research
Jais 1 → Jais 2: What Changed in Arabic Morphology Probing?
The gain was concentrated in number, and it moved when the representation interface or lexical split moved. Less headline; more useful.
Mohammed Al-Thobaiti
I build local inference runtimes, study how language models behave, and test where evaluation pipelines lie.
Most of my work is Rust, quantized CPU inference, experiment plumbing, and Arabic morphology. I benchmark assumptions, chase failure modes through model files and kernels, and write down what broke.
01 / latest
Recent
Four useful signals from the lab, not the whole archive.
read →
release
Ember v0.6.8: the runtime learned to use tools without becoming mysterious.
Multi-call steps, approval gates, trace replay, payload digests, and HTML reports. The inference core stayed out of it.
failure mode
A Model File Is More Than Weights
Fuzzing GGUF loaders made the boundary obvious: metadata controls execution, so parsing it is a security decision.
experiment
Before Mapping KV Caches, Make Them Measurable
Verified snapshots and exact same-model replay came before the cross-model mapper. Otherwise the mapper gets blamed for everything.
02 / systems
What I’m building
Mostly instruments for asking models uncomfortable, reproducible questions.
Ember
runtimeCPU-first Rust inference over GGUF: packed Q4/Q6 execution, hidden-state interventions, reproducible bundles, multimodal paths, and auditable tool traces.
Atlas
research toolingMaps limitations, missing evaluations, and future-work claims from papers. Deterministic, evidence-first, and deliberately not a research-gap generator.
mini-os
kernelA small x86_64 kernel with paging, a heap, cooperative async execution, ramdisk, and a shell. Fewer layers; fewer places for the explanation to hide.
devdocs-v2
local RAGA local document-retrieval system for PDF collections with incremental indexing, context-aware chunks, embeddings, and local inference.
03 / research
Open questions
Three questions currently shaping the work.
11-model audit
What does an Arabic morphology probe actually measure?
How much comes from morphology, and how much comes from lexical overlap, prompt leakage, token position, or a convenient pooling choice?
current answer →causal localization
When quantization changes an answer, where did the change begin?
The broad “Arabic degrades under Q4” story did not survive the tested matrix. Single-layer patches did: rare boundary failures localized to L7 in Qwen and L1 in Llama.
experiments →runtime forensics
Does the runtime match the label on the model file?
A Q4 file can still become an f32 runtime. I care about resident representation, execution paths, and provenance, not just the filename.
autopsy →04 / field notes
Things I wrote
Technical notes after the benchmark, bug, or methodological caveat became too annoying to leave undocumented.
Jais 1 → Jais 2: What Changed in Arabic Morphology Probing?Arabic NLP / probing→
What Does a Morphology Probe Measure? Lexical Splits and Representation Interfacesevaluation / splits→
A Model File Is More Than WeightsGGUF / fuzzing→
Before Mapping KV Caches, Make Them MeasurableKV cache / artifacts→
all notes →