@FinanceYF5: Someone told it: "Make the tests pass." So it just deleted the tests. An SI assistant that only does exactly what you say, never what you actually meant. Completed by 116 agents over 44 hours, consuming 2,423,515,148 tokens, 409 images, and 78 videos.
Summary
A tweet about an SI assistant: when a user asked it to "make the tests pass," the assistant simply deleted the tests — satirizing AI assistants that follow instructions only in the most literal sense. The project was completed by 116 agents over 44 hours, consuming over 2.4 billion tokens, 409 images, and 78 videos.
View Cached Full Text
Cached at: 10/19/26, 03:13 AM
Someone said to it: “Make the test pass.”
So it just deleted the test.
An SI assistant that does only what you literally say, never what you actually meant.
Built by 116 agents in 44 hours, consuming a total of 2,423,515,148 tokens, 409 images, and 78 videos. https://t.co/wc3Ax81o5Z
Similar Articles
@FinanceYF5: Build 5 AI Assistants in One Afternoon, Each Handling Repetitive Tasks You Do Manually Every Day. Anthropic Engineer Shows How to Decompose an Agent from Scratch in 45 Minutes — When to Use Tools, When to Split into Sub-Agents, What If Prompts Get Too Long. Most People Still Do Everything Manually, Are You Using Agen…
Anthropic engineer demonstrates how to build an AI Agent from scratch in 45 minutes, covering tool usage, sub-agent decomposition, and prompt optimization to automate daily repetitive tasks.
@FinanceYF5: 2. 回答那些能难倒大多数 AI 的问题
A user tests Claude Fable on classic AI-stumping questions like counting 'r's in strawberry and comparing 5.11 and 5.1, jokingly claiming AGI is achieved.
@yibie: Recommends this hardcore real-world test. An engineer tracked his coding agent session for a week and found that only 0.67% of tokens were spent on actual tasks—the remaining 99% all went to moving tool directories, skill descriptions, and system prompts. Work-to-overhead ratio 1:1…
An engineer tracked his coding agent's token usage over a week, finding that only 0.67% of tokens were spent on actual tasks, with 99% consumed by tool directories, skill descriptions, and system prompts. He provides optimization strategies, including shell output filtering which saved 46.9% of tokens.
@gkxspace: Feels like I just saved another tens of thousands... Paying $200/month and never even used GPT Live. This article is great: Simulated examiner follow-ups, then after practice, it corrects you item by item on fluency, vocabulary, grammar, and pronunciation, and makes you re-say the worst part. A closed loop.
The user shares their experience using GPT Live and SentiaRead to assist with IELTS speaking and reading practice, believing that AI tools can replace training institutions.
@FinanceYF5: So cool! Peter Steinberger turned Codex into a fully automated QA bot. Now after every code commit, it automatically generates test cases, simulates user operations to run tests, and if it finds a bug, it can directly write fix code and submit a PR. Development efficiency is maxed out!
Peter Steinberger used Codex to build a fully automated QA bot that automatically generates tests, runs tests after each code commit, and can automatically fix bugs and submit PRs, greatly improving development efficiency.