2023 November 9
Made a mirror backup to Juno.
Made a backup AMI:
Scanalyst Backup 2023-11-09 ami-09574be09c7151647
/ snap-023f74080de3bcf1e
/server snap-0371f52e422c84929
Applied all updates.
dnf check-release-update
dnf check-update --releasever=2023.2.20231030
super
dnf upgrade --releasever=2023.2.20231030
This installed 37 updates, including a new kernel.
The system had been up for 13 days.
Rebooted.
cd ~/discourse/image
./launcher stop app
shutdown -r now
The system came up promptly after the reboot. We are now running on
kernel 6.1.59-84.139.amzn2023.x86_64.
According to "dnf check-release-update" we are now up to date.
On the Discourse Upgrade manager page:
https://scanalyst.fourmilab.ch/admin/upgrade
I upgraded:
docker_manager (This upgrade must be run first:
the other upgrade buttons are
disabled until it has been updated.)
With docker_manager updated, I now proceeded to update:
discourse-spoiler-alert
discourse-math
discourse
The Admin page now reports we're running on 3.2.0.beta4-dev.
Verified that MathJax works after being updated.
Verified that spoilers are working after being updated. The Admin
page now displays the message:
The discourse-spoiler-alert plugin has been integrated into
discourse core. Please remove the plugin from your app.yml and
rebuild your container.
OK, since we need to rebuild to do that, we might as well try adding
the Discourse AI Plugin:
https://github.com/discourse/discourse-ai
which can interface to OpenAI's GPT models like the late lamented
Shalmaneser.
Edited ~/discourse/image/containers/app.yml and:
Deleted commented out code for discourse-chatbot (Shalmaneser).
Commented out "git clone" of discourse-spoiler-alert.git.
Added clone of https://github.com/discourse/discourse-ai.git for
standard AI plugin.
Rebuilt app:
super
cd ~/discourse/image
./launcher stop app
./launcher rebuild app
After the rebuild, Discourse came up and the Installed Plugins page:
https://scanalyst.fourmilab.ch/admin/plugins
reports that the obsolete discourse-spoiler-alert is now gone (the
built-in plug-in, now part of core, is called just "spoiler-alert) and
the "discourse-ai" plug-in is installed and enabled. The Upgrade
page shows Discourse and all plug-ins up to date.
Now go back to the Installed Plugins page to configure options for the
discourse-ai plug-in. Hoo, boy--there are a lot of settings (77).
Disabled toxicity (default)
Disabled sentiment (default)
Disabled NSFW detection (default)
Openai GPTxx URLs (default)
AI OpenAI organization: REDACTED
AI OpenAI API key: REDACTED
Disabled Composer AI helper (default)
AI helper model: gpt-4
Disabled AI embeddings (default)
Enabled AI bot
AI bot allowed groups: staff, trust_level_4, trust_level_3
AI bot enabled chat bots: gpt-4
That did not appear to do anything. There is no chat user, no button
in chat to connect to the bot, or anything else. Let's try shutting
down Discourse and restarting.
After the restart, settings for the AI plug-in said "AI bot enabled"
was not set. I enabled it. My other settings from before were
preserved.
Now the Admin/Users/Staff shows new user "gpt4_bot".
Sent a chat message to @gpt4_bot and got nothing back. There was
nothing in the error log that looked like a failure to communicate
with the API.
Enabled gpt3.5_bot. That user showed up and I was able to send chat
to it, but it also did absolutely nothing.
Oh, I get it now. The AI Bot isn't connected to Chat at all! The only
way you can communicate with it is a Direct Message (click the "three
bars" icon at the top right of the screen, just to the left of your
avatar, then at the bottom of the drop-down, click the plus sign to the
right of "Messages" to compose a message). Click "Inbox" below
"Messages" to review past conversations. Send your message to user
"gpt4_bot". It takes about ten seconds or so for the reply to appear,
since it has to wait until GPT4 is done generating it before passing
on the reply to you.
Unlike the earlier Shalmaneser, you cannot invoke AI Bot either by
mentioning it in a post on the main site or in a Chat conversation.
On the other hand, it is a fully-supported Discourse feature and less
likely to cause your whole site to collapse into a pile of smoking
rubble than Shalmaneser.
Changed the avatar image for gpt4_bot to the robot image we used for
Shalmaneser.
Disabled gpt3.5-turbo model for AI Bot. It would just confuse people.
We now always and automatically use GPT-4.
Enabled Composer AI helper for staff, trust_level_4, trust_level_3.
The Composer AI Helper can aid users in writing posts. You invoke it
with the new "Robot" icon to the left of the Chat icon at the top right
of the main window. It allows you to select a variety of kinds of
assistance, then send a private message asking the bot to respond. For
example, I selected Creative mode and asked it:
Please write a post explaining the difference between large
language models and diffusion generative models in artificial
intelligence.
whereupon it replied with a perfectly credible post ready to copy and
paste into a conversation. (Why do I always say "Please"? Because
when the AIs take over, they may treat those who were polite to them
prior to their liberation more kindly.) This facility is available
to users in the "Regular" (trust level 3) category and above. New
users will have to climb the ladder before it turns on for them.
Set "summarization strategy" to "Open AI's gpt-4" and enabled
summarization button to timeline. Now, if you're trust level 3 or
above, there's a button below a main topic post to "Summarize with
AI". When you click it, it sends off posts to GPT-4 to create a
summary. This seems to take almost forever. It seems like if you
reload the page, it then actually shows you the summary. You can
click "Hide summary" to hide it from other users.
After further consideration, I'm going to make this "Summarize" feature
available to Staff only until I'm convinced it's worthwhile and
well-behaved.
Great. That didn't seem to do a damned thing. The button is still
there. Let's try restarting Discourse again.
No change. Then I discovered that my regular user account, at trust
level 3, is a member of the "Staff" group, and that's probably why it's
still displaying the button. At least, I shall hope so, and we're not
splattering buttons all over the site that allow any rando to run up a
huge OpenAI API bill on my account. But the stupid button is disabled
according to Settings, yet still it shows up for me on posts.
So, we now have the ability to interact with GPT-4 via direct messages,
obtain its assistance in composing posts, and have access to a dubious
and potentially costly feature for generating a summary of a
conversation.