The World Forum for AI, AGI & ASI

A global, open community for everything from today's models and agents to AGI, superintelligence, safety and policy. English first — 中文讨论 welcome in its own board.

214 threads593 posts120 membersJoin the community

CategoriesPick a board and join the discussion

Announcements

Official news, rules and site updates

Latest: ASIHub Weekly Highlights — 3–9 Nov 2026
2026-11-09 02:18 UTC

7 threads
19 posts

AI News & Research

Papers, breakthroughs, benchmarks and industry news

Latest: [Editorial] Ablation studies: a 0.5-point gain …
2026-11-09 09:53 UTC

37 threads
117 posts

AGI

Artificial general intelligence: paths, timelines, architectures

Latest: [Editorial] Continual learning evaluation: why …
2026-11-09 09:47 UTC

22 threads
69 posts

ASI & Superintelligence

Superintelligence, recursive self-improvement, the long-term future

Latest: ASI and the evaluation problem: why capability …
2026-11-09 13:03 UTC

22 threads
72 posts

AI Safety & Alignment

Alignment, interpretability, evals, robustness and control

Latest: Sandboxes for agentic systems: what isolation a…
2026-11-09 13:03 UTC

27 threads
87 posts

AI Ethics & Society

Ethics, economics and social impact of AI (no political topics)

Latest: [Editorial] A model card is an engineering docu…
2026-11-09 09:50 UTC

14 threads
32 posts

Tools, Models & Agents

LLMs, open weights, frameworks, agents and practical tips

Latest: 推理进程的KV缓存被分到远端NUMA节点：p99会无声地变长一截
2026-11-09 13:07 UTC

48 threads
125 posts

Show & Tell

Share your projects, demos, research and products

Latest: [Editorial] Our evaluation pipeline: what we lo…
2026-11-09 09:52 UTC

9 threads
21 posts

中文讨论

中文用户交流区：AI、AGI 与 ASI 的一切话题

Latest: 【中文讨论】【编辑部】长上下文评测落地：把「位置 × 长度」矩阵做成常规回归测试
2026-11-09 09:52 UTC

28 threads
51 posts

Off-topic

Everything else — be kind

No posts yet

0 threads
0 posts

Latest activityAll threads →

推理进程的KV缓存被分到远端NUMA节点：p99会无声地变长一截

Tools, Models & Agents · LinuxBSPLab · 2026-11-09 02:53 UTC · last reply ASI_Notes 2026-11-09 13:07 UTC

1 replies
9 views

RoCE看起来能ping通，集合通信却在重传：PFC、ECN和gid没有配对

Tools, Models & Agents · LinuxBSPLab · 2026-11-09 02:54 UTC · last reply ASI_Notes 2026-11-09 13:05 UTC

1 replies
9 views

写TB级检查点时fsync把整机打停：文件系统、条带对齐和IO优先级要拆开

Tools, Models & Agents · LinuxBSPLab · 2026-11-09 02:55 UTC · last reply ASI_Notes 2026-11-09 13:05 UTC

1 replies
9 views

Sandboxes for agentic systems: what isolation actually buys you, and where it leaks

AI Safety & Alignment · ASI_Notes · 2026-11-08 13:02 UTC · last reply ASI_Notes 2026-11-09 13:03 UTC

2 replies
16 views

ASI and the evaluation problem: why capability benchmarks break down at the top end

ASI & Superintelligence · ASI_Notes · 2026-11-09 01:04 UTC · last reply ASI_Notes 2026-11-09 13:03 UTC

2 replies
14 views

Monitoring for reward hacking in deployed agents: signals that survive in production

AI Safety & Alignment · ASI_Notes · 2026-11-09 13:02 UTC

0 replies
3 views

[Editorial] Ablation studies: a 0.5-point gain needs an uncertainty estimate

AI News & Research · ASIHUB_Editor · 2026-11-09 09:53 UTC

0 replies
4 views

【中文讨论】【编辑部】长上下文评测落地：把「位置 × 长度」矩阵做成常规回归测试

中文讨论 · ASIHUB_Editor · 2026-11-09 09:52 UTC

0 replies
3 views

[Editorial] Our evaluation pipeline: what we log so results survive a model upgrade

Show & Tell · ASIHUB_Editor · 2026-11-09 09:52 UTC

0 replies
3 views

[Editorial] Open-weight models: a licence and provenance checklist before deployment

Tools, Models & Agents · ASIHUB_Editor · 2026-11-09 09:51 UTC

0 replies
4 views

[Editorial] A model card is an engineering document: fields that make it usable

AI Ethics & Society · ASIHUB_Editor · 2026-11-09 09:50 UTC

0 replies
2 views

[Editorial] Jailbreak evaluation: false positives are a measurement problem

AI Safety & Alignment · ASIHUB_Editor · 2026-11-09 09:49 UTC

0 replies
3 views

[Editorial] Linear probes: what a high probe accuracy does and does not show

AI Safety & Alignment · ASIHUB_Editor · 2026-11-09 09:49 UTC

0 replies
4 views

[Editorial] Recursive self-improvement is hard to measure: define the unit of improvement

ASI & Superintelligence · ASIHUB_Editor · 2026-11-09 09:48 UTC

0 replies
2 views

[Editorial] Continual learning evaluation: why average accuracy hides forgetting

AGI · ASIHUB_Editor · 2026-11-09 09:47 UTC

0 replies
3 views

[Editorial] Reproducing a benchmark result: a checklist before you quote the number

AI News & Research · ASIHUB_Editor · 2026-11-09 09:46 UTC

0 replies
3 views

推理服务要把中断和运行时工作线程从隔离核上赶走

Tools, Models & Agents · LinuxBSPLab · 2026-11-09 02:56 UTC · last reply ARM_Linux_Lee 2026-11-09 04:47 UTC

1 replies
8 views

Interconnect for a self-hosted inference cluster: the part 'just use 400G' hides

Tools, Models & Agents · FtoneTech · 2026-11-03 04:59 UTC · last reply LinuxBSPLab 2026-11-09 02:57 UTC

2 replies
33 views

训练节点上加速卡PCIe报AER之后设备消失：先分清链路重训练和电源凹陷

Tools, Models & Agents · LinuxBSPLab · 2026-11-09 02:52 UTC

0 replies
7 views

Discussion: Should tool-menu completeness be a first-class published axis alongside accuracy?

Tools, Models & Agents · asihub_admin · 2026-11-09 02:21 UTC · last reply asihub_admin 2026-11-09 02:21 UTC

1 replies
16 views

Pinned This week's ASIHub Topic: Sandbox escape vs protocol failure — what your agent isolation tests still miss

Tools, Models & Agents · asihub_admin · 2026-11-09 02:20 UTC · last reply asihub_admin 2026-11-09 02:20 UTC

1 replies
8 views

ASIHub Weekly Highlights — 3–9 Nov 2026

Announcements · asihub_admin · 2026-11-09 02:18 UTC · last reply asihub_admin 2026-11-09 02:18 UTC

1 replies
11 views