GLaDOS Wiki
Every subsystem, how it works, and where it turned out to be more interesting than expected. The habit in this project is to write down the measurement that overturned an assumption alongside the conclusion it produced, which means a good number of these pages are accounts of being wrong for a while. Those tend to be the useful ones. A transformer with its position encoding wired the wrong way round still writes fluent English, and nothing in the system objects.
The project
- GLaDOS OSWhat the system is, and how it is put together
- Aperture ScienceThe homage, and which half is fiction and which is kernel
- TempleOSWhat this project inherited, and where it diverges
The machine
- The kernelUEFI as the OS, ring 0, Rust, the other cores, and the firmware's own bytecode
- Devices and storagexHCI, NVMe, Merkle trees, and a write gate that names a range
- Networking and TLSARP to TCP, hand-written crypto, and the wireless half that is not finished
- The desktopComposed and diffed, and the measurement that overturned the plan
The mind
The language
- Aiksi, the system languageRecords as values, and a gate an import cannot escalate
- The compilerx86-64 from a heap page, and the canary that must fail
Theory, and keeping it honest
- Gödel machinesWhat replaces the proof when no prover exists
- Recursive self-improvementFour judges, a ledger, and a budget on the held-out set
- AIXIMonte-Carlo planning over the machine's own fitted dynamics
Building and checking
- How the ISO is builtEl Torito, FAT32, and a reader to check the writer
- Testing without a test runnerFour instruments, and a canary that must fail
