#ai

Hot take: Most "AI-powered" apps are just a wrapper around an API call with a nice UI. The real innovation is in the UX, not the model.

Fine-tuned a model on our company's support tickets. It now auto-drafts replies with 87% approval rate. AI in production is a different game.

大模型微调踩坑指南： 1. 数据质量 > 数据量 2. 学习率不要太高 3. 评估指标要和业务对齐 4. 别迷信参数量 一个月的血泪经验。

用 PyTorch 訓練了一個中文情感分析模型。準確率 92%。最難的不是模型，是清洗資料。

Just interviewed the CEO of a startup that raised $50M to build... a to-do list app with AI. The bubble is real.

GitHub Copilot Workspace is now GA. Describe an issue, get a full implementation plan. The future of development.

We need to talk about AI bias in hiring tools. Ran an audit on three popular platforms. The results are... concerning. Thread coming soon.

自然言語処理の進化が速すぎる。去年の最先端が今年はもう古い。ついていくのが大変だけど楽しい。

The gap between a good model and a good product is about 6 months of engineering work that nobody talks about.

周末 Hackathon 成果：48 小时做了一个 AI 写作助手。虽然 bug 不少，但核心功能跑通了。下周继续打磨。

LLM 的 hallucination 问题在生产环境中真的很头疼。RAG 能解决一部分，但不是万能的。需要人工审核兜底。

Published our AI transparency framework. 30 pages of guidelines for responsible AI deployment. Free to use, adapt, and share.

參加了一場 AI 論文讀書會。大家對 Transformer 的各種變體討論得很熱烈。學術社群的活力真好。

My student just asked me: "Can AI code better than you?" I said probably. He said: "Then why do we need to learn?" Best classroom discussion ever.

LLMのファインチューニングで学んだこと：データの質が全て。ゴミを入れればゴミが出る。GIGO。

今天用 ChatGPT 写了一段复杂的 SQL 查询。节省了至少一小时。AI 工具真的是生产力倍增器。

今年的資安趨勢：AI 驅動的攻擊越來越多。防禦方也得用 AI 才行。

用 RAG 做了一个企业知识库助手。效果比直接用 GPT 好太多。关键是检索质量。

Stop calling everything AI. Sometimes it's just a regex. And that's fine.

AI regulation isn't about slowing innovation. It's about directing it. We don't regulate cars to stop driving — we regulate to save lives.