Time for tinfoil hats?

I’m having a dabble with AI in the forum. Don’t panic though, I’m not going to use it to distill your soul and sell the extract to anyone! I really want to get an idea of:

  1. How it can be used within the context of this forum
  2. How much it could cost to use

At the moment the features that are enabled are only accessible by forum admins so you can’t run up a big bill for me! There are a number of features to explore and I’m already sure that a few of them aren’t appropriate so whilst I might test them I can’t see why we’d need them. However, some of the other features look like they might have some benefit.

The AI features can be enabled for specific user groups within the forum and I’m pretty sure that to avoid spiralling costs they will never be made available to all users. I’ll add further posts giving examples of the features so you can see what’s possible.

The AI features are provided by a Discourse plugin and this is then configured to use a LLM (Large Language Model). I’ve chosen to use the Google Gemini Pro 1.5 LLM for now, but I may experiment with other models if AI seems useful.

I’ll add more posts describing experiments with the different features.

2 Likes

The first feature I tried is Topic Summarization. I tried it with this WxSim related topic: New release coming soon - chance to beta test - #19 by meteosanjuan

I got a new banner at the bottom of the topic, e.g.

This gives a basic summary for the topic, e.g. how many views it’s had, how many links are included and the avatars of those who participated. This doesn’t need AI to generate it. There’s also the ‘Summarize’ button. When I clicked that the topic content was sent to the AI engine which generated the following summary after about 10 seconds:

New WXSIM Build (Version 2024, Build 1.1)

Tom_Ehrensperger announced a new WXSIM build with bug fixes and invited beta testing. weatherbee confirmed the new version worked fine.

Marian_Peteanu encountered issues with white backgrounds in windows and error messages related to “wdatamixed.txt”. Tom_Ehrensperger assisted by explaining the message and requesting log files.

After initial troubleshooting, Marian_Peteanu successfully resolved the issue by correctly importing “wdatamixed.txt”.

The discussion concluded with users, including broadstairs and meteosanjuan, confirming the update worked, and sharing their WXSIM forecast pages. Marian_Peteanu shared a thoughtful reflection on the Olympic spirit and its connection to life.

The summary doesn’t really provide massively useful facts, but it could give you an idea of whether a topic was relevant to what you were looking for.

I’m not sure how much such a summary would cost, but I can see issues with this type of summarization. AI costs are typically based on the size of the input and the size of the output.
So, if you did this on a topic with hundreds (or thousands) of posts the input size would be huge and likely expensive. This gives you an idea of the kind of things I need to investigate before I could make something like this widely available!

It would be interesting to know if you now see the banner and summary for this topic now that I’ve generated it.

1 Like

Another summarization example…for this topic…

The banner is a little different for this. There are more users who have posted, so to avoid filling the screen with avatars the list is truncated with the number of users also noted. The banner also gives a predicted time to read (6 min).

The topic summary is as follows…

Topic Summary

This forum thread discusses different webcams and methods for displaying webcam imagery on personal weather websites using PWS Dashboard.

jaltman initiates the conversation seeking advice for replacing their malfunctioning Ambient camera. Several users, including swright1957 and BKS97, share their webcam setups. blainec recommends Foscam cameras, specifically the FI9902P model, and outlines their method of capturing images directly from the camera’s web server using Weather Display (WD).

jaltman expresses confusion regarding static filenames for image retrieval. hcorrin clarifies that WD fetches images via the camera’s web server URL. They elaborate further in #18, providing a sample URL structure and suggesting the use of iSpyConnect to determine the camera’s JPEG options.

bitsostring shares their experience using a budget-friendly ieGeek IP cam with iSpy software for capturing and uploading images, while novawx details their setup using a Reolink camera and IPTimelapse software.

robloglisci provides a comprehensive overview of their setup, utilizing a Panasonic BL-C160A Pet Camera with CameraFTP and IPTimelapse.

Following the advice received, jaltman reports success with a Foscam G4 camera, confirming its compatibility with Ambient and expressing satisfaction with its performance and ease of integration with WD and PWS Dashboard.

SarahRTaylor presents an alternative solution using a Raspberry Pi equipped with a camera module for capturing and uploading images to their PWS Dashboard website.

The discussion concludes with jaltman and blainec providing further guidance on configuring WD’s webcam capture functionality and utilizing CGI commands for interacting with Foscam cameras.

Summarized with AI on 4 Aug

This summary seems pretty good and gives you some conclusions that would help you decided if you wanted to read the entire thread.

1 Like

I’ve investigated some other AI features this morning and found 3 that I won’t be testing.

Toxicity

This checks post toxicity. Each message that’s posted is checked for hateful, offensive or disrespectful content. There are a few reasons for not looking at this:

  1. Toxicity isn’t really a problem on the forum. There have been a few instances over the years, but it happens so rarely that it’s not worth investing time and money to do anything about it.
  2. As each post has to be scanned it would prove expensive in the long term.
  3. It requires an additional container to be set up to function. Given (1) and (2) it’s not worth the extra work even just test this once.

Sentiment/Emotion

This can check posts for how positive/neutral/negative they are sentiment) and/or whether they express anger/disgust/fear/joy/neutral/sadness/surprise (emotion). This is an interesting concept, but I think it would be more useful for a forum where sentiment/emotion runs higher. We’re very much more a technical community where these concepts are less relevant. In addition (2) and (3) from ‘Toxicity’ also apply and I other than general interest I think this wouldn’t be time/money well spent.

NSFW Image Detection

NSFW = Not Suitable For Work, e.g. sexual/gross content that wouldn’t normally view at work

This feature scans uploaded images to check for NSFW content. I won’t be testing this because…

  1. I don’t recall there ever being a sexual/gross image uploaded to the forum in the 20+ years it’s been running, or perhaps I’ve just blanked the memory out. In any case it’s a vanishingly rare ocurrence.
  2. Each uploaded image has to be scanned and image AI processing is considerably more expensive than text processing, so it would likely prove very expensive in the long term.
  3. This also requires an additional container to be set up to function.
1 Like

It’s been a busy week so I’ve had little time for more experimentation.

One change I’ve made though is to change the LLM model that I’m using. I originally used the Google Gemini Pro model which worked fine. After a bit more investigation I’ve found that the Gemini Flash model works just as well, but is considerably cheaper. If I’ve understood the pricing model Flash seems to be about 20 times cheaper than Pro and it seems to do the same for what I’m trying to do.