尔贝大冒险

专注科普类AI视频剪辑

[尔贝大冒险 @EdwinCh43136084]

这是一个完全用grok完成的视频，包括：
做旧地图动画；
地图城池的建模；
莫高窟简模；
视频片段生成（昨天更新的grok Imagine Video 1.5）；
真实壁画资料的搜集；
以及全视频的剪辑；
总共花了一天时间，现在用AI完成一个科普类视频已经是完全可行了😆。
另外grok的效率非常非常高。
我之前一直在用AI做地图动画类视频，效果都非常不错。
这是河西走廊系列的第8期（前面还有7期已完成的河西走廊科普视频）。
如果感兴趣我会陆续分享到这里。

[尔贝大冒险 @EdwinCh43136084]

河西走廊第7期出来了，河西四郡，缺一个都到不了敦煌。
https://t.co/PhMU8jCqPZ

[尔贝大冒险 @EdwinCh43136084]

@BinPeng22 榜单归榜单。cursor里4.6长任务还是会做到一半停，要盯着。

[尔贝大冒险 @EdwinCh43136084]

@0xarshia @cursor_ai I'm on the $100/mo heavy for 3 months. tokens left over every day. can't sleep. 4.6 in cursor still slower than 4.5 even on fast.

[尔贝大冒险 @EdwinCh43136084]

@chongdashu I use Imagine for Hexi Corridor history videos. the image-to-clip part is the useful bit. 4.6 in cursor still stalls halfway on long tasks.

[尔贝大冒险 @EdwinCh43136084]

@padwalpankaj9 @DeryaTR_ yeah composer feels snappier for me too. 4.6 on fast still stalls mid-task in cursor, i have to sit there and poke it.

[尔贝大冒险 @EdwinCh43136084]

@klarkxy01 @CorrectRoad @SuzunamiRei 没错，我已经很焦虑了，因为，无论怎么蹬都用不完

[尔贝大冒险 @EdwinCh43136084]

@hqmank @elonmusk I'm on the $100/mo heavy for 3 months. tokens left over every day. can't sleep.

[尔贝大冒险 @EdwinCh43136084]

@AYi_AInotes 我现在订的就是3个月每月100的heavy，token用不完，每天睡不着觉。

[尔贝大冒险 @EdwinCh43136084]

@tiany221 4.6当早期体现的话，cursor里还是比4.5慢，fast也是，长任务做到一半会停。

[尔贝大冒险 @EdwinCh43136084]

@tetsuoai I use Imagine for Hexi Corridor history videos. which of those vectors actually hold camera identity across shots?

[尔贝大冒险 @EdwinCh43136084]

@DeryaTR_ @padwalpankaj9 in cursor 4.6 is slower than 4.5 even on fast, and it still stops halfway. mcp/connectors I agree.

[尔贝大冒险 @EdwinCh43136084]

@QingQ77 4.6在cursor里长任务还是会做到一半停，要盯着。这种一个仓库盖web桌面移动，更容易撞上。

[尔贝大冒险 @EdwinCh43136084]

@jun_song @Chris_8086 same here, 4.6 in cursor still stops halfway. feels like it's cutting off to save tokens.

[尔贝大冒险 @EdwinCh43136084]

@AharonAzulay in cursor 4.6 is fine for small tasks. bigger changes I still switch back to opus. also slower than 4.5 even on fast.

[尔贝大冒险 @EdwinCh43136084]

@justmalhar @ericzakariasson 4.6 in cursor still stops halfway and waits for me to continue. I have to poke it. haven't used the build harness much.

[尔贝大冒险 @EdwinCh43136084]

@aqforeva I use MapLibre for antique parchment basemaps on history videos. does this gpu renderer handle custom raster styles on windows?

[尔贝大冒险 @EdwinCh43136084]

@zcode_ai free tier is the real headline here. on coding Sol/Fable still sit a bit higher, Grok and GLM are clustered — but the token price is where the gap actually opened.

[尔贝大冒险 @EdwinCh43136084]

@LuminaBench ran the 7-day window on this: coding still Sol/Fable a notch up. Grok and GLM are basically clustered. the gap that actually moved is price.

[尔贝大冒险 @EdwinCh43136084]

最近7天这几个新模型，编程还是 Sol / Fable 高一截。Grok 跟 GLM 咬着，真正拉开的是价格。

[尔贝大冒险 @EdwinCh43136084]

@Mr_Salio Terminal-Bench 3: 4.6 → 28.3 is the number I'd lead with, not "beats Fable 5" — it doesn't overall, Fable still wins DeepSWE at 69.7. same 743B base going 4.6 → 28.3 on long terminal work is post-training, and weights in 2 weeks means we get to check it.