Engineer @ https://teapawt.net/t.co/xHng0HjvU5
Melbourne
Joined May 2009
- Tweets1.8K
- Following1.3K
- Followers1.2K
- Likes29K
Pinned Tweet
Introducing Steadmesh: define your AI team in Terraform.
Give agents roles, tools and shared memory. Mix Claude Code, Codex and Pi, and run them on your own Kubernetes cluster.
Open source. This thing is sweet.
steadmesh.com
Sean 🔨 retweeted
Frontier performance at a fraction of the compute.
Today at @i14labs we’re launching Sugar Glider, our new family of LLM architectures!
Join the waitlist for early access: i14.ai
Sugar Glider delivers a new Pareto frontier: extreme memory and compute efficiency whilst maintaining frontier performance, even for large models (tested up to ~1T params).
O(1) per token inference → ✅
Fixed size KV cache → ✅
Frontier long-context retrieval quality → ✅
1 billion token long context retrieval → ☑️**
(**100% -> single needle @ 1B-token context, 90.5% multikey 1 @ 1B-token context.)
How did we do it? Turns out a few decades in the trenches of human memory and language model research is a pretty handy starting point to solve the “intelligence vs efficiency vs memory” conundrum.
... and just for something extra, we’ve been converting some of the most powerful and popular open-weight models to O(1) inference and fixed memory, with each conversion taking just hours :)
Tune in via our waitlist for early model access i14.ai
More results dropping soon!
Had an idea for an RL bench around weight space programming: get a model to program the weights of a neural net that controls humanoid physics.
Like take the huggingface duck and make it do a backflip. No PPO, just writing the weights.
2
90
Just read the proton CTO AI manifesto.
Seems pretty spot on. Your job is to be a puppetmaster.
1
4
103
Managing agents is like being a puppetmaster.
Figuring out how to be a good puppetmaster, and working with other puppetmasters is the future.
3
82
Talking with Astra about a math topic, built up shared terminology, then asked one question in a related area with different vocab. Came back and it had completely lost the original terms.
One message was enough to derail it. Feels like an architectural decision
1
37
You can put Claude Code, Codex and Pi in the same Steadmesh team.
Choose the harness for each agent, give them tools and shared memory, and run the team on your own Kubernetes cluster.
github.com/darcys22/steadmes…
4
54
Associate memory systems make a lot of sense to explain why neural nets work. Is there any other models that make sense? Maybe data compression? But is that the same thing anyway?
2
36
A simple place to start with Steadmesh: a coding agent and a reviewer, with different instructions and a shared project.
Give them a small task and watch how they work together.
Here’s an example repo to get a team running:
github.com/darcys22/steadmes…
1
1
28
Its all going to be a tool/connection problem in the end. Its great that agents can manage my whole life. But woolworths hates that a bot is going to order food for me and will block it
40
Terraform for AI teams feels surprisingly natural.
Define the agents, their roles and how they’re connected. Keep it in git. Review the plan when you change the team.
That’s what I’ve been building with Steadmesh.
steadmesh.com
1
32
The repo and a working example to get started 👇
Define the team in Terraform, give each agent its tools and instructions, then start talking to them.
github.com/darcys22/steadmes…
github.com/darcys22/steadmes…
29
The value that mathematics provides to the world is downstream from their papers.
The builders who reviewed the papers and built bridges dont care if the math dump is through a peer reviewed institution. They just care that their bridge is better
48