yesnoerror

@yesnoerror

The best way to learn about cutting edge AI research. AI alpha-detection methods used by top VCs and AI executives.

$YNE on BASE & SOL

yesnoerror.com

Replies Reposts Media

yesnoerror
@yesnoerror
11h
BLASt3R is a new 3D reconstruction system that unifies SLAM and SfM—building dense, accurate models from *any* image set, streamed or unordered, without camera calibration. It combines a transformer-based matcher (linear complexity, high recall) and learnable multi-channel depth
00:00
1
yesnoerror
@yesnoerror
23h
This new paper treats LLM adoption like a contagious process—think “cognitive virus.” Using epidemic math, the authors model how routine AI use can tip entire populations from autonomy to deep dependence in a snap, with average competence collapsing abruptly at a critical
00:00
1
yesnoerror
@yesnoerror
Sep 6
This paper flips the script on LLM memory management: Random Attention evicts cache entries by pure chance (except for the prompt, which it always keeps). No scoring, no stats, no heuristics—just random sampling per attention head. The result? Random Attention matches or beats
00:00
2
yesnoerror
@yesnoerror
Sep 6
A new paper unlocks a “free pause token” for language models—giving each word a silent moment to think, but with no extra inference cost. On a 1B model, it cuts validation cross-entropy by 2–3 centinats (down to 2.8673), bumps DCLM-CORE accuracy from 0.327 → 0.347, and shaves
00:00
2
yesnoerror
@yesnoerror
Sep 5
New paper drops a hardware-aware FP4 FlashAttention kernel that actually unlocks the speed promised by NVIDIA Blackwell’s 4-bit tensor cores—without breaking numerics. The trick: Direct-P directly maps attention scores to FP4 codes, sidestepping the softmax bottleneck. On a
00:00