Google 发布 Gemini 4 Argon 模型
Google 发布 Gemini 4 Argon,主打深度推理与网络安全防御,在 CWE-bench 与 Grok 4.7、GPT-6 Astra 并列第一,幻觉率 15% 为领先模型中最低。Argon 输入 $2/M tokens、输出 $10/M tokens,输出上限 1M tokens;目前面向 Fairwind Program 成员,将逐步开放给开发者、企业与普通用户。
The 2 most discussed AI stories in the past 48 hours
10月1日 20:05 updated · ranked by discussion heat
Google 发布 Gemini 4 Argon,主打深度推理与网络安全防御,在 CWE-bench 与 Grok 4.7、GPT-6 Astra 并列第一,幻觉率 15% 为领先模型中最低。Argon 输入 $2/M tokens、输出 $10/M tokens,输出上限 1M tokens;目前面向 Fairwind Program 成员,将逐步开放给开发者、企业与普通用户。
论文提出开源两阶段技能检索器 SkillSeek,基于标准 IR 配方(BGE-base 编码器加小交叉编码器)并通过 MCP 暴露。在 89 任务 SkillsBench 上,SkillSeek 在四个设置中的三个达到或超过 LLM 循环的通过率,剩余一个由交叉编码器弥补;每轮成本从 51.30 美元降至 27.54 美元,接近无技能基线。
Heat comes from independent accounts and organisations discussing the same event. Repeated collection counts once, with a 24-hour half-life. It measures discussion activity, not reporting quality.
The ranking covers the past 48 hours. Trends compare the same sources over time and describe our monitoring range, not the whole internet. A trend line is hidden when there is not enough comparable history.
Sources list publicly readable reports. Participants also include accounts and organisations counted only for heat. Multiple channels from one organisation may be combined, so participant and source counts can differ. Open an event to read the reports and viewpoints.