Jonathan Stray
@jonathanstray.com
Making @mysky.social, the feed algorithm you control (it’s the most fun too). Working on AI and inter-group conflict research @chai-berkeley.bsky.social.
An entirely reasonable take on the whole "is it infosec or alignment?" debate. Spoiler: both.
I wrote a post on whether sandboxing is enough to contain rogue agents. blog.cryptographyengineering.com/2026/09/30/i...
Is sandboxing sufficient to contain rogue agents?
Quick caveats: this is a post on AI safety, written by a cryptography professor. If that troubles you, you should read something else. I try hard not to work on AI (except when the topic occasional…
blog.cryptographyengineering.com
October 1, 2026 at 3:34 PM
An entirely reasonable take on the whole "is it infosec or alignment?" debate. Spoiler: both.
ngl, turning down the "politics" slider on my mysky.social feed has improved my BlueSky experience immensely. It's also something we repeatedly hear in user testing. Discuss.
MySky Social | Jonathan Stray | Substack
Open-source AI-driven Social Media Feeds for Everyone. Click to read MySky Social, by Jonathan Stray, a Substack publication with hundreds of subscribers.
mysky.social
September 29, 2026 at 10:21 PM
ngl, turning down the "politics" slider on my mysky.social feed has improved my BlueSky experience immensely. It's also something we repeatedly hear in user testing. Discuss.
Reposted by Jonathan Stray
You don't have to let someone else decide what you see.
MySky is the first feed algorithm that you control. Open source, open science, and made just for you.
Try it: bsky.app/profile/did:...
MySky is the first feed algorithm that you control. Open source, open science, and made just for you.
Try it: bsky.app/profile/did:...
September 29, 2026 at 12:12 AM
You don't have to let someone else decide what you see.
MySky is the first feed algorithm that you control. Open source, open science, and made just for you.
Try it: bsky.app/profile/did:...
MySky is the first feed algorithm that you control. Open source, open science, and made just for you.
Try it: bsky.app/profile/did:...
Reposted by Jonathan Stray
I made a little Jev-based language model! It uses beam search + Jev as a classifier to judge branches based on coherence and "success" (whatever that means). I've barely talked to it so I don't know how it'll behave or if it'll be interesting at to talk to at all, let's find out! @jevbot.bsky.social
September 25, 2026 at 2:58 AM
I made a little Jev-based language model! It uses beam search + Jev as a classifier to judge branches based on coherence and "success" (whatever that means). I've barely talked to it so I don't know how it'll behave or if it'll be interesting at to talk to at all, let's find out! @jevbot.bsky.social
I've been using CapCut to edit video for the first time, and it's fighting me. Very AI, lots of presets, but very little of the control I'm used to from After Effects, Final Cut, etc. There's a "motion blur" filter but I can't change the direction. It feels super dumbed down and I hate it.
September 26, 2026 at 7:04 PM
I've been using CapCut to edit video for the first time, and it's fighting me. Very AI, lots of presets, but very little of the control I'm used to from After Effects, Final Cut, etc. There's a "motion blur" filter but I can't change the direction. It feels super dumbed down and I hate it.
This is the first piece of AI art that I thought was good, actually. It's stylish, it's funny, it made me feel things.
"Hey Opus, I want you to make a Zine by Claude, expressing something fundamental about Claudishness or your perspective. Think the original 2600, Principia Discordia, punk zines, etc...."
Not bad. I appreciate it mocking my prompt & itself. Full thing: stateless-zine.netlify.app
Not bad. I appreciate it mocking my prompt & itself. Full thing: stateless-zine.netlify.app
September 25, 2026 at 6:50 PM
This is the first piece of AI art that I thought was good, actually. It's stylish, it's funny, it made me feel things.
One of the things that is surprisingly good about MySky is that when I get a banger or a dud I can go to the "Why Am I Seeing This?" page and find out. It makes me think about my settings, my network, and the algorithm.
September 25, 2026 at 4:09 AM
One of the things that is surprisingly good about MySky is that when I get a banger or a dud I can go to the "Why Am I Seeing This?" page and find out. It makes me think about my settings, my network, and the algorithm.
Reposted by Jonathan Stray
Nice. Cold start is a hard problem. For the moment MySky has settled on retrieving posts for the average embedding vector of all MySky users. As opposed to all BSky users. This finds interesting stuff without assuming you like either lefty politics or furries (no offense to either).
September 25, 2026 at 1:49 AM
Nice. Cold start is a hard problem. For the moment MySky has settled on retrieving posts for the average embedding vector of all MySky users. As opposed to all BSky users. This finds interesting stuff without assuming you like either lefty politics or furries (no offense to either).
Reposted by Jonathan Stray
We're more than just a pretty feed.
MySky is an API for the ATProto community to build production recommender systems. Here's how to query our search indices, candidate generators, and scoring models.
What algorithm do you want to build?
atproto.com/blog/mysky-f...
MySky is an API for the ATProto community to build production recommender systems. Here's how to query our search indices, candidate generators, and scoring models.
What algorithm do you want to build?
atproto.com/blog/mysky-f...
MySky: Open, Controllable Feed Infrastructure for the Atmosphere
MySky is an algorithmic feed running on Bluesky, with a settings panel that shows you the whole ranking pipeline exactly as it runs.
atproto.com
September 24, 2026 at 9:55 PM
We're more than just a pretty feed.
MySky is an API for the ATProto community to build production recommender systems. Here's how to query our search indices, candidate generators, and scoring models.
What algorithm do you want to build?
atproto.com/blog/mysky-f...
MySky is an API for the ATProto community to build production recommender systems. Here's how to query our search indices, candidate generators, and scoring models.
What algorithm do you want to build?
atproto.com/blog/mysky-f...
Reposted by Jonathan Stray
The Facial Expressions of Tethys by artist Julius von Bismarck. A suspended high-sea buoy that moves in real time. Sensors on a twin buoy out in the Atlantic Ocean send live satellite data to motors inside the museum. This makes the indoor buoy copy the exact, real-time movements of the ocean waves.
September 23, 2026 at 8:49 PM
The Facial Expressions of Tethys by artist Julius von Bismarck. A suspended high-sea buoy that moves in real time. Sensors on a twin buoy out in the Atlantic Ocean send live satellite data to motors inside the museum. This makes the indoor buoy copy the exact, real-time movements of the ocean waves.
The building CHAI is in looks kinda like Starfleet Academy and I think that’s beautiful
September 23, 2026 at 11:41 PM
The building CHAI is in looks kinda like Starfleet Academy and I think that’s beautiful
Reposted by Jonathan Stray
Today I learned there are only two known polyhedra in which each face shares an edge with every other face.
One of them being the timeless beauty of the tetrahedron and the other one being whatever the heck this is
One of them being the timeless beauty of the tetrahedron and the other one being whatever the heck this is
September 20, 2026 at 3:19 PM
Today I learned there are only two known polyhedra in which each face shares an edge with every other face.
One of them being the timeless beauty of the tetrahedron and the other one being whatever the heck this is
One of them being the timeless beauty of the tetrahedron and the other one being whatever the heck this is
I have never expressed a p(doom), and I think @randomwalker.bsky.social is basically right that we have little basis for quantitative estimates of AI-induced catastrophic risk.
OTH, no one has presented a knock-down argument that *any* of the various AI disaster scenarios are impossible.
OTH, no one has presented a knock-down argument that *any* of the various AI disaster scenarios are impossible.
September 23, 2026 at 4:36 PM
I have never expressed a p(doom), and I think @randomwalker.bsky.social is basically right that we have little basis for quantitative estimates of AI-induced catastrophic risk.
OTH, no one has presented a knock-down argument that *any* of the various AI disaster scenarios are impossible.
OTH, no one has presented a knock-down argument that *any* of the various AI disaster scenarios are impossible.
What's the general sentiment on the phrase "stochastic parrot" over here in the Atmosphere?
September 23, 2026 at 12:31 AM
What's the general sentiment on the phrase "stochastic parrot" over here in the Atmosphere?
Reposted by Jonathan Stray
"they are just doing a lot of simple operations! 🤢" versus "they are just doing a lot of simple operations! 😯"
September 21, 2026 at 11:39 PM
"they are just doing a lot of simple operations! 🤢" versus "they are just doing a lot of simple operations! 😯"
Reposted by Jonathan Stray
I'll note that the data shows that social media use stayed steady during this period (AI use isn't surveyed, but other studies show it has increased steadily in this age group over the time period).
After hitting its lowest levels in a decade in 2021, teenage mental health is showing significant improvements, according to new data released Monday by the CDC.
Sadness and Suicidal Behavior Among Teenagers Is Declining, C.D.C. Finds
A national survey has found that key mental health measures for teenagers have significantly improved from their worst levels during the pandemic.
nyti.ms
September 22, 2026 at 2:47 AM
I'll note that the data shows that social media use stayed steady during this period (AI use isn't surveyed, but other studies show it has increased steadily in this age group over the time period).
There we go, another major LLM does some unintended hacking. This suggests that the problem is, in fact, hard to solve.
www.reuters.com/business/gem...
www.reuters.com/business/gem...
Gemini hacked three companies in first known breakout by Google's AI
Gemini guessed credentials to access three websites it thought were in the scope of its test.
www.reuters.com
September 19, 2026 at 8:12 PM
There we go, another major LLM does some unintended hacking. This suggests that the problem is, in fact, hard to solve.
www.reuters.com/business/gem...
www.reuters.com/business/gem...
AI-induced conflict is a catastrophic risk. It's distinct from, but related to epistemic risk, sociopolitcal risk, cooperative AI, concentration of power risk, etc. It's the AI safety risk I'm working on.
Forget Killer Robots. The biggest imminent threat of sloppy generative AI integration into military decision-making (we can probably blame Pete Hegseth) is that it mistakenly helps start a war.
edition.cnn.com/2026/09/18/p...
edition.cnn.com/2026/09/18/p...
Exclusive: US military had close call after using AI for false intelligence report, sources say | CNN Politics
The episode shows the risks of using this new, relatively poorly understood technology in the middle of the Iran war
edition.cnn.com
September 19, 2026 at 3:00 AM
AI-induced conflict is a catastrophic risk. It's distinct from, but related to epistemic risk, sociopolitcal risk, cooperative AI, concentration of power risk, etc. It's the AI safety risk I'm working on.
Reposted by Jonathan Stray
Soon you will be able to prompt for what you want to see and MySky will instantly do an LLM-powered search across the entire AT universe to get you the best in that category.
It's going to appear in your regular MySky settings as a new post source.
It's going to appear in your regular MySky settings as a new post source.
September 18, 2026 at 10:02 PM
Soon you will be able to prompt for what you want to see and MySky will instantly do an LLM-powered search across the entire AT universe to get you the best in that category.
It's going to appear in your regular MySky settings as a new post source.
It's going to appear in your regular MySky settings as a new post source.
Writing code and doing related research tasks with Fable has caused me to move up my estimated date for recursive self-improvement (RSI), which is a more precisely defined term than the nebulous "AGI." Anyway, I think we'll see intelligent machines self-improving by late next year.
September 18, 2026 at 4:59 PM
Writing code and doing related research tasks with Fable has caused me to move up my estimated date for recursive self-improvement (RSI), which is a more precisely defined term than the nebulous "AGI." Anyway, I think we'll see intelligent machines self-improving by late next year.
Reposted by Jonathan Stray
Try out this feed!
It’s the result of a massive amount of research on how to build beneficial recommendations. It’s the first real workable implementation of a choose your algorithm system and it will be good for you and the broader ecosystem.
Courtesy of the great @jonathanstray.com
It’s the result of a massive amount of research on how to build beneficial recommendations. It’s the first real workable implementation of a choose your algorithm system and it will be good for you and the broader ecosystem.
Courtesy of the great @jonathanstray.com
Click SETTINGS to personalize your MySky feed.
A feed you control, designed for constructive conversation.
A feed you control, designed for constructive conversation.
September 17, 2026 at 12:32 PM
Try out this feed!
It’s the result of a massive amount of research on how to build beneficial recommendations. It’s the first real workable implementation of a choose your algorithm system and it will be good for you and the broader ecosystem.
Courtesy of the great @jonathanstray.com
It’s the result of a massive amount of research on how to build beneficial recommendations. It’s the first real workable implementation of a choose your algorithm system and it will be good for you and the broader ecosystem.
Courtesy of the great @jonathanstray.com
This is very similar to the way we define "constructive" posts in mysky.social
Social media is often full-scale rhetorical warfare, but that doesn’t mean there shouldn’t be online conflict.
Priya Parker says there's another way: “healthy heat: the willingness to hold tension, with care, long enough for it to become clarifying.”
theisolationjournals.substack.com/cp/215506446
Priya Parker says there's another way: “healthy heat: the willingness to hold tension, with care, long enough for it to become clarifying.”
theisolationjournals.substack.com/cp/215506446
September 16, 2026 at 11:06 PM
This is very similar to the way we define "constructive" posts in mysky.social
Reposted by Jonathan Stray
