Hy4 추론 최적화, GLM-5.3 안전성 검증, 대규모 벡터 검색과 zHBM

One more thing about Hy4 preview that is particularly interesting. In AI research. Hy4 preview found inference bottlenecks on its own and lifted e2e throughput 31.8% via operator fusion and comms opts — stable across context lengths and concurrency. MORE ON: https:// hy.tencent.ai

Tencent just compressed Hy4-preview from 1.5TB down to 200GB and keep SWE-Bench Multi at 81.3 A 770B frontier MoE, quantized down to 1.3–2.1 bits per layer, and the scores almost hold. - MCP Atlas 83.7 to 83.2 - SWE-Bench Multi 82.9 to 81.3 - MRCR 81.3 to 81.1 A strong example of how adaptive low-bit inference is making large open models more deployable

Z .ai's GLM-5.3 just hit 84.5% on the CyberGym vulnerability benchmark, beating top proprietary models, a huge gain over the performance of its predecessor GLM-5.2. The kicker? http:// Z.ai’s AI engineers did it purely through fine-tuning and optimization of the model’s agentic capabilities, without changing the base model. The model grew so capable at finding and targeting potential exploits that http:// Z.ai held back the open weights for safety testing. Read the full analysis in The Batch: https:// hubs.la/Q04w3GkF0 #DeepLearningAI #Cybersecurity #LLMs

1/ Production vector search: billions of vectors, thousands of RPS, tight tail latency. Public benchmarks don't come close. So we built what was missing and gave it away. Write-up: https:// qdrant.tech/blog/qdrant-fi neweb-10b-release/ … Dataset on @huggingface : https:// huggingface.co/datasets/Qdran t/FineWeb-10B … Supernova Code: http:// github.com/qdrant-labs/su pernova …

7/ Compute from @Vultr . Distribution with @skypilot_org . Corpus and hosting from @huggingface . The era of 1M-vector datasets and synthetic approximations is over. Real data, real ground truth, open tooling. All links in the first post.

Md Ismail Šojal

@0x0SojalSec

3일 전

Samsung just flipped the AI memory game. Samsung unveiled zHBM stacks high-bandwidth memory directly on top of the GPU instead of next to it. The result they’re targeting: - Up to 8× performance vs HBM4E - 3× performance per watt - Heat problem Thermal resistance cut 75-90% Shorter distance to faster data, less power, less heat. This is how they plan to break the current HBM bottleneck.

seems like it

Yun-Ta Tsai

@yunta_tsai

3일 전

Real-world safety report from real-world users.

“The Illusion of What If: Evaluating the Breakdown of Counterfactual Reasoning in LLMs” LLMs can give convincing counterfactual answers, but that doesn’t mean the causal chain behind them is coherent. This paper introduces WhatIfBench and PRISM, which turn open-ended what-if

Learning Human Health and Diseases from 24-hour Wrist Movement "Here, we present Sensori, a self-supervised foundation model that learns general-purpose health representations directly from 24 hours of raw tri-axial wrist movement." Dataset: 122,640 participants contributing

MR-JEPA: A General Purpose Video Foundation Model for Cardiac MRI "We present MR-JEPA, a self-supervised video foundation model for Cardiac MRI that extends LeJEPA to 3D spatiotemporal inputs through tubelet tokenization, spatiotemporal masking augmentation, and initialization