#psa No board meeting tomorrow. I'll be up in North Alabama for a family reason. Back next week!
Remote
Dave
@dave@podcastindex.social
mastodon 4.7.3Running systems at podcastindex.org
2243 Followers
437 Following
50 Posts
Joined September 14, 2020
lud16:
dave@getalby.com
nostr:
npub1pvdw7mm7k20t9dn9gful8n5xua5yv8rmgd9wul88qq5dxj80lpxqd39r3u
I'm going to start closing channels on the Podcast Index lightning node starting today. The $240/month is no longer self sustaining and we lose about $200 on the node every month. It won't stop anyone from boosting in a podcast. We just always kept the channels free as a service so now if you open your own channel to somewhere you might have to pay for it.
Our node's database is so humongous that all the normal tools (bos, rtl, etc) just error out. I'll be doing it from command line (lnd).
A "weight" is just a form of heuristic. Regression is literally built in to the LLM at a fundamental level. Re-weighting a model will always regress other weights. One slider gets pushed up and another goes down. It's impossible to get 100% unit test coverage in such a system. It's a beta test that never ends.
Replying to
@js@podcastindex.social @cisene@podcastindex.social @james@bne.social .xyz domain. Forever a sign of quality.
#podping The gossip-listener is now called podping-gossipwatcher.
You guys would be proud of me. I used docker across the board for all this stuff. :-)
#podping The second podping.cloud host is in the gossip swarm now. Three more to go.
Filed a bug in Safari/WebKit:
https://bugs.webkit.org/show_bug.cgi?id=320657
This has been a multi-day rabbit hole that started with a poorly behaving CORS on Live365 and culminated with filing this Webkit bug report.
If a server/cdn streams an unterminated body in response to an OPTIONS fetch (CORS preflight), Safari hangs indefinitely, waiting for the body to end. That isn't standards compliant.
Chrome and Firefox properly look at the response headers only, close the request and move on.
The entirely foreseeable rug pull begins:
Next up on the roadmap will be to neuter the $100/mth Max plan down to what the $20/mth Pro plan used to be. This won't be a reduction in hard token limits. It will be the normal "our new model is so awesome it just takes more tokens to use it" which is what they have said every time they release a new model.
podping gossip is holding up well after 96 hours of uptime across the swarm. This is encouraging.
We’re in the mainframe/timeshare era of LLM’s right now. The PC era is coming. There were demand induced shortages back then too:
https://www.nytimes.com/1988/03/12/us/shortage-of-memory-chips-has-industry-scrambling.html
Boosted by @tchambers@indieweb.social
I think @theDanielJLewis@podcastindex.social proposal makes a ton of sense:
https://github.com/Podcastindex-org/podcast-namespace/discussions/770
Currently finishing up the dead feeds export. I don't think it can be a public link because it contains feeds killed through DMCA requests so it will have to be a by request, research purposes only type thing.
That, combined with the normal db export, will form the basis of the new model training data set which is what comes next.
Nice!
Time for the board meeting! We're running with scissors live with episode 257. Too many fd's!
🍻
Replying to
@chrislw@podcastindex.social My recent experience with Qwen3.6 35B running locally on a DGX Spark tells me that the local models are virtually on par with the hyperscale models for many tasks like coding. A developer with a 128 gig Mac mini/studio/MacBook can be pretty much self sufficient and not pay those guys a dime.
I think that’s what we are seeing the beginning of now. You can’t find a big RAM Mac anywhere right now. All sold out.
Replying to
"One thing coding agents are great for is generating tests..."
Replying to
@mike@libertynode.net Peertube showed the way forward on this. Full podcasting 2.0 tag support in their feeds.
Replying to
Replying to
@admin@mastodon.foxfam.club @csb@podcastindex.social @adam@podcastindex.social The 35B moe model with reasoning enabled works really good in Opencode. It’s the first time I’ve been impressed with local coding. The 27B dense model is too slow and I don’t see the tradeoff as worth it. Doing side by side identical prompting with it and CC-Opus4.7, they are almost neck and neck for coding tasks.
Replying to
Replying to
@csb@podcastindex.social @adam@podcastindex.social Yes!
That is EBITDA profit “projection” based on a two month annualized run rate where Elon gave them a sweetheart deal on the colossus compute data center. They even said in the investor deck the story was based on, that they would go negative again immediately. They are, and will remain, deeply unprofitable.
I have a clip related to this tomorrow for the board meeting.
I’ll get with you soon on guesting. I haven’t forgotten.
Replying to
I think we're ready to start moving this code into a new repo so it can be built properly for podping.cloud integration.
I have to get @js@podcastindex.social AT proto work done though.
Replying to
@steven@podcastindex.social Strange. I don't use S3. We use Linode Object Storage. I'll check it.
Replying to
And some Vietnamese gambling:
https://podcasts.apple.com/us/podcast/dh88art-promote/id1834074347?i=1000740935761
Replying to
[sigh] Garbage:
Replying to
@russell@podcastindex.social Playwright is amazing. Only discovered it recently and wish I had known about it so much earlier.
Replying to
@paul@tapbots.social Trust us, we will send you a check for that higher electric bill you're gonna start getting. And, some ear plugs for that high pitch whine you hear day and night. And, an Uber to get you to the therapist because of the madness you experience from said perpetual whine.
One thing you can always take to the bank is Silicon Valley promises. They have an unassailable track record.
Replying to
@HeyCitizen @theDanielJLewis That's as close to a definition as possible probably. The fact is some things are just not definable. Something can be true and also not have a fully coherent definition.
I’m going to modify the csv upload endpoint to accept larger files. I’m thinking some are going to be a lot larger than 50mb. #api
Replying to
Replying to
@russell@podcastindex.social Oh, that's horrible. There are going to be so many problems with low quality vibe coded stacks holding PII. It's going to be bad brother. And, customers can never know in advance what the back-end of these systems look like to make an informed decision. 😞
Replying to
Replying to