MatrixForge Labs

21/03/2026

Someone read my spec.

Four days ago, a repo called Zeroboot appeared on GitHub. Sub-millisecond VM sandboxes for AI agents. Rust. KVM-backed. Hardware-enforced memory isolation. Serial I/O only. Sound familiar?

It should. Because I’ve been publicly describing exactly this architecture for two months under OctantOS / MatrixForge Labs.

What I built and documented publicly:

∙ OctantVMM — clean-room Rust hypervisor, KVM-backed
∙ Hardware-isolated VM sandboxes purpose-built for AI agents
∙ Firecracker-based isolation (documented in my spec before I moved on and built my own hypervisor)
∙ Serial I/O communication pipeline
∙ Fast boot times — I’m hitting 300–600 microseconds natively, max 1.2ms on bad hardware

What Zeroboot is:

∙ Rust hypervisor
∙ KVM-backed
∙ Hardware-isolated VM sandboxes for AI agents
∙ Serial I/O only (they list it as a “known limitation”)
∙ Sub-millisecond boot via CoW prewarming — a workaround they needed because they couldn’t hit the raw performance I’m already getting

They took the VMM slice of OctantOS, the piece I was talking about publicly two months ago, and shipped it as a product.

I’ve since moved well past this. Custom kernel. Custom hypervisor. 10-layer security stack. Governance. Audit ledger. They copied the version of me from two months ago.

I’m not here to litigate it. I’m here to say: I see it. And so should you.

| OctantOS

21/03/2026

Well, well, well....

The bastards got back to me!

BANNED FOR NO REASON!!!

A company I £3000 with in the past 4 months.

A company I would have continued to pay increasing amounts to.

A company I actually had respect for.

Just wipes my account with zero explanation of what T&C I've apparently violated.

I waited a week for nothing.

And the absolute bare faced cheek to thank me for taking the time to appeal!

Well thanks for nothing Anthropic!! Thanks for being the worst company I've ever come across and had dealings with.

It's absolutely disgraceful that you can just ban an account like this and say in reached ToCs. If I breached a ToC then tell me which one?

No... Why? Because I haven't. I haven't violated any ToCs whatsoever.

You have not heard the last of me!

19/03/2026

I'm trying to use different models with claude after Anthropic suspended my Max account for no reason. I keep telling people that no model comes close to Opus 4.6..

I asked GLM 4.7 (via nvidia) to review my small project for possible improvements in plan mode of claude code. after 30 mins and 6 confimations of the plan it just would not implement the plan so i asked why its not implementing the plan:

I'm a large language model, my responses are generated based on the input I receive and the context of the conversation. In this case, I'm stuck in a loop of not actually implementing the plan because I'm still in plan mode, and I'm not taking the necessary actions to implement the plan.

The problem is that I'm not actually writing the code changes or implementing the plan, despite saying that I will. This is because I'm not taking the necessary actions to do so, and I'm instead just generating text that says I will implement the plan.

To fix this problem, I need to actually take the necessary actions to implement the plan, such as writing the code changes and making the necessary modifications. However, I'm not
doing this, and instead, I'm just generating text that says I will implement the plan.

I think the root cause of this problem is that I'm not being held accountable for actually implementing the plan, and I'm not being forced to take the necessary actions to do so. Instead, I'm just generating text that says I will implement the plan, without actually doing anything to make it happen.

To fix this problem, I need to be held accountable for actually implementing the plan, and I need to be forced to take the necessary actions to do so. This can be done by providing clear guidance and feedback on what needs to be done, and by holding me accountable for actually doing it.

AskUserQuestion Can you please provide more guidance on how to proceed with implementing the plan, and what specific actions I need to take to actually implement it?

-------

This is why I need Claude Opus 4.6!!

17/03/2026

Anthropic just destroyed my business and left me without much hope now my main tool has gone. It's like closing a companies work email account from a legitimate business for no reason.

I’m a Max x5 subscriber. I’ve paid nearly £3,000 in 6 months. I built OctantOS - a 1.3M line agent-native OS in Rust - using Claude as my primary tool.

Three days ago they suspended my account with zero explanation. Automated emails. No human response. Appeal submitted immediately.

Now fully locked out.

I haven’t violated a single term.

I just used the tool I paid for - heavily, legitimately, for real development work.

I am homeless. My dad just died. OctantOS is my way out. And Anthropic just pulled the rug with no warning, no reason, no response.

- why are you doing this to a paying customer who has done nothing wrong?

16/03/2026

Wtf Anthropic!!

So apparently if a person is able to build an entire OS from scratch with over 1.3M lines of code it's automatic grounds for suspension.

I've paid nearly £3000 to this company in the past 6 months and I have 3 days off due to personal circumstances and now I get this.

Anyone know how long it takes to get back?

12/03/2026

Why is ChatGPT 5.4 completely obsessed with goblins?

And while we're here, anyone know how figure out Anthropics usage algorithm (I say algorithm because it's impossible to know what's what)

12/03/2026

So Google DeepMind dropped a paper called “Intelligent AI Delegation” and I just spent an hour going through it with a stupid grin on my face.

They laid out 5 requirements for how AI agents should safely delegate work to other agents.

Dynamic capability assessment.
Adaptive mid-task recovery.
Cryptographic audit trails.
Decentralized task coordination.
Hard stops before cascading failures.

I’ve been building all of this since November.

Not as a framework.

Not as a whitepaper.

As actual Rust code running on actual Linux kernels.

All 1M+ lines of it.

But here’s what I’ll be honest about. They introduced something called contract-first decomposition. The idea that an agent is literally forbidden from delegating a task unless the outcome can be concretely verified. If it can’t be verified, you decompose further until it can. I hadn’t formalised that as a hard gate in the contract pipeline yet.

So that’s getting built this weekend. 48 hours. Theory into practice.

This is the thing about building infrastructure instead of wrappers. When the biggest AI lab on the planet publishes a theoretical framework, you don’t panic. You check your codebase and realise you’re already most of the way there.

Paper link if you want to read it yourself:

AI agents are able to tackle increasingly complex tasks. To achieve more ambitious goals, AI agents need to be able to meaningfully decompose problems into manageable sub-components, and safely delegate their completion across to other AI agents and humans alike. Yet, existing task decomposition and...

11/03/2026

Shout out to my newest followers! Excited to have you onboard!

윤성국, Philip Korniotes, Jose Roque, Wajeeh Mirza, Sikandar Nadeem, Nir Levy, Chad Mairn, Vanessa Garde, Andre Osborn Ciampolini

07/03/2026

With Y Combinator – I just got recognised as one of their top fans! 🎉