Re: ActivityPub and AI bots?

Bob, Melvin, Daiki,

On Sun, 2 Feb 2025 at 17:56, Melvin Carvalho <melvincarvalho@gmail.com>
wrote:

> ne 2. 2. 2025 v 10:58 odesílatel Daiki Mizukami <tesaguriguma@gmail.com>
> napsal:
>
>> You may be interested in `@FediChatBot@fedichatbot.deno.dev`
>> <https://fedichatbot.deno.dev/ap/actor/bot>. Excerpt from its profile:
>>
>> > An LLM-powered chatbot for the fediverse, powered by BotKit and Gemini
>> 2.0 Flash (experimental)
>>
>
> Thanks for sharing.  That is very cool!  Gemini being free right now makes
> it a great time for experimentation!
>

Last time I checked on my very expensive Pixel 7a I was led to believe was
AI hardware, it could not even read back the second paragraph it spoke and
typed, but now seems to have that ability on just checking it does seem to
have some level of contextual memory.

> Originally, when I imagined "Semantic Inbox" (which evolved into Solid
> Inbox → LDP Inbox → ActivityPub Inbox), the idea was to accept all sorts of
> HTTP POST objects—essentially an email inbox, but better, because HTTP
> makes it more expressive.  And Linked Data can have many "Shapes", Activity
> Streams being one.
>
I have been looking at a generalization of this.

> It wouldn’t be too hard for Solid, ActivityPub, and Nostr servers to
> support MCP (Model Context Protocol) over HTTP POST (in this case,
> JSON-RPC). Both ActivityPub and Solid already have inboxes for POST.  That
> wouldn’t just make ActivityPub more agentic but could also provide a
> unifying interface for the broader social web.
>
> https://spec.modelcontextprotocol.io/specification/2024-11-05/
>
Nice spec, very useful, OpenAPI is the way to go on this ! Having the
TypeScript request/response interface generated is very nice.

> Not for everyone, of course, but the potential use cases are incredibly
> powerful. I think we should do it!
>
AI can waste a lot of time, especially for trying to use it for semi
complex programming problems or algorithms. I have run it through a range
of programming tasks and it does get better at some of them. It is a bit
overly verbose and without direct feedback it has a tendency to produce
garbage. I don't think it is anywhere near ready for any form of
architectural tasks due to it context buffer but if that context buffer is
semantic rather than token based I may well be fairing a lot better. Again
for architectural level programming it really requires some level of
experience or experiments over time to come to solve real world problems in
a manner where they will test the tests of time in a real world
environment, rather than just applying oversimplified patterns as it
tends to do.

For non programming tasks such as sorting through toot's for sense, or
SPAM, or whether it is another machine's output, that is quite an
interesting open question.

>
>>
>> On 2025/02/02 2:56, Bob Wyman wrote:
>> > Has anyone yet used the API's of Llama, or something else, to link
>> > ActivityPub to AI? Is there a bot that can receive and answer user
>> > queries via ActivityPub?
>>
>
If this is going to happen as it will I think its worth doing properly and
efficiently, and we should attempt to get ahead of the curve so we know
what we are facing.

Yes this does deserve our attention !

Aaron

>
>> > bob wyman
>> >
>>
>> --
>> Daiki "tesaguri" Mizukami
>> <mailto:tesaguriguma@gmail.com>
>> OpenPGP: <openpgp4fpr:f706f9bbb8f63cde5f40298fdf5eb02be5c2542e>
>> Fediverse: @tesaguri@fedibird.com <https://fedibird.com/@tesaguri>
>> Matrix: @tesaguri:matrix.org <https://matrix.to/#/@tesaguri:matrix.org>
>> GitHub: <https://github.com/tesaguri>
>> Keybase: <https://keybase.io/tesaguri>
>>
>>
>>

-- 
Aaron Gray - @AaronNGray@fosstodon.org | @aaronngray@threads.net |
@AaronNGray@Twitter.com

Independent Open Source Software Engineer, Computer Language Researcher and
Designer, Amateur Type Theorist, Amateur Computer Scientist,
Environmentalist and Climate Science Researcher and Disseminator.