[Opensim-users] Using GPT3 and openAI to create NPC bots

[Dr Ramesh Ramloll]

I have applied for GPT3 three times already but yes it is hard to get
access. Though I know a few people who got access recently and they did not
even have to submit project proposals, so I am still hopeful. I agree with
Teravus assessment. In fact, I am researching GPT2 at the moment. So in
short, at the moment, it's accessing these services that is the challenge,
getting them to work with NPCs in world is very doable. I do think it will
be a game changer when it happens. I hope all of you will stay in this
game, good things ahead. I know this virtual world space can feel hard
sometimes. It's a bit like the stock market, we need to dig in to try to
remain solvent... 'The market can remain irrational longer than you can
remain solvent.' so our job is to stay in the game as long as we can ;)
Don't know why am saying this, I am comforted to see all these names
above.. love you all.

On Mon, Feb 22, 2021 at 3:04 AM Teravus Ovares <teravus at gmail.com> wrote:

> Just to follow up on my own email;
>
> Have a look here for how someone else got access to GPT-3
> https://outline.com/RXMbpS
>
> The beta comes with 300k 'free' api tokens for building your app.   After
> that, it costs between $0.008 and $0.06 per 1k tokens.
> A token is approximately 4 characters of text, so it's pretty easy to use
> up 1k tokens in a few requests.
>
> OpenAI will have to review your app before it can be used publicly..   and
> they're very careful about allowing it to be used ethically and in a way
> that doesn't impersonate people in a whole (which it *can* do from very
> little input text).
>
> Also, remember, GPT-3 isn't, necessarily, conversational by default.  It is
> a prediction service.   Given X was said in the input tokens, the most
> likely thing to 'complete' a sentence or paragraph is Y. It takes some text
> structure munging and careful prompting to get it to 'complete' what you
> typed in a way that seems conversational.   Unlike with GPT-2, you can't
> fine-tune GPT-3 with conversational text because...   the model isn't
> available for you to download.   Even if you did have the complete trained
> model..   you wouldn't have the computing resources to run a query on it,
> let alone fine-tune it with conversational text.
>
> In GPT-2 you can prompt it like this and it works (sometimes)
>
> Want:  You want it to tell you where the virtual world navigation beacon
> is.
> Prompt: A computer gamer sat down in the chair and put on a virtual reality
> headset.  The headset lit up and the gamer peered into the virtual world
> looking for the hub navigation beacon.   The place the gamer found it was:
>
>
>    - Set the stage (tell the story and set the context)
>    - Write the prompt as if the character in the story already knows or is
>    just about to discover the thing.
>    - It often helps to prompt a completion with :, because, in actual
>    writing, a colon and semicolon is just before examples of whatever
> precedes
>    the colon.
>
>
>
> I'm not excited about the way that OpenAI has set up GPT-3 and access
> through an API.  As far as I'm concerned, the 'OpenAI' name is..
>  misplaced at best with GPT-3 ..   but..     I understand the need to make
> money so the project can continue to create new and innovative neural
> networks.
>
> Regards
>
> Teravus
>
>
>
>
>
>
>
>
> On Sun, Feb 21, 2021 at 10:17 PM Teravus Ovares <teravus at gmail.com> wrote:
>
> > I would also say..  the biggest barrier is getting access to GPT-3.
> > When you do get access to it..    you have to get right in and build your
> > app.  They don't give you much time.     Also..   after a short testing
> > period, OpenAI charges you to use GPT-3 by the token.   Cost wise..
> > probably not effective for most people.
> >
> > On Sun, Feb 21, 2021 at 6:44 PM Gwyneth Llewelyn <
> > gwyneth.llewelyn at gwynethllewelyn.net> wrote:
> >
> >> Hi! Now talk about serendipity... for two upcoming projects, I need some
> >> chatbots, something always described by clients/end-users as 'we just
> need
> >> relatively simple chatbots' when in reality they expect natural language
> >> understanding...
> >>
> >> Anyway, I'm pretty sure I've been using llHttpRequest() with HTTPS
> >> requests
> >> recently, and I believe that they work quite nicely. As Nick suggested,
> >> the
> >> quickest way to implement something is to have an object attached to the
> >> bot/NPC and let it talk 'through' it, for instance, using osNpcSay().
> >>
> >> To do very complex things I'd suggest having a 'proxy' running outside
> >> OpenSim which talks back to whatever AI/NLU/NLP platform you're using
> and
> >> connects to the in-world script attached to the bot.
> >>
> >> And if you're planning to use OpenAI's GPT-3 — meaning that you *have*
> >> managed to get an invite to the beta!! — I envy you profusely.
> >>
> >> Cheers :-)
> >>
> >>    - Gwyn
> >>
> >> --
> >> "I'm not building a game. I'm building a new country."
> >> -- Philip "Linden" Rosedale, interview to Wired, 2004-05-08
> >>
> >> [image: Image] <http://feeds.feedburner.com/~r/GwynethLlewelyn/~6/3>
> >> _______________________________________________
> >> Opensim-users mailing list
> >> Opensim-users at opensimulator.org
> >> http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-users
> >>
> >
> _______________________________________________
> Opensim-users mailing list
> Opensim-users at opensimulator.org
> http://opensimulator.org/cgi-bin/mailman/listinfo/opensim-users
>


-- 
'Consider how the lilies grow. They do not labor or spin.'
*Rameshsharma Ramloll* PhD, CEO CTO DeepSemaphore LLC, Landisville, PA;
Affiliate *Research Associate Professor*, Idaho State University,
Pocatello, ID 83209 Tel: 208-240-0040
LinkedIn <http://www.linkedin.com/in/rameshramloll>