Jake Gold
@jacob.gold
Infra eng leader, SF Bay Area. Helped launch and scale Bluesky. Prev: Nuro, Docker, Google, and founder.
Obsessed with history, computers, and open systems.
Happy to chat: j@jacob.gold
Obsessed with history, computers, and open systems.
Happy to chat: j@jacob.gold
Pinned
Jake Gold
@jacob.gold
· Nov 11
An unordered / non-exhaustive list of things that helped scale Bluesky's infra efficiently:
+ Exiting the cloud (colocation)
+ HAProxy w/many Node backends
+ Go w/clever code
+ ScyllaDB
+ SQLite w/per user databases
+ Redis w/many instances
+ AMD servers w/many cores
+ Purchasing bandwidth directly
+ Exiting the cloud (colocation)
+ HAProxy w/many Node backends
+ Go w/clever code
+ ScyllaDB
+ SQLite w/per user databases
+ Redis w/many instances
+ AMD servers w/many cores
+ Purchasing bandwidth directly
Seems like the right frame of mind to me. We probably want more of our own code to look like it's an open source library created by a third-party, that way it's not tightly coupled to our own environment, etc.
February 10, 2026 at 9:01 PM
Seems like the right frame of mind to me. We probably want more of our own code to look like it's an open source library created by a third-party, that way it's not tightly coupled to our own environment, etc.
Coding agents are good at gluing together libraries but also good at creating them.
Keeping things self-contained and modular was already one of the best ways of taming complexity but it's become even more important now.
Keeping things self-contained and modular was already one of the best ways of taming complexity but it's become even more important now.
February 10, 2026 at 8:58 PM
Coding agents are good at gluing together libraries but also good at creating them.
Keeping things self-contained and modular was already one of the best ways of taming complexity but it's become even more important now.
Keeping things self-contained and modular was already one of the best ways of taming complexity but it's become even more important now.
Agreed. I think "make something people want" might be even harder now that more people can "make".
February 10, 2026 at 6:40 PM
Agreed. I think "make something people want" might be even harder now that more people can "make".
It's easier to generate lots of good code but it's also easier to generate lots of bad code.
The new skill to learn is how to maintain discipline and quality while using coding agents.
The new skill to learn is how to maintain discipline and quality while using coding agents.
February 10, 2026 at 6:34 PM
It's easier to generate lots of good code but it's also easier to generate lots of bad code.
The new skill to learn is how to maintain discipline and quality while using coding agents.
The new skill to learn is how to maintain discipline and quality while using coding agents.
Now that we have LLM coding agents:
* The value of simple code is ~$0.
* The value of high quality and necessarily-complex code is higher than ever.
* The value of simple code is ~$0.
* The value of high quality and necessarily-complex code is higher than ever.
Software is created for accumulation of knowledge. AI is not going to cancel this fact. Forgot the idea that programs will be prompts (specifications). The details is what really matters, and they are hard to capture textually than in the code. New projects will be spec + code, evolving.
February 10, 2026 at 6:22 PM
Now that we have LLM coding agents:
* The value of simple code is ~$0.
* The value of high quality and necessarily-complex code is higher than ever.
* The value of simple code is ~$0.
* The value of high quality and necessarily-complex code is higher than ever.
That's possible but I don't think so. It largely depends on how AI progress develops.
I expect inference for most of what users want day-to-day to become (almost) too cheap to meter within a few years.
And with more time, a lot will move to local devices/models and become essentially free.
I expect inference for most of what users want day-to-day to become (almost) too cheap to meter within a few years.
And with more time, a lot will move to local devices/models and become essentially free.
February 9, 2026 at 9:52 PM
That's possible but I don't think so. It largely depends on how AI progress develops.
I expect inference for most of what users want day-to-day to become (almost) too cheap to meter within a few years.
And with more time, a lot will move to local devices/models and become essentially free.
I expect inference for most of what users want day-to-day to become (almost) too cheap to meter within a few years.
And with more time, a lot will move to local devices/models and become essentially free.
OpenAI's self-deception and rationalization about injecting ads into ChatGPT is something to behold.
Ads are corrupting and the proof is traditional media, social media, and Google.
Fortunately OpenAI has serious competition so users won't have to live with ads for very long.
Ads are corrupting and the proof is traditional media, social media, and Google.
Fortunately OpenAI has serious competition so users won't have to live with ads for very long.
Testing ads in ChatGPT
OpenAI begins testing ads in ChatGPT to support free access, with clear labeling, answer independence, strong privacy protections, and user control.
openai.com
February 9, 2026 at 9:44 PM
OpenAI's self-deception and rationalization about injecting ads into ChatGPT is something to behold.
Ads are corrupting and the proof is traditional media, social media, and Google.
Fortunately OpenAI has serious competition so users won't have to live with ads for very long.
Ads are corrupting and the proof is traditional media, social media, and Google.
Fortunately OpenAI has serious competition so users won't have to live with ads for very long.
Yeah, Google is competing effectively now. Gemini is in the race. But Gemini only exists because ChatGPT exists.
February 8, 2026 at 1:19 AM
Yeah, Google is competing effectively now. Gemini is in the race. But Gemini only exists because ChatGPT exists.
Feels like Google would have gotten here, easily, if it had 10% of the competition that OpenAI attracted with ChatGPT.
In retrospect, it's obvious how much Google sucked in comparison to what was possible.
It's good that Google's decades of stagnation and taxing the internet are coming to an end.
In retrospect, it's obvious how much Google sucked in comparison to what was possible.
It's good that Google's decades of stagnation and taxing the internet are coming to an end.
thank u mr claude for searching 500 forum posts in 7 minutes to identify the exact issue with the specific wifi chip in my specific motherboard that causes it to keep dying randomly
star trek moment
star trek moment
February 8, 2026 at 12:39 AM
Feels like Google would have gotten here, easily, if it had 10% of the competition that OpenAI attracted with ChatGPT.
In retrospect, it's obvious how much Google sucked in comparison to what was possible.
It's good that Google's decades of stagnation and taxing the internet are coming to an end.
In retrospect, it's obvious how much Google sucked in comparison to what was possible.
It's good that Google's decades of stagnation and taxing the internet are coming to an end.
Yeah, I think that's a good way to do it.
February 7, 2026 at 10:04 PM
Yeah, I think that's a good way to do it.
Using a subdomain for TXT can be ideal. For non-technical users it might be a little harder though so not sure it's worth it.
The CNAME on the apex was just an example but you're right not everywhere allows that or supports the "virtual" ALIAS type so maybe not the best example.
The CNAME on the apex was just an example but you're right not everywhere allows that or supports the "virtual" ALIAS type so maybe not the best example.
February 7, 2026 at 8:42 PM
Using a subdomain for TXT can be ideal. For non-technical users it might be a little harder though so not sure it's worth it.
The CNAME on the apex was just an example but you're right not everywhere allows that or supports the "virtual" ALIAS type so maybe not the best example.
The CNAME on the apex was just an example but you're right not everywhere allows that or supports the "virtual" ALIAS type so maybe not the best example.
Apps that get DNS verification functionality wrong really cause a lot of pain for their users.
Having explained the problem a number of times, I finally decided to write it down so now I can just share a link.
Having explained the problem a number of times, I finally decided to write it down so now I can just share a link.
Stop Telling Users Their DNS Is Wrong
If you’re like me, you’ve had the experience where you add a DNS record for a service. You triple check it. The app still says it’s invalid. You wait. You check again. Still invalid. You start wonderi...
jacob.gold
February 7, 2026 at 6:02 PM
Apps that get DNS verification functionality wrong really cause a lot of pain for their users.
Having explained the problem a number of times, I finally decided to write it down so now I can just share a link.
Having explained the problem a number of times, I finally decided to write it down so now I can just share a link.
Don't think the TUI lets you do it but `max` effort is wildly expensive FYI.
February 6, 2026 at 6:12 PM
Don't think the TUI lets you do it but `max` effort is wildly expensive FYI.
I assume it's mostly just information overload. World is too busy for people to track it all accurately.
February 6, 2026 at 6:10 PM
I assume it's mostly just information overload. World is too busy for people to track it all accurately.
IMHO this is nothing like Elon's highly deceptive robotaxi/robot demos which were remotely operated.
When a Waymo gets stuck, which is a small % of its driving time, a remote assistant can click waypoints on a map to get it unstuck.
I'd favor local assistants but even 500+ ms RTT wouldn't matter.
When a Waymo gets stuck, which is a small % of its driving time, a remote assistant can click waypoints on a map to get it unstuck.
I'd favor local assistants but even 500+ ms RTT wouldn't matter.
Once again, it turns out “fully autonomous” means “a guy in the Philippines.”
It Turns Out Waymos Are Being Controlled by Workers in the Philippines
During a Congressional hearing, Waymo's chief safety officer, Mauricio Peña, was grilled over the company's reliance on overseas workers.
futurism.com
February 6, 2026 at 5:35 PM
IMHO this is nothing like Elon's highly deceptive robotaxi/robot demos which were remotely operated.
When a Waymo gets stuck, which is a small % of its driving time, a remote assistant can click waypoints on a map to get it unstuck.
I'd favor local assistants but even 500+ ms RTT wouldn't matter.
When a Waymo gets stuck, which is a small % of its driving time, a remote assistant can click waypoints on a map to get it unstuck.
I'd favor local assistants but even 500+ ms RTT wouldn't matter.
For now LLM coding agents are capable of generating torrents of tech debt that they're not capable of fixing.
Team #1 bets on software engineering discipline keeping tech debt at a minimum and not slowing them down
Team #2 bets on the coding agents bailing them out before tech debt slows them down
Team #1 bets on software engineering discipline keeping tech debt at a minimum and not slowing them down
Team #2 bets on the coding agents bailing them out before tech debt slows them down
February 5, 2026 at 11:39 PM
For now LLM coding agents are capable of generating torrents of tech debt that they're not capable of fixing.
Team #1 bets on software engineering discipline keeping tech debt at a minimum and not slowing them down
Team #2 bets on the coding agents bailing them out before tech debt slows them down
Team #1 bets on software engineering discipline keeping tech debt at a minimum and not slowing them down
Team #2 bets on the coding agents bailing them out before tech debt slows them down
An open source project could replace it really easily.
The API is public and very simple.
I never felt like I could justify spending more than like 15 minutes at a time on it (too much was going on) so the code is bad enough that I would just replace it anyway.
The API is public and very simple.
I never felt like I could justify spending more than like 15 minutes at a time on it (too much was going on) so the code is bad enough that I would just replace it anyway.
February 5, 2026 at 5:26 PM
An open source project could replace it really easily.
The API is public and very simple.
I never felt like I could justify spending more than like 15 minutes at a time on it (too much was going on) so the code is bad enough that I would just replace it anyway.
The API is public and very simple.
I never felt like I could justify spending more than like 15 minutes at a time on it (too much was going on) so the code is bad enough that I would just replace it anyway.
It's 2026, so I now say "If you believe that, I've got a datacenter in space to sell you."
February 4, 2026 at 2:04 AM
It's 2026, so I now say "If you believe that, I've got a datacenter in space to sell you."
Even Zuck can barely make a new social network happen with a billion+ captive users.
Threads is sort of hollow and fake because of how forced it was. It's more like a feature of Instagram than a proper standalone network.
The fact that Bluesky is self-sustaining, open, and scaled is unprecedented.
Threads is sort of hollow and fake because of how forced it was. It's more like a feature of Instagram than a proper standalone network.
The fact that Bluesky is self-sustaining, open, and scaled is unprecedented.
January 28, 2026 at 10:11 PM
Even Zuck can barely make a new social network happen with a billion+ captive users.
Threads is sort of hollow and fake because of how forced it was. It's more like a feature of Instagram than a proper standalone network.
The fact that Bluesky is self-sustaining, open, and scaled is unprecedented.
Threads is sort of hollow and fake because of how forced it was. It's more like a feature of Instagram than a proper standalone network.
The fact that Bluesky is self-sustaining, open, and scaled is unprecedented.
Yeah, Bluesky is already important enough for spammers and scammers, so we got that going for us which is nice.
January 28, 2026 at 10:09 PM
Yeah, Bluesky is already important enough for spammers and scammers, so we got that going for us which is nice.
We really need subcommunities so that groups can form in a self-sustaining way. One of the strongest kinds of groups are geographic/language-based.
January 28, 2026 at 10:09 PM
We really need subcommunities so that groups can form in a self-sustaining way. One of the strongest kinds of groups are geographic/language-based.
It's counterintuitive, but Bluesky could be bigger than X/Threads in ~18 months of a change that fixes retention.
The non-replicable act of bootstrapping the network already succeeded!
Now we're just a product change away (e.g. adding subcommunities, modal video/image feeds) of exponential growth.
The non-replicable act of bootstrapping the network already succeeded!
Now we're just a product change away (e.g. adding subcommunities, modal video/image feeds) of exponential growth.
helpful heuristic: actual active user statistics for Bluesky that accounts for lurkers is firehose-visible active users times two
so actual MAU is around 12M, DAU around 3M
bskycharts.edavis.dev/edavis.dev/b...
so actual MAU is around 12M, DAU around 3M
bskycharts.edavis.dev/edavis.dev/b...
January 28, 2026 at 8:06 PM
It's counterintuitive, but Bluesky could be bigger than X/Threads in ~18 months of a change that fixes retention.
The non-replicable act of bootstrapping the network already succeeded!
Now we're just a product change away (e.g. adding subcommunities, modal video/image feeds) of exponential growth.
The non-replicable act of bootstrapping the network already succeeded!
Now we're just a product change away (e.g. adding subcommunities, modal video/image feeds) of exponential growth.
Zed seems like it's doing a good job of implementing modern features (like LLM-based predictions) without me having to mess with plugins.
But I might end up back at vim or neovim.
But I might end up back at vim or neovim.
January 27, 2026 at 4:26 PM
Zed seems like it's doing a good job of implementing modern features (like LLM-based predictions) without me having to mess with plugins.
But I might end up back at vim or neovim.
But I might end up back at vim or neovim.
I want to bind some ctrl keys, like ctrl-o to OpenSelectedFilename (or whatever opens the fuzzy file open search)
January 27, 2026 at 2:05 AM
I want to bind some ctrl keys, like ctrl-o to OpenSelectedFilename (or whatever opens the fuzzy file open search)