Jun 22, 2025

• Debunking NIST's calculation of the Kyber-512 security level (2023)
  NIST's calculation of Kyber-512's security level is fundamentally flawed, as it mistakenly multiplied costs that should have been added, leading to a gross overestimation of its security against AES-128.

• AbsenceBench: Language Models Can't Tell What's Missing
  AbsenceBench reveals that while large language models (LLMs) excel at recalling information, they struggle significantly with identifying missing elements, achieving only a 69.6% F1-score on tasks designed to test this ability.

• TPU Deep Dive
  TPUs excel in scalability due to their co-designed hardware and software, notably the XLA compiler, which optimizes performance by minimizing memory access and maximizing throughput.

• Scaling our observability platform by embracing wide events and replacing OTel
  LogHouse has scaled from 19 PiB to over 100 PB, managing nearly 500 trillion rows of uncompressed logs, while achieving a 20x increase in event volume with less than 10% of the CPU previously required, thanks to the development of a specialized tool called SysEx.

• AllTracker: Efficient Dense Point Tracking at High Resolution
  AllTracker introduces a novel model for long-range point tracking, achieving high-resolution dense correspondence fields by estimating flow between a query frame and multiple subsequent frames, rather than just the next one.

• Load Test GlassFlow for ClickHouse: Real-Time Dedup at Scale
  GlassFlow demonstrated robust performance by processing over 9,000 records per second with sub-0.12ms latency, effectively handling deduplication for 20 million records without crashes or message loss.

• [D] RL/GRPO for lossless compression of text passages into 'least token representation', then using this emergent 'language' as the basis for reasoning instead of english
  This entry discusses a novel approach for compressing text.