On the little-lamented Google+, I'd long noted that larger groups often had abysmally poor discussion quality. Most were overrun with self-promoters, scammers, and to an extent, advertising (not officially supported, but certainly present). You can see a similar trend on group-discussion sites such as Reddit.
I don't know what the hard-and-fast rule for "best discussion quality" is, and of course that metric will vary by topic and community. I'm pretty sure that this is strongly at odds with the interests of advertisers, for which the goal seems to be to aggregate the largest collection of eyeballs possible at a single venue or channel.
On Reddit, I've noted that discussion quality seems to tank strongly at some point between ~10k and 1m subscribers. Note that subscribers != active participants, but it's a rough proxy. Looking right now at a subreddit I'd found at least somewhat useful at 100k members, it's now boasting north of 500k and ... while not a cesspit as occurs elsewhere, the overall quality is markedly lower.
I had the opportunity to do some detailed analysis of overall statistics of Google+ groups (over 8 million in total at the time of the site shutdown), and activity vs. other parameters. In particular, it seems that measureable interactions (+1s, shares) scale most directly with post activity rather than total community members, which makes some sense. G+ was around long enough that "subscribers" was only very vaguely related to actual participants (as is undoubtedly the case on Reddit as well). In particular, I noted that the most active profiles also seemed highly prone to going inactive or being deleted or removed (I'm not sure which). I suspect that that was a mix of spam accounts (whether created as such initially or converted / hijacked later), and of simple burnout or disaffection. A recent post about interactions on a blog offered a similar observation, and my read is that each individual interaction is an opportunity to push someone over the edge, and that audience cultivation is a balance of bringing aboard new members whilst minimising loss of old. Blocking dynamics on other platforms (I've witnessed this on G+, Diaspora*, Mastodon/Fedivers, Ello, Slashdot, and Usenet, amongst others) is somewhat similar, and seems to me to great interesting network topologies as connections are made and broken.
I've also experimented with small groups of active members and found that ~50 people really does seem to be a sweet point, though even there perhaps 15 or fewer were highly-active participants. (I was the administrator of that group, and knew all the participants, if largely online-only.)
I've also seen small groups persist over years or decades, and watched those stale and spin off members with time. Unfortunately, without fresh blood, that seems all but inevitable, and recruiting new members is a challenge. See groups such as The Well (still extant) as long-lived examples. Many members there are now entering their dotage, if not having passed through the other side of it...
Dave Winer's observation still stands: conversation doesn't scale very well. I think that's innate, and part of the magic when it does emerge.