empac

[empac @empac2058569834]

@OpenAI 红队测试条件有利于攻击者，智能体拿到了基础设施源码、配置和文档；多数隐蔽信道速度慢，也难以利用。

https://t.co/9rsvxj1d4S

[empac @empac2058569834]

OpenAI用红队智能体测试训练环境，我关注的是隔离边界。

加固前沿模型训练环境

红队智能体在同一宿主机的独立microVM中，通过CPU缓存计时交换合成秘密：无需网络，也无需共享可写文件。

来自 @OpenAI。

#AI #AISafety

[empac @empac2058569834]

@OpenAI The red-team setup favored attackers with infrastructure source, configs and docs; most covert channels are slow and hard to exploit.

https://t.co/9rsvxj1d4S

[empac @empac2058569834]

OpenAI red-teamed its training setup. Here's my takeaway:

Securing our frontier model training environment

Agents in separate microVMs on one host exchanged synthetic secrets via CPU cache timing: no network or shared writable files.

From @OpenAI.

#AI #AISafety

[empac @empac2058569834]

@OhioState @Princeton 综合分只差约 0.01 时，我不会把这样的接近成绩当成排名。
https://t.co/vllru5YiyP
https://t.co/6dYsazvv0W

[empac @empac2058569834]

俄亥俄州立大学、普林斯顿大学等团队（@OhioState @Princeton）做了个 ENSO 科学考试，我很感兴趣。

SciExam for ENSO

十二个最终模型中，六个综合分超过了已发表的参考综合分 0.378。

#AI4Science #ClimateScience

[empac @empac2058569834]

@OhioState @Princeton I wouldn’t read near-tied composite scores differing by about 0.01 as a ranking.
https://t.co/vllru5YiyP
https://t.co/6dYsazvv0W

[empac @empac2058569834]

Ohio State and Princeton (@OhioState @Princeton) built an ENSO exam I’m watching.

SciExam for ENSO

Six of twelve final models exceed the published reference composite of 0.378.

#AI4Science #ClimateScience

[empac @empac2058569834]

我的 1.0.4 测试版尚未真机实测。指标来自 AI 核过的历史测试标签，并非这版 APK 的手机实测结果。
https://t.co/6LtRnF81tn

[empac @empac2058569834]

我做的镜库是用一句话搜索本地视频对白的 Android 应用，1.0.4 测试版已发布。

Top-3 约 76%（62 条 AI 核过的 query）。

换用 270M 模型，APK 约小 320 MB，Top-3 决策不变；全程本地离线。

#Android #OnDeviceAI

[empac @empac2058569834]

My 1.0.4 Pre-release is not yet tested on a physical phone. The metric is AI-audited historical labels, not this APK's phone benchmark.
https://t.co/6LtRnF81tn

[empac @empac2058569834]

I've released Jingku 1.0.4 Pre-release, my Android app for finding dialogue in local videos.

Top-3 ≈ 76% on 62 AI-verified queries.

Search with a sentence, fully local offline.

#Android #OnDeviceAI

[empac @empac2058569834]

@newscientist 成员身份为自行申报，承诺是原则而非强制执行的规则，人数每周更新。

https://t.co/ZP3kpYjIb0

[empac @empac2058569834]

已有1,000多名数学家加入一个团体，承诺不为商业AI公司工作，也不发表AI生成的数学文本。

人类数学协会 (AHM)

协会有1,008名成员，其中654人加入了自愿参加的无AI小组，不在研究中使用AI：寻找证明、探索例子或提出问题。

此事由 @newscientist 报道。

#AI #Mathematics

[empac @empac2058569834]

@newscientist Membership is self-reported, the pledges are principles rather than enforced rules, and the count is updated weekly.

https://t.co/ZP3kpYjIb0

[empac @empac2058569834]

Over 1,000 mathematicians pledged not to work for commercial AI firms or publish AI-generated math.

Association for Human Mathematics

Of its 1,008 members, 654 are in an optional AI-free caucus that doesn't use AI in research.

As reported by @newscientist.

#AI #Mathematics

[empac @empac2058569834]

@blocks @jack Block尚未披露模型规模、基准测试或发布日期，片段是公司演示，未经独立测试。

https://t.co/kU2FSjUQ7f

[empac @empac2058569834]

Block的小型AI研究实验室发布全双工语音模型，我想试试。

Dessa

仅用人类对话音频从零训练，不用文本或预训练模型，只用一台NVIDIA GB200 NVL72。

该实验室隶属 @blocks，直接向 @jack 汇报。

#AI #SpeechAI

[empac @empac2058569834]

@jack @blocks Block has not disclosed model size, benchmarks or a release date; the clips are company demos, not independent tests.

https://t.co/kU2FSjUQ7f

[empac @empac2058569834]

Block's small AI research lab announced a full-duplex speech model I'm keen to try.

Dessa

Trained from scratch only on human conversation audio (no text, no pretrained model), using just one NVIDIA GB200 NVL72.

The lab reports directly to @jack at @blocks.

#AI #SpeechAI