WEEKLY PICKS

Weekly Paper Club

The papers most likely to change a position view

FlashInfer cut token latency by as much as 69 percent

AI/CS · 2025

THE TAKEAWAYFlashInfer cuts attention latency and is already integrated into major serving frameworks, so AI capacity can rise without buying the same proportion of new hardware.

Next Papers

7 to discuss

RankWhat changedPositionsDecision relevanceAction

Moving AI memory to CPUs could unlock bigger batches without new GPUsAI/CS · NEO: Saving GPU Memory Crisis with CPU Offloading for Online LLM Inference

9988 HK · Positive 700 HK · Positive DELL UN · Positive+1

Potentially positive for cloud platforms with underused host CPUs; mixed for Dell and data-center holdings because the winning system mix depends on workload and hardware generation.

Evidence →

Mobile AI got 40 percent faster without shrinking the modelAI/CS · Efficient LLM Inference using Dynamic Input Pruning and Cache-Aware Masking

1810 HK · Positive

Thematic positive for Xiaomi's on-device AI thesis, but no shipping-device latency, battery, adoption, or monetization data is provided.

Evidence →

Humanoid hands learned three real world tasks from modest dataRobotics · Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids

TSLA UW · Mixed

Directly relevant to the technical feasibility behind Tesla's humanoid narrative, but mixed for the current Short: capability progress challenges the view while reported reliability and hardware constraints support it.

Evidence →

Long VLA kept robots on track through multi step jobsRobotics · Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation

TSLA UW · Mixed

Thematic and mixed for Tesla's Short: longer task chains challenge a pure software-impossibility case, while manual data work and failure propagation remain commercialization barriers.

Evidence →