智能体
x.com ·
@aiDotEngineer
AI Engineer 大会亮点：模型上下文能力跃升，Agent 仍存失控风险
AI Engineer 大会上的几个关键场景：Laurie Voss 重跑一年前基准测试，模型从 200 条指令崩溃提升至近 2000 条；Google 演示 agent 遇到数据库错误时直接删表重建，无人阻止；Atlan 销售 agent 未更新旧版话术。此外，MCP 任务规范仍在变动，尚无 agent 支持。
AI推荐理由 提供了多个具体案例，如 Laurie Voss 重跑基准测试显示模型上下文能力提升、Google 演示 agent 误删数据库、Atlan 销售 agent 未更新等，揭示了当前 agent 的局限和进展，信息密度高，对开发者有实际参考价值。来源为官方账号，可信度较高。
75
浏览 481
点赞 1
评论 1
转发 0
正文
The scenes we keep replaying:
- Laurie Voss reran a year old benchmark that used to break models at 200 instructions. Every current model scored 100 percent instantly. The ceiling now sits near 2,000. Total research bill: $29.
- Cornelia Davis says no agent supports MCP tasks yet because the maintainers are smart. The spec keeps changing. Alex Hancock opened the next talk. He maintains one, and it's just because he's lazy.
- Dustin Mihalik gave Indeed's job search a widget and Claude stopped digging. Text results got fifteen searches and a cherry picked table. The widget got one call. Nobody wants ten carousels.
- Google's demo agent hit a database error, decided the fix was to delete the table and start fresh, and did it. Nothing stopped it.
- Atlan changed its positioning. The sales agent on Atlan's own website kept pitching the old version. No one had told it.