科技新闻 2026-2-28
AI and systems news this week: LLMs ingest CI logs, Parakeet.cpp brings ASR to C++ with GPU speed, Hyprland 0.54 launches, and concerns grow over ChatGPT’s medical safety. Meanwhile, secure agent sandboxes, stack allocation, and a Smalltalk rewrite highlight developer innovation.
We gave terabytes of CI logs to an LLM 94
Tags:
AI AgentSQLCI/CDObservabilityClickHouseDebuggingData QueryingSource:
HackerNews| 阅读原文
[摘要]
AI agent uses dynamic SQL queries across 1.5B CI log lines in ClickHouse to rapidly trace flaky tests via contextual analysis of metadata and raw logs, enabling novel debugging beyond predefined queries.
We deserve a better streams API for JavaScript 94
Tags:
JavaScriptStreamsPerformanceWeb StandardsRuntimeSource:
HackerNews| 阅读原文
[摘要]
Web Streams标准因历史设计限制导致现代JS应用中存在性能与可用性瓶颈,新方案基于语言原语实现2-120倍性能提升。
Parakeet.cpp – Parakeet ASR inference in pure C++ with Metal GPU acceleration 94
Tags:
C++ASRApple SiliconGPU AccelerationFast InferenceAxiom Tensor LibrarySource:
HackerNews| 阅读原文
[摘要]
NVIDIA Parakeet模型实现纯C++部署,基于轻量级Axiom张量库,支持Apple Silicon GPU加速,110M模型10秒音频识别仅需27ms,性能远超PyTorch MPS。
Building secure, scalable agent sandbox infrastructure 92
Tags:
Unikraft微虚拟机沙箱安全控制平面无服务器架构容器化Source:
HackerNews| 阅读原文
[摘要]
通过控制平面架构将浏览器代理完全隔离于Unikraft微虚拟机中,实现无密钥、可弹性伸缩的安全部署,从AWS Lambda演进至全栈沙箱化。
Allocating on the Stack 92
Tags:
Go编译器优化内存管理性能提升Source:
HackerNews| 阅读原文
[摘要]
Go 1.22+优化栈分配,通过编译器自动将固定大小切片分配至栈,减少堆内存使用与GC压力,提升性能。
Working on Pharo Smalltalk: BPatterns: Rewrite Engine with Smalltalk Style 92
Tags:
Smalltalk重构工具AST语言设计编程语言Source:
HackerNews| 阅读原文
[摘要]
BPatterns 提供 Smalltalk 原生语法的模式匹配 API,简化了重构引擎的复杂语法,保留 AST 级别重写能力。
LiteLLM (YC W23): Founding Reliability Engineer – $200K-$270K and 0.5-1.0% equity 92
Tags:
AI基础设施可靠性工程Python异步性能优化生产稳定性Source:
HackerNews| 阅读原文
[摘要]
LiteLLM招聘首位可靠性工程师,负责保障全球顶级企业AI基础设施的稳定性与性能,应对高并发、内存泄漏、锁竞争等复杂生产问题。
Hyprland 0.54 Released As A "Massive" Update To This Wayland Compositor 87
Tags:
WaylandCompositorLinuxPerformance OptimizationUI/UXSource:
Phoronix| 阅读原文
[摘要] Hyprland 0.54引入了工作区/显示器级布局支持、核心化hyprscrolling插件、新monocle布局及50%-500%性能提升,尤其优化集成显卡表现。
Experts sound alarm after ChatGPT Health fails to recognise medical emergencies 87
Tags:
AI SafetyHealthcare AIClinical Decision SupportRisk AssessmentSource:
HackerNews| 阅读原文
[摘要]
ChatGPT Health failed to identify urgent medical needs in over half of tested cases, including missed suicide risk and delayed emergency care, raising serious safety concerns.
A Chinese official’s use of ChatGPT revealed an intimidation operation 85
Tags:
AI安全信息操控政治监控人工智能伦理Source:
HackerNews| 阅读原文
[摘要]
中国执法官员用ChatGPT记录跨国镇压行动,包括伪造美国移民官员身份威胁海外异议人士,暴露AI被用于系统性政治压制。