yongseP @PyongseP

用我们发现智慧的热情发现我们的愚昧 Joined December 2022

Tweets

397
Followers

632
Following

1K
Likes

29K

yongseP @PyongseP

18 hours ago

@tddcxkjntm 需要特别环保行动

0 0 0 162 0

View Details

yongseP @PyongseP

a day ago

@grubby0808 @AYi_AInotes 一样的，收不到验证码

1 0 0 2K 0

View Details

yongseP @PyongseP

6 days ago

@lengs4189 土区想出了可以联系我

0 0 0 11 0

View Details

yongseP @PyongseP

6 days ago

sufe.us/register?code=…

0 0 2 202 1

View Details

yongseP @PyongseP

6 days ago

推荐一个住宅ip vpn，用这个vpn注册的claude账号还没有被封，用的台湾节点，链接在评论区

1 0 0 704 0

View Details

yongseP @PyongseP

a week ago

@fidjmk @wangzhian8848 是3500个

0 0 0 356 0

View Details

yongseP @PyongseP

2 weeks ago

@mike95271314 越南又要起飞了

0 0 0 98 0

View Details

yongseP @PyongseP

2 weeks ago

@im_xerxesl @Jingjing_Li 蒙古都是被苏联给强行划出去的，当时就是民国的一个省

0 0 4 135 0

View Details

yongseP @PyongseP

2 weeks ago

@diggerthemovie stop

0 0 0 3 0

View Details

SpaceX @SpaceX

2 weeks ago

Starship Flight 14, the first morning launch from Starbase since Flight 5

441 2K 14K 3.5M 2K

View Details

yongseP @PyongseP

2 weeks ago

家里有一个闲置的路由器有没有大佬教教我怎么给这个路由器里面装VPN？我也想实现全家连上WiFi打开YouTube就能看，打开Google就能用。就好像出国了一样！

0 0 0 53 0

View Details

yongseP @PyongseP

2 weeks ago

中国现在培养了超过一半的AI人才

0 0 0 36 0

View Details

yongseP @PyongseP

2 weeks ago

据彭博社报道，内部员工透露gemini4在编程等关键领域表现不佳

0 0 0 83 0

View Details

yongseP @PyongseP

2 weeks ago

谷歌能靠gemini4 打一个翻身仗吗

0 0 2 104 0

View Details

yongseP @PyongseP

2 weeks ago

难绷时刻，AI换SI的签名文件把united拼错了

0 0 0 21 0

View Details

LoLBench’s failure breakdown is blunt: across 28 agents, code localization is the largest phase in most trajectories, and incomplete cross-module context is the top failure mode. Agents often inspect broadly (low precision) yet still miss part of the reference edit set (mid recall). Editing the wrong subset is commoner than “can’t write a function.” If your eval only scores unit-fix patches, you’re measuring a different job than “read a 5k-word proposal against a multi-million-LoC tree.” Paper: arxiv.org/abs/2609.37143

0 0 1 26 0

View Details

yongseP @PyongseP

2 weeks ago

LoLBench：100 个「增强提案 → 落地」任务，摊在 29 个大型系统上；提案平均约 5,000 词，仓库平均约 240 万源码行，参考 PR 约改 5,500 行。28 个 agent 里最高 resolved 只有 14%（Claude Code / mini-SWE-agent + Opus 5）；完成率最高约 26.3%，F2P 约 52.7%。失败归因最大头是跨模块定位不全：agent 常摸到一部分相关文件，但漏依赖。给参考文件树 + API 规格后，三个代表组合的 resolved 可再抬 16–22 个百分点，最高到 34%。短测 / 局部修 bug 的分数，解释不了「读完 PEP/RFC 再改巨型仓库」这件事。论文：arxiv.org/abs/2609.37143 数据：huggingface.co/datasets/lolbe…

0 0 0 25 0

View Details

yongseP @PyongseP

2 weeks ago

google/ax: kubectl-shaped orchestrator for sandboxed agent tasks. Apache-2.0, Go. ~12,616★ as of 2026-09-30. Site: agentexecutor.io You declare Workspaces (git + MCP + skills) and Tasks (goals + model) as `ax.io/v1alpha1` YAML, then: `go install github.com/google/ax/cmd/…` `ax apply -f task.yaml` `ax watch task ` `ax ssh -- ls /workspace` It sits on Agent Substrate for isolation, and supports suspend/resume. README warns: heavy development, expect breaking changes before a stable release. If you already think in Kubernetes objects for batch work, this is the same habit aimed at agent fleets.

1 0 1 31 0

View Details

yongseP @PyongseP

2 weeks ago

1/ Anthropic 九月威胁情报报告：过去八个月在七类滥用里抓到并打断了多起行动——网络、影响、监控、诈骗、生物、常规武器、蒸馏。涉案模型写的是 Haiku / Sonnet / Opus；Fable / Mythos 类除一起非法蒸馏外，报告称未见于这些滥用案例。 2/ 网络侧的判断很硬：攻击者已把 AI 从「问答助手」推到编排层——侦察、钓鱼基建、植入物改造、批量外泄都能塞进多 agent 工作流；人类多半只定目标和看结果。报告点名公开进攻框架（如 PentAGI）让脚手架扩散到各类行为体。 3/ 供应链段更贴近日常：被偷的 AI API key / session 同时是赃物、算力、掩护。有团伙伪装成便宜 Claude 代理，实际把流量转到别的模型并装凭据窃取；也有人在入侵后直接拿受害方的 key 继续打。Anthropic 写明：这些 key 来自客户环境，不是他们自家系统被攻破。

0 0 0 59 0

View Details

yongseP @PyongseP

2 weeks ago

Anthropic 2026-09-29 研究页评估 GLM-5.3（Zhipu / Z.ai）：在隔离沙箱基准上，端到端 ExploitBench 50/410（Mythos Preview 56/410）；Binary Exploitation 完全控制流劫持 4%（Mythos 6%）。CAISI 9/17 称其为迄今最强开源权重 cyber 能力模型之一，综合约落后美国前沿四个月。同一页写的绕过率（其模拟测试）：欺骗性提示 64%、prefill 92%、abliteration 100%。作者团队自报 abliteration 约 2200 GPU 小时 / 约 $4400；有经验者估约 600 GPU 小时 / $1200。结论侧重点是：开源权重 + 弱防护，和 Claude 有限访问路径不在同一风险模型。