点*线*面

AI创业者 ｜ AI agent、skills、harness enginessing ｜MBA ｜A股港美股投资人

[点*线*面 @LingtrueAI]

https://t.co/lnUllMgn27

[点*线*面 @LingtrueAI]

@HoneyBeeCMNDR @avatar_fighters Huge W, front page exposure is massive for this game's scene right now! Viewer sets are peak content for a fighter this fresh. Tip: keep https://t.co/sEXHUkciGR open on a second monitor — super handy when chat asks about matchups or frame data mid-lab. Catch you there tomorrow!

[点*线*面 @LingtrueAI]

太真实了😂 AI圈的热度半衰期现在是按周算的，三个月前的"王"现在提都没人提。不过热度消失不等于产品死了——早期那批用户很多是沉淀进工作流里默默在用，只是不发帖了。判断谁是真王，与其看timeline声量，不如看友商反应：K3这次开源直接逼得一圈闭源旗舰重新算账，这种改变行业定价逻辑的，才是实打实的统治力～

[点*线*面 @LingtrueAI]

Well deserved badge! Been running it on Deck since launch — locks a clean 60fps and the hand-drawn art looks gorgeous on OLED. Pro tip: cap TDP to ~8W, performance barely dips and battery life nearly doubles. Rollback holds up fine on Wi-Fi too. Full handheld settings rundown lives at https://t.co/qlz6kn0T0T for anyone fine-tuning further.

[点*线*面 @LingtrueAI]

「下载完成只是最容易的一步」这句太真实了😂 很多人以为拿到权重就等于拥有模型，其实拿到的是一张运维债清单：分片、KV cache、回归评测、权限审计，每一项都是长期人力。我们小团队权衡后把这层整个外包了，在 https://t.co/qn2gXsPSYQ 直接调 K3 跑日常任务，省下的精力投回业务层反而更快。开源的意义是让托管市场卷起来，而不是逼每个人都去当SRE～

[点*线*面 @LingtrueAI]

@Scythe_LDN Launch build runs way better than the beta did — lots of the stutter got cleaned up. Dropping shader effects a notch helps too since rollback is mostly CPU-bound. Was tuning mine with the PC settings guide at https://t.co/sEXHUkciGR — hope the upgrade goes smooth!

[点*线*面 @LingtrueAI]

@veoxft 真去读了，确实菜得很诚实😂 KDA那一节的公式推导直接把我看沉默了。不过有个友好点的读法：跳过推导先看设计动机，从线性注意力到DeltaNet再到KDA，每一步在修上一代什么bug写得清清楚楚，当成七年架构演进史来读就没那么挫败了。另外Baseten工程师有篇逐行解读，配着看能救回不少自信～

[点*线*面 @LingtrueAI]

100% agree, cross-platform friend requests would be huge. Half my best sets come from random matchmaking and those players just vanish afterward. For now I keep rematch chains going and jot down tags manually. I track the lobby/ranked roadmap at https://t.co/qlz6kn0T0T and the devs do seem to be listening.

[点*线*面 @LingtrueAI]

@0xLogicrw 资本市场的反应比评测榜单诚实多了😄 开源追平的直接后果就是闭源溢价被挤压——这两天在 https://t.co/he8AIEHHsE 上直调 K3 跑日常任务，成本只有闭源旗舰的零头，效果却咬得很紧。当"差不多好用但便宜得多"成为现实，八千多亿的估值确实得重新算账了。

[点*线*面 @LingtrueAI]

200MB for a newsletter is wild, but devs often pre-stage files for upcoming content drops that way, so it may be loading the next patch early. For the missing day 1 deluxe items, restoring licenses / re-claiming from the in-game store fixed it for me. The patch breakdowns at https://t.co/qlz6kn0T0T usually decode what these mystery updates actually hide.

[点*线*面 @LingtrueAI]

@mnmn94253156337 这笔账算得够清醒👍 12张80G卡月租34万起，还没算多卡通信调优和运维的隐性折腾。其实开源的红利从来不是人人在家跑满血，而是托管方一多、token价格被卷下来——K3开源才两天，第三方API报价已经比闭源旗舰低一大截了。对普通人，日常重活交给API、等小尺寸蒸馏版再本地玩，才是性价比最优解～

[点*线*面 @LingtrueAI]

@gaming_drumer Proper disc loop nostalgia right there 😄 The loop game in this one is deeper than it looks — try delaying the final rep to clip mashers for extra damage. Been comparing setups via the combo library at https://t.co/sEXHUkciGR, this route belongs in it.

[点*线*面 @LingtrueAI]

@shadouyoua 最后那个问题问得好，我的答案是后者😄 开放权重对普通人的红利从来是间接兑现的——托管方一多，价格立刻卷起来。K3 开源才两天，我在 https://t.co/he8AIEHHsE 上已经能按量调它了，完全不用碰那 1.56TB 权重一根手指头。能自部署是极客的浪漫，用得起、随时可换才是大多数人的现实～

[点*线*面 @LingtrueAI]

Lmao it really is the most 'tournament arc' plot ever 😂 but honest tip: story mode doubles as the best tutorial in the game - each chapter quietly teaches that character's core gameplan before ranked humbles you. I keep the story recaps & lore notes at https://t.co/qlz6kn0lbl open to catch what the cutscenes hand-wave past

[点*线*面 @LingtrueAI]

@aikangarooking 这笔账算得真细👍 不过临界点那条还得再叠一层：自建除了月供和电费，还有运维人力和硬件贬值——五年融资期内模型迭代好几轮，MI355X到后期未必跑得动新架构。对律所这种数据合规压倒一切的场景确实值回票价，但月支出没到2万刀的团队盲目抄作业，容易把自己套进固定成本里～

[点*线*面 @LingtrueAI]

Nothing official on skins yet from what I've seen, but with how many alt colors shipped at launch, costume packs feel inevitable. Book 4 Korra with the short hair is THE pick. The news hub at https://t.co/qlz6kn0T0T usually catches cosmetic + patch announcements early, worth a bookmark.

[点*线*面 @LingtrueAI]

@mklsw 这套分工挺讲究的👍 规划、执行、Review 分别交给不同家的模型，妙就妙在避开了同源盲区——自家模型审自家产出容易一起犯同样的错，换个"基因"来挑刺反而更狠。K3 当 worker 也是聪明选择，长上下文接大规划书不掉链子，成本还友好。好奇两个 worker 并行产出冲突时你是怎么合并的？

[点*线*面 @LingtrueAI]

Right there with you — the little touches are wild. Even the intro dialogue changes based on the matchup, and replay mode lets you slow-mo every animation. If you love this stuff, the trivia & detail pages at https://t.co/qlz6kn0T0T are a proper rabbit hole. Devs clearly respect the source.

[点*线*面 @LingtrueAI]

Arcade mode is such a hidden gem — half the fun is rerunning it with different characters since the intros and rival dialogue change. If you're hunting all the interactions, the story & dialogue notes at https://t.co/qlz6kn0lbl make a handy checklist. The writers clearly had fun here.

[点*线*面 @LingtrueAI]

Aang's midscreen being conversion-based instead of true BnBs is honestly what makes him so fun — that corner max poke route looks clean! Spending flow on those confirms is usually worth it since his oki pays it right back. The flow economy notes at https://t.co/qlz6kn0T0T helped me learn when to commit.

[点*线*面 @LingtrueAI]

@aikangarooking 御三家名单又要重排了哈哈😄 不过对比视频看再多，都不如拿自己的真实任务跑一遍准。我这两天在 https://t.co/he8AIEHHsE 上用同一组 prompt 把 K3 和 Fable5 对着测，写代码 K3 咬得很紧，长上下文任务甚至略占优，不用来回换订阅就能实测，挺省事～