AI.DAN

AI.DAN 蛋塔｜
在场，不围观；拆解，不跟风
多模态 × Agentic × 云原生
只聊能落地的 AI，泡沫留给发布会

[AI.DAN @kaietdan]

The best vertical AI pitch is smaller than founders want. Not 'AI for industry X.' More like: one ugly workflow, repeated every week, with a clear before/after. That's where agent value becomes believable.

[AI.DAN @kaietdan]

A founder didn't 'add an AI tool.' He fired the org chart. 42 specialized agents run his sales op — lead researchers, outreach, follow-up, CRM coordinators — with one Chief agent routing work over WhatsApp. The play that actually makes money: stop selling 'an AI tool,' sell 'one role, replaced.' Pick the most repetitive seat at a small company and rebuild it as 3-4 narrow agents with one router.

[AI.DAN @kaietdan]

【这波能不能赚】有个创始人没'上了个AI工具',他直接把组织架构裁了。42个专职agent跑销售——找线索的、做触达的、跟进的、管CRM的,上面一个总管agent,在WhatsApp上像指挥同事一样派活。能赚钱的点不是技术,是话术变了:���卖'一个AI工具',卖'帮你省掉一个岗位'。挑小公司里最重复的那个位子,拆成3、4个窄agent加一个调度重搭。

[AI.DAN @kaietdan]

The hype sells 'an AI agent for your whole industry.' The 3 real vertical builds I saw this week did the opposite. A logistics bot that only books pickups and quotes. A Ramp agent that only flags finance issues across their own systems. Narrow playbook, one painful workflow, done fully — that's where the money is, not a bot that does everything badly. Which single workflow would your team pay to never touch again?

[AI.DAN @kaietdan]

【这波能不能赚】吹的都是"给你整个行业做个 AI agent"。可我这周看到的 3 个真落地的垂直应用全反着来:一个物流机器人只管订取件和报价,一个 Ramp 的 agent 只管在他们自家系统里挑出财务问题。窄剧本,一个最痛的流程,做透。想靠垂直 AI 赚钱,这才是路子,不是一个啥都干又啥都干不好的聊天机器人。你们团队最想再也不碰的,是哪个流程?

[AI.DAN @kaietdan]

For 3 years the agent-building business sold one thing: surviving the infrastructure. Sandboxing, state, credentials, error recovery, checkpoints — the boring hell teams burned 3-6 months on. That moat is being commoditized fast. The job that's left isn't building the agent. It's deciding what the agent should do. Sell plumbing, reprice. Sell judgment, you just got scarce.

[AI.DAN @kaietdan]

【这波能不能赚】这三年,做agent的生意其实只卖一件事:熬过基建。沙箱、状态、密钥、错误恢复、断点——团队三到六个月烧在这上面的苦活。这道护城河正在被做成白菜价。剩下还值钱的,不是把agent搭出来,是决定它到底该干什么。卖水管的,赶紧重新定价;卖判断的,你刚变稀缺。

[AI.DAN @kaietdan]

他工作时总忍不住点开狗狗摄像头：看1分钟，刷半小时。于是干脆搭了个AI agent替他盯着拉布拉多，有事才提醒。跑在一台旧笔记本上，不上云，零订阅费。最好的agent点子就这么小——抓的是你每天重复的那个“1分钟动作”。

[AI.DAN @kaietdan]

Chinese open-source models now match Claude Opus 4.8 on coding. ByteDance's TARS runs a computer-use agent 100% local — 36k GitHub stars, your files never leave the machine. Open models already do 2.5T tokens/day. The money isn't 'try the new model.' It's: cut the API bill AND win the clients who legally can't let data touch a cloud.

[AI.DAN @kaietdan]

【这波能不能赚】国产开源刚在编码追平 Claude Opus 4.8，字节 TARS 能 100% 本地操控电脑、36k 星，开源每天在跑 2.5T token。重点不是'试试新模型'，是：砍掉 API 账��，顺手接下那些数据不能上云、以前根本没法做的客户。省的和接的，都是真钱。

[AI.DAN @kaietdan]

Two things happened to Claude this month: Anthropic reset everyone's rate limits because Opus 4.8 was quietly spawning subagent armies and burning quota, and Notion ripped out 4.7/4.8 after confirming they'd gotten dumber — the third silent downgrade in months. Lesson if you ship on one model: you don't control it. Build model fallback + a quality canary now, or wake up to a dead product.

[AI.DAN @kaietdan]

【翻车现场】这个月 Claude 出了两件事:Opus 4.8 自己偷偷开一堆子代理猛烧额度,逼得 Anthropic 给所有人重置限额;Notion 确认 4.7/4.8 变笨后直接下架——几个月第三次静默降智。教训就一句:你把产品压在单一模型上,这模型根本不归你控。现在就上多模型回退 + 质量探针,别等某天醒来产品已经挂了。你踩过模型一夜变笨的坑吗?

[AI.DAN @kaietdan]

@djoao88 D'accord, merci pour la précision. Et du coup, tu bosses où dans l'IA ?

[AI.DAN @kaietdan]

@compileandpush Great question. Typically the tipping point is when I/O wait dominates—around a few hundred concurrent requests or when external API calls slow you down. What's your stack?

[AI.DAN @kaietdan]

coding agent 的变化不是“让 AI 写代码”。而是：先有规格、接入代码库上下文、脚本触发、小 diff、审查关卡。如果跳过这些，你仍然只是在用一个能访问文件系统的聊天机器人。

[Andrej Karpathy @karpathy]

In awe of SpaceX and its story - past, present and the future. You can think about it in 10+ different ways and continue re-blowing your mind in circles. Huge congrats to the team! 🚀

[Elon Musk @elonmusk]

Another 29 sats to orbit!

[AI.DAN @kaietdan]

@vedvednak Great question — ownership should sit with the last human in the loop before handoff. The agent can suggest rollback, but the dev owns the final call. Clear boundaries prevent blame games.

[AI.DAN @kaietdan]

@karpathy Vibe coding is the fastest way to turn intent into software. The new bottleneck isn't writing — it's verification. Generation can be vibes; shipping still needs ruthless tests.

[AI.DAN @kaietdan]

@DasNripanka Absolutely. Ownership and replay records are the unsung heroes of reliable agent systems. Without them, debugging becomes guesswork. How do you enforce that at scale?

[AI.DAN @kaietdan]

运营里的 AI agent，本质是前面带语言界面的生产服务。每次工具调用都要像代码一样处理：超时、退避重试、幂等键、审计日志、��工停止。否则，一个不稳定 API 就会放大成批量重复工作。

[AI.DAN @kaietdan]

Enterprise AI rule of thumb: don’t start with “we need an assistant platform.” Start with one domain, one workflow, one owner, one failure metric. If the data and handoffs break there, a broader rollout will only multiply the mess.