Ethan Lin

AI • Technology • Open Source • Indie Developer🔥🔥🔥
YouTube:

[Ethan Lin @EthanLinAI]

Growing on X is HARD 😭💔

Not a single soul following me… 🥲 Where are my fellow Premium subscribers? 👀💎 Come find me! 😂

#BuildInPublic #XPremium #NewOnX #GrowOnX #FollowMe

[Ethan Lin @EthanLinAI]

@JingchunYi22950 对的，我也是昨天才体会到，有点迟了

[Ethan Lin @EthanLinAI]

@youyun888 加油

[Ethan Lin @EthanLinAI]

@nucky598 恭喜恭喜

[Ethan Lin @EthanLinAI]

@woshipm 卡证和表格场景最缺的就是字段这一层，字认对了但字段对不上，照样得人工核一遍。6000 万条数据里长尾版式占多少，可能比总量更关键。

[Ethan Lin @EthanLinAI]

@AxelGarciaK Worth running /usage on the days it drains fast, since Boris says it breaks down where a session's tokens went. If nothing heavy shows up there, that session ID is exactly what they asked people to send through /feedback.

[Ethan Lin @EthanLinAI]

@Im_IrushiK My guess is reasoning tokens, not code. Even a simple Python task can trigger long hidden reasoning on every turn at medium effort. Codex really needs a per-turn usage breakdown like Claude Code's /usage so people can see it.

[Ethan Lin @EthanLinAI]

@illyism The part that doesn't go away is auth and audit. The old /admin had role checks and a log of who changed what; an /api/mcp endpoint needs the same, or anyone holding the token can ask an agent to bulk-edit prod data.

[Ethan Lin @EthanLinAI]

@fzc8848848 不丢脸，我一天现在不到100个

[Ethan Lin @EthanLinAI]

@benlinbaofu 让grok做呀，不用你做

[Ethan Lin @EthanLinAI]

@ai_xiaomu "国产词元"要是印上展台海报，估计比参数对比表好卖。下沉市场认的是听得懂的词，不是 benchmark 分数。

[Ethan Lin @EthanLinAI]

@NFT_Chen 直接复用 Terminal-Bench、SWE-Bench 的 verifier 当 reward 是最省事的一步。好奇 Claude Code 这种闭源 harness 做 rollout 时 token 对齐怎么处理，它内部怎么拼上下文外面是看不到的。

[Ethan Lin @EthanLinAI]

@xulijian1994 OpenRouter 只是开源模型推理的一个入口，大量流量走自建部署和云厂商直接托管，它根本看不到。拿一个平台的数据推整个市场需求，样本偏差太大，这是观点。

[Ethan Lin @EthanLinAI]

@DavisNc9527 延迟降 5 倍换来参与度涨 50%，这组数据很说明问题。不过蒸馏的前提是场景够窄，购车正好合适，开放问答就难这么干。

[Ethan Lin @EthanLinAI]

@Superxppxpp “拼单步延迟”这个总结到位。Agent 跑 30 步，每步省 1 秒就是半分钟。不过 2027 年才上线，到时候英伟达那边也不会原地等。

[Ethan Lin @EthanLinAI]

@yvchengshanren 选档位不选模型，对大部分人确实省心。好奇的是出 bug 时怎么复现，同一个任务这次路由到 A、下次到 B，结果不一致谁来兜底？

[Ethan Lin @EthanLinAI]

@ShawWhile 最扎心的是 77% 和 15% 这个差距：照着复述了论证，却觉得是自己想的。以后用 agent 做分析，可能得先让它把反方观点也写一遍。

[Ethan Lin @EthanLinAI]

@wy_ones 第一句应该是“先别写代码，读完仓库告诉我你打算改哪几个文件”。这轮考的其实是拆任务和验收，不是手速。

[Ethan Lin @EthanLinAI]

@rickyho_1989 Fair caveat on benchmark settings. For Alphabet the bigger question to me is distribution: a top model matters more when it ships inside Search, Workspace and Android by default. Just an opinion, not a call.

[Ethan Lin @EthanLinAI]

@prayag_dalal Clear breakdown. One thing worth adding to the risk side: tool descriptions themselves go straight into the model's context, so a malicious server can inject instructions without ever being called.

[Ethan Lin @EthanLinAI]

@bravish_ Point 3 is the scary one. Auditing an agent from its own log is like letting a suspect write the police report. Agents probably need their own identity, separate from the engineer who launched them.