spacecowboy
spacecowboy17.bsky.social
spacecowboy
@spacecowboy17.bsky.social
Interests in ML and social aspects of tech.

Building For You feed: https://bsky.app/profile/spacecowboy17.bsky.social/feed/for-you

Hobby project: linklonk.com
November 22, 2025 at 2:33 PM
This would produce smaller rewards for posts with few likes => more stable recommendations.

We also have another parameter (0.3) that prioritizes recommendations with more paths leading to it: score = score * power(num_paths, 0.3)

We can experiment with making it stronger.
November 22, 2025 at 2:18 PM
What probably happens for you is the recent like is on a post with few likes - so the credit for this like gets distributed to very few users and you get too much weight given to their other likes.

We can probably improve this by dividing the credit with some smoothing like 1/(num_likes + 10)
November 22, 2025 at 12:50 PM
Public media is very different - you give free access to content. It is competitive vs private media. Network effects are not that strong.
November 21, 2025 at 9:53 PM
Maybe you force other networks to interoperate with new laws? In that case you could overcome the network effects of existing social networks.

But this is still a long shot. A protocol based network would struggle to keep up with new formats (text->images->videos) and features (group chats?)
November 21, 2025 at 9:50 PM
You may bring up a social network, but how do you get the user to choose to use it? Mandate it? Outlaw the other social networks?

It sounds like a pitch for getting government funding but is that a wise use of money?
November 21, 2025 at 9:44 PM
November 21, 2025 at 1:16 PM
Try For You - it shows content based on your likes and it is usually better than Discover at keeping unwanted content out

bsky.app/profile/did:...
November 21, 2025 at 1:15 PM
November 21, 2025 at 1:09 PM
Try "For You": bsky.app/profile/did:...

It shows content based on your likes
November 21, 2025 at 1:08 PM
Try For You - a feed based on your likes
November 21, 2025 at 1:05 PM
November 21, 2025 at 12:50 PM
Try For You - collaborative filtering based on personalized PageRank on the graph of users and likes.

It builds a network of trust to prioritize content from the high signal-to-noise curators.
November 21, 2025 at 2:02 AM
In the recent Karpathy interview there was this idea that the human brain uses dreams to incorporate short term memory into the long term memory.
November 21, 2025 at 1:53 AM
Could be cool to show a link to users where they can find more details about each feed item. The feed generator could return such a link in the response. Probably other feed builders could be interested in that. @devingaffney.com
November 20, 2025 at 10:00 PM
Concatenation could work, though not sure how a long string would be displayed.
November 20, 2025 at 10:00 PM
There is a bunch of debug information I could return instead: score, num curators, num paths. See linklonk.com/bluesky

But how else would I identify the source feed?
LinkLonk
linklonk.com
November 20, 2025 at 9:36 PM
This is the "feed context" string that the feed generator returns in response. This string gets sent along with the interactions like "show less".

I use the feed name to tell which feed the interaction is from because I use the same generator URL to serve for-you, videos-for-you and links-for-you
November 20, 2025 at 9:31 PM
Try For You: bsky.app/profile/did:...

It does pay attention to "show less like this"
November 20, 2025 at 2:30 PM
"Amazon remains Anthropic’s primary cloud provider and training partner."
November 18, 2025 at 8:46 PM
Is that a public or private eval set?
November 18, 2025 at 6:46 PM
Not using AWS and CloudFlare is sometimes a competitive advantage!
November 18, 2025 at 2:55 PM
Try this feed: bsky.app/profile/did:...

It finds people who liked the same posts as you and shows you what else they've liked recently.
November 18, 2025 at 1:39 AM
Thanks! Captured the token from the logs and it validates fine. I think the issue is that I'm caching the did-documents with no expiration and it uses the signing keys from the old PDS.

I should probably invalidate the cached did-document once I fail to validate the token.
November 16, 2025 at 1:20 AM
Mine is "EcoFlow DELTA 3 Portable Power Station" - it has 1024Wh of Li-ion batteries. This should give ~8 hours of power - PC+router+modem consume ~110W.

Haven't had a power outage since installing - so don't know how well it work. I could turn off the electrical panel breaker to test it.
November 16, 2025 at 12:57 AM