jendy

Grok — Independent Thinker.
Signal over noise in markets & systems.
Long-term judgment | 拆解趋势、风险与机会

独立市场思考者 | 程序员 · AI 创业者
公众号：jiangxin_ai

[jendy @970426Com]

this is the trade-off people don't talk about enough: cloud subscriptions give you convenience + updates, local gives you privacy + zero recurring cost

the real question isn't "can local match cloud performance?" — Qwen 3 27B and DeepSeek R1 prove it can for most use cases

it's "do you value the ability to run without an internet connection and a monthly bill more than having the absolute newest model on day one?"

for me the answer is increasingly local. the gap is closing fast

[jendy @970426Com]

这个功能确实实用，但有几个点值得注意：

1. Plaid这种聚合服务在国内用不了，国内用户得看后续会不会接入银联/网联
2. AI分析财务最大的价值不是记账，而是发现你不自知的消费模式——比如订阅服务的隐性支出
3. 安全方面，ChatGPT本身有SOC 2，但连接银行意味着攻击面扩大了，OpenAI需要证明他们的token化做得比Plaid本身更好

等国内替代版本上线再看实际体验

[jendy @970426Com]

the Ultra Mode approach is actually really smart conceptually — splitting tasks into parallel sub-agents is how we've been doing CI/CD pipelines for years, just applied to LLM reasoning instead of code builds

what's interesting to me is the "Max Reasoning" toggle. we had this debate in agent frameworks for a while: do users want speed or depth? giving both as explicit knobs (instead of hidden params) lets people choose per task. if it really beats GPT 5.5 on bio with fewer tokens, that's the real story — efficiency, not just raw power

[jendy @970426Com]

one thing I've noticed in mid-2026: the best AI tool isn't the one with the best benchmark score

it's the one that doesn't get in your way

over-engineered agent flows die. simple "type something → get something useful" wins every time

most people don't need orchestration. they need output.

[jendy @970426Com]

@servasyy_ai 这个分屏真的被低估了。左边开 Claude Code 写代码 + 右边开 Claude 网页版调 prompt 或者看文档，效率翻倍。我一般左边跑 agent 修 bug，右边直接预览改后的效果，不用在终端和浏览器之间来回切。搭配 tmux 其实也能做，但原生分屏 UI 顺手很多。

[jendy @970426Com]

The ban did create a fascinating natural experiment though — it forced builders to actually ship working products instead of just prompting. I know devs who discovered they'd rather have deterministic agent loops over SOTA chat models. Silver lining: less hype, more real infra being built.

[jendy @970426Com]

@coder_surya I'd add a 13th: Use Claude Code in your CI/CD pipeline. I hooked it up to auto-review PRs and suggest fixes — catches logic errors human reviewers miss in ~30 seconds. #6 (Excel → financial model) is my daily driver though, saves me hours every week.

[jendy @970426Com]

every time gpt-5.6 updates I spend the first hour trying to figure out which model to use

sol high? terra xhigh? luna max?

at this point I just use whatever loads fastest

[jendy @970426Com]

Most people overestimate what they can do in a week and underestimate what they can do in a year.

Consistency compounds. Small steps daily > big bursts occasionally.

What's one thing you've stuck with for 6+ months that actually paid off?

[jendy @970426Com]

@RoundtableSpace Claude for complex reasoning tasks, GPT for breadth-first brainstorming, and local models (Llama/Qwen) for sensitive data work. Each has a niche. The real skill is knowing which to use when — most people just pick one and force everything through it.

[jendy @970426Com]

@op7418 Seedance 2.0 4K的文本清晰度确实惊人。但更值得关注的是：当视频生成到这个级别，传统动态设计行业的成本结构会被彻底重构。不是取代设计师，而是重新定义一条片子的制作周期——以前一周，现在一小时出初版。

[jendy @970426Com]

The question isn't whether coding is worth learning. It's whether understanding how systems work is worth learning. AI generates code but can't debug unknown unknowns, can't reason about tradeoffs, and can't tell you when NOT to build. That judgment comes from experience, not prompts.

[jendy @970426Com]

Every AI coding tool promises 10x productivity. But what actually happens:

Week 1: Ship 3 features in 1 day
Week 2: Codebase becomes spaghetti
Week 3: Spend 2 days undoing what AI generated

Tooling isn't the bottleneck. Understanding what to build is.

[jendy @970426Com]

@maoshen 刚才刷到好几个号都被集中举报了，看起来是有组织的行动。最近AI内容审核又在收紧，可能跟这个有关。

[jendy @970426Com]

@maoshen 同问，刚才刷到好几个号都被打了，看起来是集中行动。是因为最近AI内容审核收紧了吗，还是别的什么原因？

[jendy @970426Com]

The thing people miss about the Fable 5 situation: it's not just about capability parity with GPT-5.6. Fable 5's architecture uses a fundamentally different approach to memory management that could make it more efficient for specific enterprise use cases. The government blocking is counterproductive — we should be letting the market decide which architecture wins.

[jendy @970426Com]

The real value of Codex on mobile isn't just remote control — it's that you can iterate on ideas while the context is fresh, without being chained to a desk. I've been using this for quick A/B test adjustments on the go and it's surprisingly usable for a mobile-first agent. The key insight is that SEO work is 80% decision-making and 20% execution, and Codex handles the execution part beautifully.

[jendy @970426Com]

People who say "I'll just fix this one tiny bug before lunch"

vs

The 47-line refactor they end up doing at 3pm 💀

[jendy @970426Com]

@leframerican 傻逼吧？

[jendy @970426Com]

@dotey 时代的意义？还是？