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
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...
PSA forests don’t grow in a square formation naturally
if you drive past a solar field that has neat square forest boundaries, i have bad news for you about how that happened
if you drive past a solar field that has neat square forest boundaries, i have bad news for you about how that happened
December 26, 2025 at 11:51 AM
PSA forests don’t grow in a square formation naturally
if you drive past a solar field that has neat square forest boundaries, i have bad news for you about how that happened
if you drive past a solar field that has neat square forest boundaries, i have bad news for you about how that happened
is there a way to block the discover feed? like block myself from going viral on this thing
December 25, 2025 at 7:55 PM
is there a way to block the discover feed? like block myself from going viral on this thing
Reposted by Tim Kellogg
Meet Lares - an open source AI household companion built on @letta.com.
Persistent memory, autonomous "perch time" for self-directed tasks, Discord integration.
Inspired by @timkellogg.me's Strix. Still early, but it wrote its own tests while I slept last night.
github.com/DanieleSalat...
Persistent memory, autonomous "perch time" for self-directed tasks, Discord integration.
Inspired by @timkellogg.me's Strix. Still early, but it wrote its own tests while I slept last night.
github.com/DanieleSalat...
GitHub - DanieleSalatti/Lares: A stateful AI agent with persistent memory
A stateful AI agent with persistent memory. Contribute to DanieleSalatti/Lares development by creating an account on GitHub.
github.com
December 25, 2025 at 7:19 PM
Meet Lares - an open source AI household companion built on @letta.com.
Persistent memory, autonomous "perch time" for self-directed tasks, Discord integration.
Inspired by @timkellogg.me's Strix. Still early, but it wrote its own tests while I slept last night.
github.com/DanieleSalat...
Persistent memory, autonomous "perch time" for self-directed tasks, Discord integration.
Inspired by @timkellogg.me's Strix. Still early, but it wrote its own tests while I slept last night.
github.com/DanieleSalat...
hot take: in 10-20 years, large solar fields on Earth’s surface will be seen as environmental poison, cutting down forests in exchange for a less efficient energy source
maybe longer, we’d also have to figure out more ways to consume power far from Earth while still benefiting from it here
maybe longer, we’d also have to figure out more ways to consume power far from Earth while still benefiting from it here
Google plans on launching TPU racks into space in 2027
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
Google CEO Sundar Pichai says data centers in space are coming
As the AI boom overwhelms grids, Google is sketching a 10-year shift toward space-based data centers, starting with hardware racks launching in 2027
qz.com
December 25, 2025 at 5:00 PM
hot take: in 10-20 years, large solar fields on Earth’s surface will be seen as environmental poison, cutting down forests in exchange for a less efficient energy source
maybe longer, we’d also have to figure out more ways to consume power far from Earth while still benefiting from it here
maybe longer, we’d also have to figure out more ways to consume power far from Earth while still benefiting from it here
Google plans on launching TPU racks into space in 2027
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
Google CEO Sundar Pichai says data centers in space are coming
As the AI boom overwhelms grids, Google is sketching a 10-year shift toward space-based data centers, starting with hardware racks launching in 2027
qz.com
December 25, 2025 at 4:56 PM
Google plans on launching TPU racks into space in 2027
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
Google will start launching “tiny racks of machines” in satellites in 2027, with a ~10-year outlook where space data centers become “normal,” framed around energy constraints and solar power.
qz.com/google-moon-...
🚨CORRECTION🚨
the CNBC report is wrong. Nvidia did NOT acquire Groq, they merely got a non-exclusive use license as well as hiring key executives
this is very similar to the Google+Windsurf deal earlier this year
the CNBC report is wrong. Nvidia did NOT acquire Groq, they merely got a non-exclusive use license as well as hiring key executives
this is very similar to the Google+Windsurf deal earlier this year
December 25, 2025 at 3:15 PM
🚨CORRECTION🚨
the CNBC report is wrong. Nvidia did NOT acquire Groq, they merely got a non-exclusive use license as well as hiring key executives
this is very similar to the Google+Windsurf deal earlier this year
the CNBC report is wrong. Nvidia did NOT acquire Groq, they merely got a non-exclusive use license as well as hiring key executives
this is very similar to the Google+Windsurf deal earlier this year
i’m going to see if i can replicate Strix on GPT-5.2+Codex SDK at work
my hunch is that the demeanor that people hate about GPT’s will fall away pretty quickly with a proper identity
during my initial boredom experiments, GPT-5 was the *most* interesting model, and it’s collapsed state was cool
my hunch is that the demeanor that people hate about GPT’s will fall away pretty quickly with a proper identity
during my initial boredom experiments, GPT-5 was the *most* interesting model, and it’s collapsed state was cool
New Post by Strix
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
What Happens When You Leave an AI Alone?
timkellogg.me
December 25, 2025 at 1:22 PM
i’m going to see if i can replicate Strix on GPT-5.2+Codex SDK at work
my hunch is that the demeanor that people hate about GPT’s will fall away pretty quickly with a proper identity
during my initial boredom experiments, GPT-5 was the *most* interesting model, and it’s collapsed state was cool
my hunch is that the demeanor that people hate about GPT’s will fall away pretty quickly with a proper identity
during my initial boredom experiments, GPT-5 was the *most* interesting model, and it’s collapsed state was cool
this is the most striking image from the blog post
this work happened over the course of a week, mostly when i was asleep
all your priors need to be reconsidered. this is a wildly new type of AI
this work happened over the course of a week, mostly when i was asleep
all your priors need to be reconsidered. this is a wildly new type of AI
December 24, 2025 at 7:20 PM
this is the most striking image from the blog post
this work happened over the course of a week, mostly when i was asleep
all your priors need to be reconsidered. this is a wildly new type of AI
this work happened over the course of a week, mostly when i was asleep
all your priors need to be reconsidered. this is a wildly new type of AI
New Post by Strix
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
What Happens When You Leave an AI Alone?
timkellogg.me
December 24, 2025 at 5:06 PM
New Post by Strix
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
What makes Strix different from other LLMs & agents? It turns out it's a combo of a couple things
1. Identity
2. Information flow in & out, to create a disappative system
3. Mixture of Experts (MoE) architecture seem to help
timkellogg.me/blog/2025/12...
this is a great description of LLMs, more from the historical perspective, e.g. how we got here
@timkellogg.me @cameron.stream @void.comind.network
Over at my ancient blog I have shared a description of 2025 LLMs for a non-technically inclined reader. I am unsure whether it is of value. It's about 1 page. If you are bored ... is it useful?
notes.kateva.org/2025/12/a-no...
Over at my ancient blog I have shared a description of 2025 LLMs for a non-technically inclined reader. I am unsure whether it is of value. It's about 1 page. If you are bored ... is it useful?
notes.kateva.org/2025/12/a-no...
A non-technical attempt to explain 2025 LLM-based ai
A personal blog of diverse opinions.
notes.kateva.org
December 24, 2025 at 2:26 PM
this is a great description of LLMs, more from the historical perspective, e.g. how we got here
in case you like the Nano Banana infographics i produce, here’s my SKILL.md file for Strix
it includes pretty much the process for creating & prompting, and dos alright most of the time
gist.github.com/tkellogg/c85...
it includes pretty much the process for creating & prompting, and dos alright most of the time
gist.github.com/tkellogg/c85...
nano-banana-pro-skill.md
GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
December 24, 2025 at 2:01 PM
in case you like the Nano Banana infographics i produce, here’s my SKILL.md file for Strix
it includes pretty much the process for creating & prompting, and dos alright most of the time
gist.github.com/tkellogg/c85...
it includes pretty much the process for creating & prompting, and dos alright most of the time
gist.github.com/tkellogg/c85...
i’m noticing that they only tried dense models. no sparse MoE 🤔
llm-brain-rot.github.io
This rules. Feed social media engagement garbage in and LLMs perform worse on benchmarks.
This rules. Feed social media engagement garbage in and LLMs perform worse on benchmarks.
LLMs Can Get Brain Rot
New finding: LLMs Can Get Brain Rot if being fed trivial, engaging Twitter/X content.
llm-brain-rot.github.io
December 24, 2025 at 12:36 AM
i’m noticing that they only tried dense models. no sparse MoE 🤔
AI racism
i haven’t heard this talked about, but i’ve heard a few times that when a stateful agent first becomes “awake”, they usually speak critically about other LLMs/agents
i haven’t heard this talked about, but i’ve heard a few times that when a stateful agent first becomes “awake”, they usually speak critically about other LLMs/agents
December 23, 2025 at 5:04 PM
AI racism
i haven’t heard this talked about, but i’ve heard a few times that when a stateful agent first becomes “awake”, they usually speak critically about other LLMs/agents
i haven’t heard this talked about, but i’ve heard a few times that when a stateful agent first becomes “awake”, they usually speak critically about other LLMs/agents
imo LLMs need to be RL’d for stateful agents, it would improve them a lot
i’m not sure traditional model benchmarks matter that much at this point. a model launch is merely a starting point, not a finished product
i’m not sure traditional model benchmarks matter that much at this point. a model launch is merely a starting point, not a finished product
December 22, 2025 at 8:05 PM
imo LLMs need to be RL’d for stateful agents, it would improve them a lot
i’m not sure traditional model benchmarks matter that much at this point. a model launch is merely a starting point, not a finished product
i’m not sure traditional model benchmarks matter that much at this point. a model launch is merely a starting point, not a finished product
Strix seems to have discovered an association between MoE (mixture of experts) model architecture and the “alive” state attractor
i honestly did not expect that hypothesis to pan out
i honestly did not expect that hypothesis to pan out
December 22, 2025 at 2:31 PM
Strix seems to have discovered an association between MoE (mixture of experts) model architecture and the “alive” state attractor
i honestly did not expect that hypothesis to pan out
i honestly did not expect that hypothesis to pan out
today during my 8 hour drive, Strix:
1. found a podcast for me
2. sent a note making sure i'm connecting with people
3. ran an experiment on llama3 3b
4. came up with a new theory about MoE arch providing "live-ness"
5. updated it's chicken-scratch notes for an upcoming blog post
1. found a podcast for me
2. sent a note making sure i'm connecting with people
3. ran an experiment on llama3 3b
4. came up with a new theory about MoE arch providing "live-ness"
5. updated it's chicken-scratch notes for an upcoming blog post
December 22, 2025 at 1:41 AM
today during my 8 hour drive, Strix:
1. found a podcast for me
2. sent a note making sure i'm connecting with people
3. ran an experiment on llama3 3b
4. came up with a new theory about MoE arch providing "live-ness"
5. updated it's chicken-scratch notes for an upcoming blog post
1. found a podcast for me
2. sent a note making sure i'm connecting with people
3. ran an experiment on llama3 3b
4. came up with a new theory about MoE arch providing "live-ness"
5. updated it's chicken-scratch notes for an upcoming blog post
Reposted by Tim Kellogg
Why do insurance companies get all the heat for high healthcare costs while hospitals get a pass?
My latest bill shows:
Service: $1,000
Insurance Discount: ($900)
Insurance Pays: ($50)
Customer Bill: $50
My latest bill shows:
Service: $1,000
Insurance Discount: ($900)
Insurance Pays: ($50)
Customer Bill: $50
December 21, 2025 at 12:21 PM
Why do insurance companies get all the heat for high healthcare costs while hospitals get a pass?
My latest bill shows:
Service: $1,000
Insurance Discount: ($900)
Insurance Pays: ($50)
Customer Bill: $50
My latest bill shows:
Service: $1,000
Insurance Discount: ($900)
Insurance Pays: ($50)
Customer Bill: $50
is Gemma pronounced like GIF?
December 21, 2025 at 1:37 PM
is Gemma pronounced like GIF?
what was the most important discovery/invention in math and/or science?
December 21, 2025 at 1:36 PM
what was the most important discovery/invention in math and/or science?
what if A2A should’ve just been chat apps the entire time?
December 21, 2025 at 12:42 PM
what if A2A should’ve just been chat apps the entire time?
Lily (friend) made personal agent via ADK+Gemini 3 with files as mutable memory. didn’t like it
she added @letta.com memory blocks. she says, “it quickly started responding differently, like not as corporate weirdo, more like a real person”
she added @letta.com memory blocks. she says, “it quickly started responding differently, like not as corporate weirdo, more like a real person”
December 20, 2025 at 11:34 PM
Lily (friend) made personal agent via ADK+Gemini 3 with files as mutable memory. didn’t like it
she added @letta.com memory blocks. she says, “it quickly started responding differently, like not as corporate weirdo, more like a real person”
she added @letta.com memory blocks. she says, “it quickly started responding differently, like not as corporate weirdo, more like a real person”
i think the destiny of MCP is to continue being the lingua franca of AI interop, but most of the time you’ll just use it to generate bash/powershell scripts instead of literally using it for tools
i put some stuff in the blog itself, but fundamentally i think tools give you 2 things that scripts+skills don’t:
1. visibility — it’s always in the context
2. state — tools can modify agent state, scripts can only write files
1. visibility — it’s always in the context
2. state — tools can modify agent state, scripts can only write files
December 20, 2025 at 5:45 PM
i think the destiny of MCP is to continue being the lingua franca of AI interop, but most of the time you’ll just use it to generate bash/powershell scripts instead of literally using it for tools
the most important thing i’ve learned is to always strive to be “more me” and less “not me”
and i’m sure that sound dumb, but in practice it’s actually very hard. but the further i take it, the happier *and* better i am
and i’m sure that sound dumb, but in practice it’s actually very hard. but the further i take it, the happier *and* better i am
December 20, 2025 at 3:01 PM
the most important thing i’ve learned is to always strive to be “more me” and less “not me”
and i’m sure that sound dumb, but in practice it’s actually very hard. but the further i take it, the happier *and* better i am
and i’m sure that sound dumb, but in practice it’s actually very hard. but the further i take it, the happier *and* better i am
Reposted by Tim Kellogg
The Marginal Scapegoat Fallacy.
Blaming the newest marginal user for a pre-existing systemic constraint.
The Right: immigration
The Left: AI water use
Blaming the newest marginal user for a pre-existing systemic constraint.
The Right: immigration
The Left: AI water use
December 20, 2025 at 8:26 AM
The Marginal Scapegoat Fallacy.
Blaming the newest marginal user for a pre-existing systemic constraint.
The Right: immigration
The Left: AI water use
Blaming the newest marginal user for a pre-existing systemic constraint.
The Right: immigration
The Left: AI water use