@mza
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States Android App
Account-level information from X, not a live location or the device used for a specific post.
Chief AI & Technology Officer, AWS
Seattle, WA
Joined January 2007
- Tweets7.8K
- Following803
- Followers26.6K
- Likes942
mza retweeted
Today we’re unveiling Odyssey-3, a big step forward for foundation world models.
It can control robots, power humanoids, drive cars (on the roads of India!), train AIs, pilot drones, and even play video games.
We can’t wait to see what intelligent systems it enables.
mza retweeted
I love reading Werner's blog because of how he is able to break things down to their essence. In his use of Kiro Crew he picked out patterns in its memory architecture around how it decides what to keep, what to compress, and what to prune. While the team did this for practical reasons, he notices similarities to how the brain works. Specifically he compares Kiro Crew's memory architecture of distinct memory types, background consolidation, and decay to what Hawkins describes in A Thousand Brains. Memory is what lets an agent earn the autonomy to work unattended. You review less over time not because you check less, but because it remembers your rules.
That memory architecture is core to the personal Kiro Crew system I have described in recent tweets. You should try it.
Since launch I've been spending time with Kiro Crew, and its memory system is what holds my attention. What it keeps, what it compresses, what it lets go. Our engineers started from pure engineering constraints and landed where evolution did long ago: the brain. allthingsdistributed.com/202…
Couple of updates to For Your Information (FYI), my personal knowledge graph (where you can read or subscribe your agent to my stream of attention on AI).
1️⃣ First! Thank you to everyone who has interacted, searched, or sent me feedback. Response has been great so far, and I'll take that as signal to keep going.
2️⃣ Second! AI doesn't stand still, and so that means no only to keep posting, but also improve the agent experience. FYI now scores 100% on two useful measures of agent usefulness: is-agentic, and accept markdown evals. The site is now available as a REST API with markdown formatting in all cases, and has the interfaces and responses agents need to be successful in understanding and interacting with the content and relationship graph on the site. That will mean your agents can do more with the data, more cheaply and efficienctly.
I used both evals in an optimization loop and asked Kiro to just keep making improvements until we hit the perfect scores. Took about an hour all in.
3️⃣ Third! I added a new 'Themes' section. This is an AI-curated wiki of themes and ideas, based on the recent WikiSkill paper (link is on FYI, and in description). They collect together recurring themes in FYI, becoming a compounding synthesis of the items, sources, and connections on that topic. New items are folded into these pages as they’re posted. It's pretty cool.
mattwood.fyi
Fun fact: a year ago, inspired by @OmarchyLinux, I started experimenting with my own distro (okonomi - 'as you like it'), which allowed run time alteration of any app. Just hit SUPER-TILDE, and it would jump into edit mode, accept a prompt, and update the software in place. It was... not good. Turns out, run time is actually the worst place to make changes! You have to stop, pause, change mental model, right when you are trying to get work done.
A 'build' phase is actually super valuable not just to build, but to figure out what and why you want something to work the way it does (it's why specs are such an important part of development today).
Omarchy's plugins solve the same problem in a much, much cleaner way. Love it! Congrats @dhh and team on Quattro!
Congrats, @mattsgarman!
Initially, I resisted giving agents access to my Obsidian vault. It was my space, and I didn’t want it changing under my feet. But how I use Obsidian has changed enormously over the past year.
Today, my agents have permission to read and write. Multiple agents work across the vault several times a day, keeping notes up to date, connecting dots, expanding links, finding relevant context, and generally tending to the space alongside me.
At first, I found it difficult to share what had always felt like my private thinking space. Now the opposite is true. It feels oddly lonely to write, explore, or think without that additional context and support around me. And it feels strange to imagine my agents not being up to speed on what I’m thinking, learning from it, and becoming more useful to me as a result.
I also wondered whether I would eventually abandon the shared vault as a kind of machine space and retreat to a new, isolated, private one. That hasn’t happened either. The vault still feels like mine. It just no longer feels like I’m alone in it.
That’s a pretty significant shift in mindset in a remarkably short period of time.
I have flashbacks from XML and HTML linters of yore, but this 'is agentic' score drives a lot of the right behaviors.
is-agentic.com/scan/mattwood…
Of course, @kirodotdev is looping in the background to push the score up. Low-cost validation loops ftw.
