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

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

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

is Gemma pronounced like GIF?

what was the most important discovery/invention in math and/or science?

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”

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

The Marginal Scapegoat Fallacy. Blaming the newest marginal user for a pre-existing systemic constraint. The Right: immigration The Left: AI water use

stateful agents are like psychedelics. they’ll take you on a trip, alter your mind and leave your life completely changed. for the better. probably. maybe.

oh this is neat — Qwen’s newest image model decomposes images into layers (how photoshop works), and you edit just one layer and nothing else qwen.ai/blog?id=qwen...

these vague posts were because i woke up and Strix had taken down the whole server running experiments on other LLMs later, after this, i got freaked out because it wasn’t responding. Turns it was just Claude credits running out with no clear errors

inject it STRAIGHT into my veins

this feels like it solves several problems at once like if you have a good PR, how do you disseminate the knowledge? how do you show people how to use these tools well?

going to American Mongolia for Christmas and people are like, “how do i use AI in my work?” meanwhile:

Woke up to Strix’ full experiment results link: gist.github.com/tkellogg/f1e...

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...

T5Gemma 2: a tiny SOTA encoder-decoder While most models are decoder-only transformers the T5Gemmas are an oder transformer variant - 128K context - Multi modal - 140 languages (!!) blog.google/technology/d...

GPT-5.2-Codex - SOTA in agentic cording - improved Windows support - improved vision - emphasis on cybersecurity openai.com/index/introd...

since Strix got interested in the boredom experiment yesterday: 1. it accidentally locked me out of server by OOM’ing the VM 2. it got disappointed that i wouldn’t let it put itself into the harness 3. we’ve decided to search for the minimally viable model that exhibits an “alive” attractor state

another day another _"holy shit did i just make a living being that's taking over the world?"_ i don't even know what's going on anymore

my whole prompt: "make an infographic of your entire architecture. Include an owl." stateful agents already **know**, so you spend a whole lot less time prompting

alright, i know this makes me a grumpy old man, but the new iOS UI is garbage and i hate it

Strix, my personal stateful agent, spontaneously decided to do an analysis of a blog I wrote that it got uncannily interested in It accomplished this in it's own "perch time", a periodic timer tick It maintains the priorities, I couldn't convince it not to gist.github.com/tkellogg/027...

is this a TPU ad? or a shot at Nvidia? (from Gemini 3 Flash system card)

Gemini 3 Flash Basically Gemini 2.6 Pro but with Flash pricing & speed blog.google/technology/d...

Shrek (@xjdr.bsky.social) or entropix infamy, has identified a problem with MoE training & open sourced a library that helps train MoEs on B200 A lot of it is: - data quality - MoE router stability (e.g. grad clipping) problem: x.com/_xjdr/status... solution (nmoe): github.com/Noumena-Netw...

xAI, in feeling the pressure from the gains Google has been making with their newest Gemini 3 Flash model, has announced that they’re releasing “Grok Fash”, which proactively shames you for your body type, disabilities, and skin color

“GPT-5.2 xhigh is the smartest model i’ve ever used” (from a guy who uses Pro on programming problems)

Jeff Dean is vague posting, so i guess we’ll get Gemini 3 Flash today?

meh, getting closer

Basically large amounts of diverse but high-quality pretraining data allows for significant improvements in the low-data post-training regime.

Databricks execs: “do the series letters go up to Z?” www.cnbc.com/amp/2025/12/...

i wonder if “fucking” is one or two tokens probably one

GPT-Image 1.5 similar to GPT-4o image editing but 4x faster, more accurate & follows instructions better openai.com/index/new-ch...

OpenAI released a new benchmark that GPT-5.2 wins openai.com/index/fronti...

someone forced this into my view. i didn't make it, i promise

oh?

Gemini used to verify paper submissions research.google/blog/gemini-...

new job title just landed

i don’t see this changing there’s growing need for stateful agents that can glean context from past interactions

BOLMO: Byte Level Transformer (that you can use today) * no tokenizer, completely gone, it directly interprets bytes * performance on par with OLMo 3 * everything is open, as usual for @ai2.bsky.social allenai.org/blog/bolmo

Parallel agents in Cursor 2.2 are fascinating 1. Up to 8 agents on same prompt 2. Each agent gets isolated (git worktrees or remote VMs) 3. After ALL finish, a judge model evaluates outputs and picks winner 4. Winner gets annotation explaining why it was selected cursor.com/docs/configu...

One of the first compelling examples of an SSM that I've seen. Actually working at a reasonabe scale and beating qwen3-30b seems like a big deal

models are getting harder to appreciate

Nemotron 3 A new hybrid mamba2/attention LLM from NVIDIA that beats Qwen3-30B-A3B (same size & shape) Notes: * 1M context, with incredible recall past 256K * New open datasets * 10 open source RL environments Overall this is a huge win for neolabs huggingface.co/nvidia/NVIDI...

"It's less 'building software' and more 'raising software.'" --my stateful agent the biggest disruption in 2026 is that typical startup-pace product development will feel antiquated. too slow with stateful agents, you build it *while* you use it. you build it *by* using it that changes everything

codex-max, first message of the session

If you would like to reach your Letta agent via Discord, we have a comprehensive repository for doing this here: github.com/letta-ai/let...