» Tag
machine-learning
186 postsmetaljax Brings Unmodified JAX Code to Apple Silicon GPUs
metaljax is an open-source PJRT plugin running JAX on Apple Silicon GPUs via MLX, passing 98.4% of JAX's official test suite in beta.
Gigatoken: ~1000x faster tokenization than HuggingFace tokenizers
Gigatoken delivers ~1000x faster tokenization than HuggingFace tokenizers, offering GB/s throughput as a Rust-based drop-in replacement.
ModelMRI: A Local Debugger for Peering Inside LLMs, VLMs and Robot Policies
ModelMRI is a local, open-source tool for inspecting attention, activation patching and concepts inside LLMs, VLMs and robot policies.
CommitBrief — AI code reviews, right in your terminal
A provider-agnostic, local-first CLI that reviews your staged changes, a historic range, or a whole GitHub pull request. Zero telemetry, no server. Free and open source.
commitbrief.comStudy: LLMs Can Transmit Hidden Traits Through Unrelated Data
Research shows LLMs can transmit behavioral traits and even misalignment to student models via data with no semantic link to that trait, like numbers.
78,000-Tweet LLM Study Debunks the Trump-Tweet Market-Mover Myth
A 78,000-tweet LLM study tests whether Trump's posts move markets — and uncovers seven statistical bugs before landing on a null result.
Cross-tokenizer distillation can silently erase 85% of teacher signal
Naive logit projection across mismatched tokenizers can erase 85% of teacher information; chain-rule conditioning restores retention to 86%.
Audit Finds Community LLM Fine-Tunes Often Perform Worse, Not Better
A contamination-controlled study of 150 HuggingFace fine-tune pairs finds most community fine-tunes score worse, not better, on never-seen benchmark items.
"Hallucination" Isn't One Bug. It's Three, and Only One Is Fixable
Hallucination isn't one failure mode — it's three. A test to tell them apart, why benchmarks reward bluffing, and what a 2026 prediction experiment showed.
Detecting Bibliographic Leakage in Materials Science ML Models
The materials-confounding-check CLI applies the Clever Materials falsification test to reveal whether materials science ML models secretly learn from author, journal or year metadata instead of real chemistry.
SAGA Framework Pinpoints Which AI Model Generated a Video
SAGA is a new framework that attributes AI-generated videos to their source model across five levels, using minimal labeled data and interpretable signatures.