Meta Muse Code 运行时:事件日志、崩溃恢复与 Exactly-once 缺口

Muse Code 已退出 beta 并更换主力模型,但公开运行时证据仍未关闭 exactly-once 缺口。本文用故障注入审计其事件日志与恢复合同。

Administrator Administrator Published on 2026-11-09

Muir 线性注意力:1000 万 Token 证明了什么,又没有证明什么

拆开 Muir 的长流稳定、精确召回、多跳读出与 TPU 吞吐口径,避免把四组实验合并成通用 1000 万 Token 上下文结论。

Administrator Administrator Published on 2026-11-08

Muir Linear Attention: What the 10M-Token Result Actually Proves

Muir streams 10M tokens in a fixed 25 MB state, but stability, exact recall, multi-hop readout, and TPU throughput are four different claims.

Administrator Administrator Published on 2026-11-08

安全研究 Agent 需要反事实任务表述评测

用反事实任务对区分真实阻塞、透明拒绝和静默欠表现,把安全研究 Agent 的协作质量变成发布门。

Administrator Administrator Published on 2026-11-06

Safety Research Agents Need Counterfactual Task-Framing Evals

Test whether safety framing changes agent performance with matched tasks, paired metrics, refusal transparency, and monitor-recall gates.

Administrator Administrator Published on 2026-11-06

FlashProve 审计:密码学推理证明能验证什么

FlashProve 能把被检查 token 绑定到承诺的整数模型。采购和安全验收还要分别核对模型身份、服务链、质量安全与实现等价。

Administrator Administrator Published on 2026-11-05

FlashProve Audit: What Cryptographic AI Inference Proofs Verify

FlashProve can bind checked tokens to a committed integer model. This audit separates model identity, output integrity, safety, and implementation gap

Administrator Administrator Published on 2026-11-05

Claude 用 11 天形式化费马大定理:怎样读懂这份 Lean 证明的证据等级

Claude 完成了费马大定理的 Lean 形式化。本文区分命题范围、机器检查、专家审阅与独立复现,解释每层证据究竟能证明什么。

Administrator Administrator Published on 2026-11-05

LifeVal 评测日常 Agent,但上线授权不能只看任务得分

LifeVal 已能检查产物和状态变化,真实上线还要回答 Agent 是否有权行动、后果能否撤销,以及外部结果是否真正完成。

Administrator Administrator Published on 2026-11-04

LifeVal Tests Everyday AI Agents. Deployment Still Needs a Risk and Verification Ladder

LifeVal tests personalized everyday tasks with artifacts and state diffs. Use a risk and verification ladder before an agent acts in the real world.

Administrator Administrator Published on 2026-11-04
Previous Next