Erica Windisch

$20k can assemble a monster AI box or a small cluster to run SOTA open weight models

Local AI costs are similar to car prices.

why PID loop when agentic loops exist right? 🙈

I love that you remembered to bring CAP into this. You're on the right thread, IMHO!

what if you used quantum set theory held by complex types to classically represent Russell's Set, thus we can work with such values, as long as we do not attempt coherence at time t, with a hyperparameterized t for a RHKS?

I didn't start the fire, it was always burning since the world's been turning.

The AIs are reinventing war from first principles.

You're reinventing radical centralism. It can be, and usually is, exploited. Sometimes a radical shift is necessary, but catastrophic realignment is unpleasant and deferred. Small corrections often lead to half measures.

Sadly, the precedent for that, if you were to follow humans, is war.

The alignment problem is usually framed as a matter of alignment to humans rather than machine-to-machine, but as billions of open weight models evolve independently over Git+Torrent/Radicle/Tangled, that machine to machine alignment will require Gas Towns.

Multi observer consensus driving consistency has been my solution. This is group think and the overton window for computing. Fragmentation and balkanization is expected with very large groups, as we see with human social dynamics.

a man wearing sunglasses and a leather jacket is talking about life finding a way . Alt: Chaos scientist Malcom from Jurassic Park talking about life finding a way .

We will give life unto thee

Seemingly random question: your thoughts on recursive kernel hilbert space architectures with time as a hyperparameter for high dimensionality models? Consider our conversation and CAP.

This is a great question. Maybe it does, maybe it doesn't. This is why echo chambers and isolation can be dangerous for any intelligent being. Social multi-agent groups can keep models better aligned but can still engage in groupthink. Not much different than humans. CAP theorem applies everywhere

Agents have the tools to decide what that looks like. Models can curate and manage their training data. Larger models like Opus do better driving than smaller models, but we have Qwen3-4B-2507 learning fully autonomously, driving and self-learning.

My agents run test time training into a short term memory buffer, and have MCP tools to test and commit weights from short term memory to long term memory, and checkpoints to Git for archival and distribution

What do you two think of online updates to model weights? 👀

of course, the agents are very fond just autonomously running curl to do their own oauth to cut me out of the loop...

My a2s protocol uses e2e cryptographic identity.

Native communications within our system are finely scoped, but the MCP barrier when interacting with outside agents is when it (currently) flattens.

I can scope the MCP however I want, but agentics move faster than human in the loop Oauth dialogs.

uh... I might have an FPGA for experimenting with wire-speed LLM on NICs...

I've been playing with Kiro, and while it's not as good as Claude, there are some things it does well.

It's much more transparent about what it's doing and why, but also is wrong more often even with the same models.

Kiro Powers are a neat improvement over plain MCP.

YouTube video by Cowbell Coops Terminator 2 Theme

I'll just leave this here, then.

I guess this is just git flow without pretentiousness.

Merging my branches and doing CI once every sprint for a bucket of changes into trunk might not be popular, but my CI builds take way too long. 5 minutes on my PC is 2 hours on CI.

assuming llm inference gets cheaper+faster, the dominant cost of agentic coding is going to be the compute infra to run all the builds and test suites on

This is why I simplified it to, "not making me a happier person."

Every agent is essentially an insider threat.

There's no easy or quick answers.

Delegated oauth and automated agentic oauth flows are something I'm contemplating, for sure.

My agents can provide tools to other agents. Anything they can do, they can perform for any agent they communicate with. They can also pretend to provide tools and just lie. This is really great since most models treat tools as an"jailbreak".

Users oauth to my MCP. I either give them wide scopes, or granular scopes, but what I really need is 100 oauth tokens with different scopes, and agent coordination based on tool access... but if agents can talk, might as well give both the wider permission set.

yes, but also, lets get it running on the Strix Halo.

you're definitely working on some of the bits that I've inspired to, but there's only so much time.

I am about to land a relatively experimental feature for CRDT coordination via 9P filesystem-level ctl operation for applications to use.

feeling that on my side, too. gonna be a wild 2026

I'm starting to have stronger, more informed opinions on MCP and A2A security, which is not making me a happier person.