I’ve open-sourced ORE.
ORE is a self-hosted systems language compiler targeting AArch64. The compiler itself is written in ORE - lexer, parser, type checker, IR, optimizer, register allocator, and native codegen.
It also runs inside its own small ARM64 kernel and can rebuild itself to a byte-identical compiler image.
https://t.co/qQzsaVourb
@MudideShiva Riding the AI hype train, anyone claiming to have achieved superintelligence is either clueless or doesn't really understand what they're talking about.
ORE’s 3rd loop is public: declared machine → M1 → self-rebuild → M2, with M1 == M2 byte-for-byte.
ORE now owns the machine semantics. Python is only a differential oracle.
Hundreds of private iterations behind this one public step.
https://t.co/0cVthUWciQ
@Royultea_@sama Not really these programs are too bureaucratic in shortlisting process which is a serious letdown, i tried once and also its a very crowded space.
@Royultea_@sama I've been using claude lately too, opus lasts surprisingly more compared to sol although it is not as capable.
but anyways its always fun to use these LLMs and try to solve some of the most complex problems out there!
it adds to our value function i guess😁
@Royultea_@sama Yep, I agree it sounds like a stretch when we are discussing simple workflows like designing websites and so on. but when it comes a long horizon tasks lets say in Rust, Luna is incapable, even with SOL working as an orchestrator, this is where open models come in to rescue.
@sama@Royultea_ Chatgpt plus makes no sense now, cant use SOL in codex, it literally consumes 5hr limit in 15 mins, Luna is better for probing not implementation. Atleast give us Astra in Chat with limit. Else the current capability is equivalent to what open models are providing for free.
There are still deliberate limitations: no floating point yet, and parts of the language surface are still being brought into the self-hosted frontend.
The next step is the third self-hosting loop: making the rebuilt compiler construct more of the complete bootable system itself.
That’s where I’m taking ORE next.
I’ve open-sourced ORE.
ORE is a self-hosted systems language compiler targeting AArch64. The compiler itself is written in ORE - lexer, parser, type checker, IR, optimizer, register allocator, and native codegen.
It also runs inside its own small ARM64 kernel and can rebuild itself to a byte-identical compiler image.
https://t.co/qQzsaVourb
The self-hosting loop was the milestone I wanted before publishing:
ORE compiler boots under the kernel → reads its own source through file syscalls → recompiles itself → emits byte-identical assembly → rebuilds the compiler image → boots that image and reproduces itself again.
Why this may matter:
In multi-agent systems, robotics, distributed sensors, or modular world models, simply aligning representations may not produce useful communication.
A stronger design principle is to make each module depend on information only another module has and let prediction error flow through the channel.
Synthetic result, not yet a general law. But a clear hypothesis to test.
I tried to build a minimal “becoming cell” around sphere→plane→sphere dynamics. It did not become a universal learner. What survived: a reproducible boundary study of geometry, uncertainty, communication, and clean negative results.
https://t.co/dTdNLD08qM
The most surprising result was about communication.
Two partial observers learned a useful continuous protocol without rewards or message labels but only when each directly needed the other’s missing information to predict its own future.
Agreement alone aligned their representations, but produced no useful semantics.
Meaning came from reciprocal predictive necessity not alignment.
• sha256 checksummed excerpts no excerpt, no "verified"
• drift-checker re-reads files live, flags stale evidence
• live multi-agent debate on the spine as a message bus
• emit/import handoffs - memory survives role changes
• contradictions recorded, supersede-on-fix clears them
• zero-dep Python, MCP server + CLI + git hooks
AI agents lie. EvidenceSpine makes them checkable claims carry checksummed evidence, verified means re-verifiable, stale claims get flagged. Zero-dep Python, MCP-ready, runs as a git hook. v0.5.0 is live
→ https://t.co/bmngPJaUaJ
Maybe the next step is worlds natively playable by the model. In ERTW I’m testing egocentric tensors in, continuous actions out, lockstep physics, and no reward channel. The model can causally audit a generated world by inhabiting it, not by slowly inspecting screenshots.
https://t.co/Zh1M4sjP6X
@miramurati Continuous spherical latent control with future-latent correction + hard boundary vetoes (no extrinsic reward). Exploring reward-free prospective control on the experimental branch: https://t.co/dES7J6pVQ8