Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-09-16 · source-backed
Measuring symbol distribution across 29 ternary models found zeros account for up to 51.5% of weights, invalidating the equiprobable assumption behind the log2(3) reference and the 1.625 bits/weight that five-trit packing reaches in practice. The BITCOS layout (dense presence bitmap plus compacted sign vector) costs 2 minus z bits per weight at zero density z, storing weights more compactly than five-trit packing in 26 of 29 models and reaching 1.485 bits per weight on the sparsest.
Each link below shares sources, entities, or timing with this story.
This comparison ran the baseline the retrofitted-linear-attention literature skipped. Across multiple LLMs and downstream tasks SWA with sinks matches or beats post-trained linear attention, and on Needle-in-a-Haystack and BABILong it scores 2 to 10 times higher. The recommend...
Submitted August 30, the claim is empirical across small networks trained on list manipulation and across LLMs on arithmetic, logic, code and natural language (arXiv 2608.29530). It's a direct attack on the assumption that continuous vector representations and symbolic account...
Thirteen authors ran a generational genetic algorithm over specialized agents that separately handle mechanistic argument, assumption reconsideration, and evidence and testability assessment (arXiv 2609.15938). Evaluated against DepMap and Open Targets across 34 cancer types,...
First benchmark of off-the-shelf LLMs against expert-derived ground truth built on INCOSE criteria, ten models across two families and five generations each, one hundred independent runs, two requirement sets, five temperatures. The error profile is asymmetric, and performance...
The first systematic measurement of PyPI import cost covers the 500 most-downloaded packages sampled quarterly over five years, under CPython 3.9 through 3.14, on Apple M5/macOS and Intel Xeon/Linux. Half of packages import in under 6 ms but p99 is 354 ms. First import after i...
Measuring 8,234 revisions over nine years, with suppression detected semantically and validated against blinded hand labelling at 0.828 precision and 0.911 recall, exclusions were added 1,642 times and withdrawn 304 (arXiv 2608.31062). Per individual rule the ratio climbs to 1...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.