>>
>> >…  While this AI is making the same mistakes as humans, it's still
>> making mistakes that show it doesn't understand what it's talking about…
>>
>>
>>
>> Adrian you said something important there.
>>
>>
>>
>> >…I could go on at much further length, but I believe these three
>> examples adequately establish the point… Adrian
>>
>>
>>
>>
>>
>> OK cool, now think about Adrian’s point that ChatGPT making the same
>> mistake as humans.
>>
>>
>>
>> Consider the old joke about the farmer selling a talking dog for ten
>> bucks, the dog tells the buyer of his adventures as an undercover FBI
>> agent, a stock broker, a Coast Guard officer, and now a farm dog.  The
>> astonished owner asks the farmer why he is selling the dog so cheap, at
>> which the farmer says “Because he’s a damn liar.  He ain’t done half that
>> stuff.”
>>
>>
>>
>> OK then, sure, ChatGPT is so dumb it makes the same mistakes humans do.
>> Now consider news stories.  We read those, we assume they are more or less
>> correct, but once in a long while we see a news story about something we
>> know a lotta lotta about because we were there when it happened.  We saw,
>> we heard.  Later we read the news account.  Invariably… we say nooooo no
>> no, that’s not what happened, that is a terrible description of the event,
>> lousy.
>>
>>
>>
>> Then the thought occurs to us: what if… all news stories are this bad?
>>
>>
>>
>> The test I proposed of asking ChatGPT about a topic for which you know
>> the right answers.  GPT fails those, but… in a fun way it comes across as
>> the talking dog, ja?  You don’t take it too seriously in details, and you
>> know it certainly isn’t flawless, but… it’s a talking dog fer cryin out
>> loud, we don’t expect it to be Mr. Peabody (younger crowd, ask ChatGPT who
>> is Mr. Peabody.)
>>
>>
>>
>> ChatGPT has its uses, and has demonstrated itself to be a marvelous
>> teacher and trainer.  I have used it recently to come up to speed on legal
>> terms, and I am convinced it will have enormous impact on society in many
>> ways.  I think it produces insight-like comments, but it is clear enough to
>> me it found them in a huge database rather than invented them.  Perhaps
>> that counts as a kind of legitimate pseudo-insight, and has its uses.  I
>> will accept that it is better than humans at many things we do and pay for,
>> resulting in some professions going away.  The one that comes to mind first
>> is paralegals.  Those guys are adios amigos methinks.
>>
>>
>>
>> ChatGPT makes the same mistakes as humans and it is a marvelous novelty
>> like a talking dog.  I haven’t been able to convince myself it is going to
>> result in the big S Singularity by rewriting itself and becoming a
>> Bostrom-style superintelligence.  That is still ahead of us.
>>
>>
>>
>> But hey, here is an optimistic parting shot: let us use ChatGPT as a
>> trainer, ask it to teach us how to set up large language models.  Then we
>> can all try our own hands at it, ja?
>>
>>
>>
>> spike
>> _______________________________________________
>> extropy-chat mailing list
>> extropy-chat at lists.extropy.org
>> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>>
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230402/be4ffa5b/attachment.htm>