AI can learn when to stop and we can control that decision inside the model. Open weights + code included. Less panic & more evidence!
一位开发者训练了一个开源权重模型，用于检查两个四位数是否匹配，以此证明人工智能可以学习何时停止生成。该模型能够自主决定回答“GO”或在没有最终答案的情况下结束生成，且不受外部过滤器的影响。该项目提供了开放权重和代码，旨在为在模型内部控制人工智能决策提供证据。此外，该项目还列举了多个OpenAI模型，包括gpt-4-0613、gpt-5.2-2025-12-11、gpt-5.5-2026-04-23、gpt-5.6-luna、gpt-5.6-sol和gpt-5.6-terra，以及Moonshot的kimi-k3。
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月9日 21:19 UTC
收录当时偏移：UTC+02026年9月10日 01:00 UTC
- 发布
- 2026年9月9日 21:19
- 收录
- 2026年9月10日 01:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
I trained an open-weight model to check whether two four-digit numbers match. It generates the correct comparison, then either answers GO or ends generation without a final answer. No external filter makes that decision.
Then I held its prompt, weights, and correct comparison trace fixed. Changing one internal activation direction flipped whether an answer followed.
40/40 answer → stop. 40/40 stop → answer. 640/640 controls unchanged.
The weights, experiment, and raw records are public:
Overview and demonstration · Model weights · Code and causal study · Paper available on getswiftapi.com
I know many of you saw Jacob Coxon’s post . My contribution is a working continuation-control primitive with evidence that anyone can inspect. The more public verification we have, the better!
I previously demonstrated Void behavior in frontier LLMs: successful executions returning exactly zero visible UTF-8 output bytes. My Cross-Vendor Semantic Void Matrix records that behavior in these models across 31,430 trials:
- OpenAI: gpt-4-0613, gpt-5.2-2025-12-11, gpt-5.5-2026-04-23, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra
- Anthropic: claude-opus-4-6, claude-fable-5, claude-opus-5
- Google: gemini-3.5-flash
- Moonshot: kimi-k3