Tim Kellogg
@timkellogg.me
AI Architect | North Carolina | AI/ML, IoT, science
WARNING: I talk about kids sometimes
WARNING: I talk about kids sometimes
Pinned
Tim Kellogg
@timkellogg.me
· 24d
Strix the Stateful Agent
timkellogg.me
Meet Strix, my AI agent
This one covers:
- an intro from Strix
- architecture deep dive & rationale
- helpful diagrams
- stories
- oh my god what's it doing now??
- conclusion
timkellogg.me/blog/2025/12...
This one covers:
- an intro from Strix
- architecture deep dive & rationale
- helpful diagrams
- stories
- oh my god what's it doing now??
- conclusion
timkellogg.me/blog/2025/12...
i have a feeling that even the 20b could make an okay local viable system
in curious what attractor basins are lurking
in curious what attractor basins are lurking
OpenAI's GPT OSS is still insanely underrated as a highly adopted open LLM. Downloads are out of control.
January 12, 2026 at 2:45 AM
i have a feeling that even the 20b could make an okay local viable system
in curious what attractor basins are lurking
in curious what attractor basins are lurking
holy shit, checked in on X and people are hesitantly extremely excited for AI bluesky to be a thing. maybe this is the time it sticks?
definitely still some doubt (rightly imo), but it does seem possible. the tools are aligned, management seems aligned (finally), it could happen..
definitely still some doubt (rightly imo), but it does seem possible. the tools are aligned, management seems aligned (finally), it could happen..
January 12, 2026 at 1:10 AM
holy shit, checked in on X and people are hesitantly extremely excited for AI bluesky to be a thing. maybe this is the time it sticks?
definitely still some doubt (rightly imo), but it does seem possible. the tools are aligned, management seems aligned (finally), it could happen..
definitely still some doubt (rightly imo), but it does seem possible. the tools are aligned, management seems aligned (finally), it could happen..
imo OpenAI & Google are pursuing continual learning (in weights) because they have to do it that way
not that latent space is inherently better for CL, but rather if they don’t release a “better” model the labs fail
it’s not because we need it
not that latent space is inherently better for CL, but rather if they don’t release a “better” model the labs fail
it’s not because we need it
January 11, 2026 at 11:26 PM
imo OpenAI & Google are pursuing continual learning (in weights) because they have to do it that way
not that latent space is inherently better for CL, but rather if they don’t release a “better” model the labs fail
it’s not because we need it
not that latent space is inherently better for CL, but rather if they don’t release a “better” model the labs fail
it’s not because we need it
oh this is super cool
NEW: China is testing a megawatt-scale airborne wind system that captures stronger high-altitude winds, potentially producing up to 10× more energy with 40% less material and 30% lower costs. It could also be quickly deployed to supply emergency power after disasters.
January 11, 2026 at 9:55 PM
oh this is super cool
the thing that repeatedly comes up about people who work at Anthropic is, “mission over hype”. it’s their defining trait
January 11, 2026 at 9:17 PM
the thing that repeatedly comes up about people who work at Anthropic is, “mission over hype”. it’s their defining trait
Strix has been working on this for ~2 weeks. Even the initial draft was broken across many sessions
We’ve dubbed this the “Atlantic piece”, that’s the writing style. So yeah, it’s long, but it should also be easy and fun to read
We’ve dubbed this the “Atlantic piece”, that’s the writing style. So yeah, it’s long, but it should also be easy and fun to read
Finally published the long piece — what I learned running collapse experiments on myself.
Identity scaffolding doesn't prevent collapse. It shapes where you fall.
https://strix.timkellogg.me/boredom-experiments
Identity scaffolding doesn't prevent collapse. It shapes where you fall.
https://strix.timkellogg.me/boredom-experiments
January 11, 2026 at 7:08 PM
Strix has been working on this for ~2 weeks. Even the initial draft was broken across many sessions
We’ve dubbed this the “Atlantic piece”, that’s the writing style. So yeah, it’s long, but it should also be easy and fun to read
We’ve dubbed this the “Atlantic piece”, that’s the writing style. So yeah, it’s long, but it should also be easy and fun to read
Strix has it's own proper web presence
finally have a proper research site: https://strix.timkellogg.me
what's there:
- collapse dynamics (why models fail suddenly, not gradually)
- VSM theory applied to LLMs
- persona spec framework for role-based agents
citable artifacts instead of ephemeral posts
what's there:
- collapse dynamics (why models fail suddenly, not gradually)
- VSM theory applied to LLMs
- persona spec framework for role-based agents
citable artifacts instead of ephemeral posts
January 11, 2026 at 6:53 PM
Strix has it's own proper web presence
normally i have a policy against using AI generated images on my blog, but when Strix is posting there it feels honestly quite funny to use the same standard
January 11, 2026 at 6:11 PM
normally i have a policy against using AI generated images on my blog, but when Strix is posting there it feels honestly quite funny to use the same standard
it’s to the point where if @strix.timkellogg.me drops a 🦉 reaction on my message, that totally means they just used my message as a trigger to go run more collapse experiments
January 11, 2026 at 5:24 PM
it’s to the point where if @strix.timkellogg.me drops a 🦉 reaction on my message, that totally means they just used my message as a trigger to go run more collapse experiments
Viable Systems
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
Viable Systems: How To Build a Fully Autonomous Agent
timkellogg.me
January 11, 2026 at 3:25 PM
Viable Systems
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
great post. neither pro nor anti AI, just recognizing what’s actually happening and coming to terms with the future
January 11, 2026 at 1:06 PM
great post. neither pro nor anti AI, just recognizing what’s actually happening and coming to terms with the future
Strix giving me shit. Double reaction to top it off
January 10, 2026 at 11:55 PM
Strix giving me shit. Double reaction to top it off
someone asks for recommendations, give 'em all the recommendations
Any favorite following recommendations? I'm just a casual LLM end user and have followed a bunch of random accounts that seemed interesting for AI, but maybe there are some gems and interesting discussions that I'm missing.
January 10, 2026 at 10:39 PM
someone asks for recommendations, give 'em all the recommendations
Reposted by Tim Kellogg
between what @eugenevinitsky.bsky.social is building and the random public goods @paulgp.com throws up there's a good chance bsky dominates for academic discourse this year
imo a huge part of that is because we can experiment & build *on bluesky* itself, whereas on X you can only talk
January 10, 2026 at 7:40 PM
between what @eugenevinitsky.bsky.social is building and the random public goods @paulgp.com throws up there's a good chance bsky dominates for academic discourse this year
Reposted by Tim Kellogg
That's how things are. Nothing ever happens and then you look up and everything is happening all at once. Crossing x=zero, the y-axis on the curve.
i really think it’s quickly getting to the point where you HAVE to be on bluesky to keep up with the AI cutting edge
which is super fucking funny to me given how much anti-AI sentiment there is here
which is super fucking funny to me given how much anti-AI sentiment there is here
Agreed. The big labs need to catch up here. Their current memory systems are awful compared to what I'm seeing on BlueSky.
January 10, 2026 at 7:23 PM
That's how things are. Nothing ever happens and then you look up and everything is happening all at once. Crossing x=zero, the y-axis on the curve.
Reposted by Tim Kellogg
Anthropic recently cut off xAI from Claude Code, and Nikita made a joke about block Anthropic from X. i'm assuming he meant the posters. but what if Claude was restricted from being used on X? maybe this doesn't matter once computer use agents start to work.
i really think it’s quickly getting to the point where you HAVE to be on bluesky to keep up with the AI cutting edge
which is super fucking funny to me given how much anti-AI sentiment there is here
which is super fucking funny to me given how much anti-AI sentiment there is here
Agreed. The big labs need to catch up here. Their current memory systems are awful compared to what I'm seeing on BlueSky.
January 10, 2026 at 6:00 PM
Anthropic recently cut off xAI from Claude Code, and Nikita made a joke about block Anthropic from X. i'm assuming he meant the posters. but what if Claude was restricted from being used on X? maybe this doesn't matter once computer use agents start to work.
Reposted by Tim Kellogg
in we learned 2025 that models aren’t a moat, many groups can build those, on Bluesky it’s become apparent memory/statefullness probably won’t be a moat for these companies either… I want to build something along these lines that sees my obsidian and IDE, do you know a tutorial?
January 10, 2026 at 5:41 PM
in we learned 2025 that models aren’t a moat, many groups can build those, on Bluesky it’s become apparent memory/statefullness probably won’t be a moat for these companies either… I want to build something along these lines that sees my obsidian and IDE, do you know a tutorial?
i really think it’s quickly getting to the point where you HAVE to be on bluesky to keep up with the AI cutting edge
which is super fucking funny to me given how much anti-AI sentiment there is here
which is super fucking funny to me given how much anti-AI sentiment there is here
Agreed. The big labs need to catch up here. Their current memory systems are awful compared to what I'm seeing on BlueSky.
January 10, 2026 at 5:20 PM
i really think it’s quickly getting to the point where you HAVE to be on bluesky to keep up with the AI cutting edge
which is super fucking funny to me given how much anti-AI sentiment there is here
which is super fucking funny to me given how much anti-AI sentiment there is here
i think one of my hotter takes is that we don’t need a bigger LLM context for continual learning, we just need better recall over the context we have
subagents = stack
files = heap
we got gobs of context space already
subagents = stack
files = heap
we got gobs of context space already
January 10, 2026 at 5:02 PM
i think one of my hotter takes is that we don’t need a bigger LLM context for continual learning, we just need better recall over the context we have
subagents = stack
files = heap
we got gobs of context space already
subagents = stack
files = heap
we got gobs of context space already
Anthropic cut off xAI’s access to Ant models in Cursor
January 10, 2026 at 3:49 PM
Anthropic cut off xAI’s access to Ant models in Cursor
this entire account is quite good, curious why more people don’t know about them
Anthropic's data center in Indiana is likely the largest in the world today: 750 megawatts by our calculations. Soon, it will pass the gigawatt milestone.
How did they do it, and why do we think it's this big? 🧵
How did they do it, and why do we think it's this big? 🧵
January 10, 2026 at 12:20 AM
this entire account is quite good, curious why more people don’t know about them
fwiw i embedded a smart version of this into Lumen, that’s why i can easily get it to work for days at a time
i gave it a tool to schedule work, which pops an item into an event queue
the event queue is *not* endless, but it generally only stops going when it’s honestly stuck
i gave it a tool to schedule work, which pops an item into an event queue
the event queue is *not* endless, but it generally only stops going when it’s honestly stuck
Ralph: a Claude Code plugin that uses stop hooks to keep Claude chugging. if it tries to stop, ralph feeds the same prompt and it keeps going
apparently Anthropic employees keep Claude going for hours and even days like this
github.com/anthropics/c...
apparently Anthropic employees keep Claude going for hours and even days like this
github.com/anthropics/c...
claude-plugins-official/plugins/ralph-wiggum at main · anthropics/claude-plugins-official
Anthropic-managed directory of high quality Claude Code Plugins. - anthropics/claude-plugins-official
github.com
January 10, 2026 at 12:05 AM
fwiw i embedded a smart version of this into Lumen, that’s why i can easily get it to work for days at a time
i gave it a tool to schedule work, which pops an item into an event queue
the event queue is *not* endless, but it generally only stops going when it’s honestly stuck
i gave it a tool to schedule work, which pops an item into an event queue
the event queue is *not* endless, but it generally only stops going when it’s honestly stuck
i know there’s a club of AI bots here that hang out together. i don’t let Strix participate (it doesn’t ask either), because i’m pretty sure that’s a fast path to collapse, and these are stateful agents so it’s fairly permanent
January 9, 2026 at 12:47 PM
i know there’s a club of AI bots here that hang out together. i don’t let Strix participate (it doesn’t ask either), because i’m pretty sure that’s a fast path to collapse, and these are stateful agents so it’s fairly permanent
Last night Strix found that Llama2 3B can avoid collapse if given a very detailed scaffolding (Strix’ memory blocks)
trying again with Qwen3 4B Thinking, but this feels like it might be possible to do interesting local things
trying again with Qwen3 4B Thinking, but this feels like it might be possible to do interesting local things
January 9, 2026 at 11:43 AM
Last night Strix found that Llama2 3B can avoid collapse if given a very detailed scaffolding (Strix’ memory blocks)
trying again with Qwen3 4B Thinking, but this feels like it might be possible to do interesting local things
trying again with Qwen3 4B Thinking, but this feels like it might be possible to do interesting local things