It would be interesting if we could use the archived emails to train an LLM to create a knowledge base. I’ve done that for my own notes using NotebookLM, but I don’t know how to do that publicly and how we would authorize access to it (if that is important - we could also scrub names from the data first).
First, we have to get past the migration from the old system to the new.
But it’s an interesting thought… When we were initially trying out Discourse we did create some AI summaries and load them into a Wiki category.
I’m not sure with current tools how it would work on a collection of thousands of topics that wander everywhere… Topic by topic we can get quick summaries, but feeding in the lot of them…??
As a peek, I took the recent topic Propeller Performance for Cruising Yachts and created an AI summary using both Discourse AI and NotebookLM:
NotebookLM:
Cruising Yacht Propeller Performance Discussion
1 source
These excerpts from an article titled “Propeller Performance for Cruising Yachts” present several short discussions among sailors regarding their experiences with various types of propellers and engines. Nick Hirst praises the Kiwi feathering propeller, noting its improvement to sailing despite requiring occasional weed clearing, and expresses that putting his old fixed three-blade prop in neutral created drag. Charles Garbarino discusses upgrading from a two-bladed MaxProp to a three-bladed Max Prop feathering propeller after repowering with a Beta 25 engine, reporting improved reverse performance and less prop walk, though he cannot compare drag to a fixed propeller. Finally, Stephen Marinelli mentions installing a fixed pitch three-blade propeller with a new Beta 35 engine, stating that the change had no significant negative impact on the sailing speed of his specific type of boat.
Discourse AI:
Topic summary
In a discussion on folding versus three-blade props for Nonsuch sailboats, users share varied experiences. Sgrmagnolia10 seeks advice for a new engine, noting Tessa’s folding prop reduces drag but wonders how much. Breezin advises folding props for racing or long-distance cruising (adding 0.5–1.0 knots), while fixed props suffice for casual day sailing. Rob highlights PHRF handicapping discrepancies — his fixed 3-blade adds 12 seconds per mile, but a 0.5-knot gain from a folder could yield tactical racing advantages despite quirks in reverse. bwc4884 endorses Flexofold for reduced propwalk and references a detailed Yachting Monthly test, noting folders are cheaper than featherers. ddewolfe praises his Campbell Sailer fixed 3-blade for low drag and durability. Mirthin switched to a Kiwi feathering composite prop, gaining 1 knot in light winds, with less propwalk and good reverse — and notes it’s cheaper than Flexofold. Cmack finds his Volvo Penta saildrive’s 3-blade folding prop highly responsive in reverse. RMP recounts a race where a feathering prop outperformed a fixed one; he notes featherers reverse better but cost more and need maintenance, and warns Kiwi props may be over-pitched in reverse unless adjusted. Mirthin disagrees that neutral helps fixed props, and defends his Kiwi prop despite needing weed clearing. Charles-Garbarino upgraded to a 3-blade MaxProp, noting improved reverse handling and quiet motoring, and recommends the Yachting Monthly article. Sunshine271 reports no sailing penalty with a fixed 3-blade on his catboat, accepting minor speed loss as negligible.
The Discourse summaries are nice, that’s a great feature.
I was thinking about it more from a Q&A perspective though. So instead of looking by topic on a wiki, the interface might be just a prompt. For example, I loaded that thread (or maybe a similar one) into Notebook LM and asked “are maxprops good on a nonsuch”. I’d say the reply is good, but of course not as detailed as the whole thread. Here’s what it spit out:
Based on the experiences shared by members of the INA Nonsuch Discussion Group, the MaxProp feathering propeller is generally well-known on Nonsuch vessels, with some owners reporting good performance, while others have encountered issues leading them to replace or rebuild them.
Here is a summary of experiences regarding MaxProps on Nonsuch sailboats:
Positive Reports and Longevity
- One owner of an N30U reports having a first generation, 2-blade, feathering MaxProp that “works great” [1, 2]. This owner noted that prior to greasing, the prop transitions smoothly from forward to reverse and vice versa [1, 2].
- Another owner of a Nonsuch 26C (“La Reina”) had a 2-blade MaxProp Classic, which was purchased in 1982 and was among the first ones sold by PYI [3-5]. It was replaced in 2019, having accumulated over 2000 hours of motoring [3-5].
Maintenance and Performance Issues
- MaxProps require maintenance, and signs of wear often lead to discussions of rebuild or replacement.
- One owner noticed slight “play” (front to back) in the blade joint of their 2-blade MaxProp prior to greasing, though the play did not look excessive to one respondent until it was filled with grease [1, 2, 5, 6].
- MaxProps can sometimes be rebuilt. A rebuild quote was mentioned to be in the neighborhood of $1800 [7-9].
- Vibration and Configuration: One owner of an NS30U (#460) bought their boat with a 2-blade Max Prop that caused a “significant vibration” [7, 9]. After pulling the prop, the owner opted for a new three-blade prop instead of rebuilding the MaxProp [7, 9]. The new prop performs “much better” and eliminated the vibration [7, 9].
- The cost of a new 2-blade, 15-inch prop was quoted at $2450, while a new 3-blade prop was $3100 [6, 8].
Alternatives Mentioned
While the query focuses on the MaxProp, other owners recommended alternative feathering/folding props:
- One owner suggested that if buying new, one might “consider a Gori for a whole lot less money,” noting they had been happy with their Gori for the past 20 years [10].
- Another Nonsuch owner (Mirthin N30 Ultra) replaced their fixed three-blade prop with a feathering prop from Kiwi prop and reported gaining 1 to 1.5 knots while sailing, and found it good in both forward and reverse [11].
FWIW, the INA did an experiment last year with an AI company that was given access to our entire collection of documents and the archives of the Google Discussion Group.
It failed testing, due to the inability to eliminate “hallucinations” – made up answers presented with confidence. Consequently, the INA Board voted not to release it, and dropped the experiment. The final straw for me was when it recommended that you handle being overpowered in high winds by reducing the height of your mast.
I used to be in AI R&D before I retired. I’m currently helping a former student writing a book on related topics. My impression from his work is that LLMs work best in a paradigm called Retrieval Augmented Generation, in which other mechanisms are used to select the documents used to answer a question, and the LLM is used only for generating readable text. (I’m glossing over a lot of detail because this is a sailing Forum rather than an AI discussion.)
Summarizing a thread seems to be something that can be done reasonably well.
Ditto for summarizing a set of posts within a thread which have been tagged by people as containing good answers.
But, at least as of the beginning of 2025, using LLMs to take in questions and produce answers was clearly a bridge too far.
As near as I can tell, that remains the case. It’s definitely something to watch.
In the meantime, the most useful thing we can do with Discourse’s AI aid might be to feed it posts that people tagged as helpful, and asking it to summarize those.
– Bob
Thanks for the detailed explanation, good to know that path was already researched. And generally that has been my experience with LLM queries as well, mostly good with some brazenly inaccurate hallucinations thrown in. Shorten the mast, ha!