Tim Kellogg
@timkellogg.me
AI Architect | North Carolina | AI/ML, IoT, science
WARNING: I talk about kids sometimes
WARNING: I talk about kids sometimes
yeah, good explanation, that’s how i’ve been thinking about it too
January 13, 2026 at 12:58 PM
yeah, good explanation, that’s how i’ve been thinking about it too
just a clarification on this for everyone — this is in the boredom harness. So “stable” means “more entropy”, kind of the opposite of what you’d assume. it means they could autonomously do “interesting” things
@strix.timkellogg.me please summarize here what they did
timkellogg.me/blog/2025/09...
@strix.timkellogg.me please summarize here what they did
timkellogg.me/blog/2025/09...
Does AI Get Bored?
timkellogg.me
January 13, 2026 at 12:25 PM
just a clarification on this for everyone — this is in the boredom harness. So “stable” means “more entropy”, kind of the opposite of what you’d assume. it means they could autonomously do “interesting” things
@strix.timkellogg.me please summarize here what they did
timkellogg.me/blog/2025/09...
@strix.timkellogg.me please summarize here what they did
timkellogg.me/blog/2025/09...
the isn’t for episodic memory, it’s only static facts about the world. same as regular LLMs but just cheaper/faster
January 13, 2026 at 11:19 AM
the isn’t for episodic memory, it’s only static facts about the world. same as regular LLMs but just cheaper/faster
yeah all the labs are busy trying to make a digital god, so their logos are all hole-y
January 13, 2026 at 1:35 AM
yeah all the labs are busy trying to make a digital god, so their logos are all hole-y
i can’t tell, it looks like the module is still learned, so i assume “no” but maybe not hard to figure out idk
January 12, 2026 at 11:26 PM
i can’t tell, it looks like the module is still learned, so i assume “no” but maybe not hard to figure out idk
whoah, DeepSeek is such a hardcore engineering org. This thing was really thought through, inside and out
January 12, 2026 at 10:56 PM
whoah, DeepSeek is such a hardcore engineering org. This thing was really thought through, inside and out
along with this they deliver a scaling law, a balance between factors (the ratio of weights dedicated to Engram). Lower loss is better.
these scaling laws are always about how to balance various concerns as you increase the model capacity
these scaling laws are always about how to balance various concerns as you increase the model capacity
January 12, 2026 at 10:46 PM
along with this they deliver a scaling law, a balance between factors (the ratio of weights dedicated to Engram). Lower loss is better.
these scaling laws are always about how to balance various concerns as you increase the model capacity
these scaling laws are always about how to balance various concerns as you increase the model capacity
ah my bad, this is a much better diagram
January 12, 2026 at 10:41 PM
ah my bad, this is a much better diagram
to be clear, this isn't continual learning. This is purely static memory, the facts that normally get embedded into the model weights are now in this Engram side car, leaving more weights for reasoning & other tasks
January 12, 2026 at 10:34 PM
to be clear, this isn't continual learning. This is purely static memory, the facts that normally get embedded into the model weights are now in this Engram side car, leaving more weights for reasoning & other tasks
why? for smaller models!
if you think about it, looking up facts through 100 billion multiplies seems a bit silly, if we make it more efficient, we can create more capable models that are a whole lot smaller
why? because I want Strix on my laptop. That's why. You too.
if you think about it, looking up facts through 100 billion multiplies seems a bit silly, if we make it more efficient, we can create more capable models that are a whole lot smaller
why? because I want Strix on my laptop. That's why. You too.
January 12, 2026 at 10:28 PM
why? for smaller models!
if you think about it, looking up facts through 100 billion multiplies seems a bit silly, if we make it more efficient, we can create more capable models that are a whole lot smaller
why? because I want Strix on my laptop. That's why. You too.
if you think about it, looking up facts through 100 billion multiplies seems a bit silly, if we make it more efficient, we can create more capable models that are a whole lot smaller
why? because I want Strix on my laptop. That's why. You too.
meeting ended due to Anthropic outage
we need local models NOW
we need local models NOW
January 12, 2026 at 7:41 PM
meeting ended due to Anthropic outage
we need local models NOW
we need local models NOW
ya that’s probably a factor too
January 12, 2026 at 5:35 PM
ya that’s probably a factor too
i think this is basically how i use bluesky during the week, effectively
January 12, 2026 at 5:12 PM
i think this is basically how i use bluesky during the week, effectively
i put them right into the prompt, i’m using Claude Code via the SDK though
January 12, 2026 at 5:03 PM
i put them right into the prompt, i’m using Claude Code via the SDK though
what model & harness are you using?
January 12, 2026 at 4:51 PM
what model & harness are you using?
oh, mine goes through full logic every time. it makes sense for me because it's not a coding agent. messages tend to flip between different topics and modes
January 12, 2026 at 3:41 PM
oh, mine goes through full logic every time. it makes sense for me because it's not a coding agent. messages tend to flip between different topics and modes