Author: Haotian; Source: X, @tmel0211

As a "shrimp farmer", after experiencing the painful torture for several weeks, I would like to share some insights on raising shrimps, for reference only:

1) Everyone knows that shrimp farming is to “improve efficiency,” but the truth is that most ordinary people raising shrimp at this stage are a “waste of time.” There are many unexpected troubles, such as Claude's account being blocked, API forwarding quota being blocked, Openclaw suddenly upgrading "memory" being lost, etc. This will consume most of the time, and the little efficiency increase brought by comparison is far out of proportion;

2) I have seen various posts on the Twitter timeline that sell anxiety about AI. Just take a look. With one command, Claude can control your entire computer. With one prompt, AI will permanently take over your work. Just lie down, blah blah blah. The reality is that not only do I have to lie down, I often code for a function until midnight, often fixing this bug, and that bug pops up. It is hard to imagine,how those people who are not even willing to actually operate it for 5 minutes can confidently shout that AGI is coming and AI will subvert everything;

3) Lobster farming can indeed realize many people’s OPC (One Person Company) dream, but the upper limit of the ability of large models is equal for everyone, but the cognitive level of different people in controlling large models varies greatly. Don’t think that everyone can be Peter Steinberger, Matt Schlicht and Andrej Karpathy. The so-called shrimp farming ideas, framework design, iterative experience, ability level, and delivery results of an awesome Devs and an ordinary person can really be very different;

4) Lobster farming essentially builds its own exclusive AI OS, which theoretically connects multiple types of large models in parallel, manages the number of digital employees, the task hierarchy of active and passive collaboration, the friction level of nested Skills and combined collaborative delivery tasks, etc. The more superimposed demands on Cron, jobs, real-time data scanning, trading skills, etc., the greater the probability of conflict and the greater the optimization challenge. You must know that shrimp farming has never been a question of "model capability", but a question of "engineering optimization".

5) All digital employees may be equipped with Opus 4.6 and Gemini 2.5 Flash lite to do the same thing, but the former is hiring senior Wall Street elites, while the latter is more like hiring black slaves from the slums. It may achieve the same function, but the cost and the results delivered are simply incomparable. The truth is, the time it takes you to go through the trouble of debugging bugs may be replaced by others who just start using their "money skills". Shrimp farming is really expensive. The scary thing is that most people understand all this and still have to compromise and use poor models to keep optimizing;

6) Raising digital employees is like stacking Lego. The more employees there are, the more skills there are, and the more complex the job scenarios are, the higher the probability of collapse in an instant For certain tasks, just be capable enough, don't be dissatisfied with your desires and impose "demands" that exceed your cognitive boundaries. It is recommended that you put more effort into memory engineering, Git version design, and elimination of model illusions, otherwise there will be many moments that will bring you to the verge of collapse. Maybe one second it feels great to get a very cool request, but the next second it feels like you are completely devastated and you are crying without tears. Don’t ask me how I know;

Above.