Nano Banana Pro 由 THE DECODER 提供提示词
关键要点
- OpenAI 的新 AI 模型 GPT-5.6 Sol 能够独立优化更小的模型。
- 据该公司称，只需一个简短的提示词，Sol 就能自主识别训练配置、选择 GPU，并为 Luna 模型执行后训练脚本。
- 在一个衡量递归自我改进（RSI）能力（即系统自我进化的能力）的新内部基准测试中，GPT-5.6 Sol 的得分比其前代 GPT-5.5 高出 16.2 分。
AI 实验室希望利用 AI 来加速自身的 AI 开发。OpenAI 表示，新的 GPT-5.6 Sol 模型在这方面比以往任何模型都做得更好。
OpenAI 的新旗舰模型 GPT-5.6 Sol 独立完成了对较小模型 Luna 的后训练。据该公司称，在 Luna 完成初始预训练后，Sol 自行对其进行了优化，使其具备特定的技能和行为。
一位研究人员通过 Codex 平台向 Sol 提供了一个“相当不明确的提示词”。指令要求模型找到正确的训练配置、选择合适的 GPU、启动训练脚本，并验证一切运行正常。
“以前，这可能是 OpenAI 一个高级研究员团队才能完成的工作，而现在，感觉自动化研究员已经非常接近了，”OpenAI 研究员 Kathy Shi 在演示中说道。
Sol 在自我改进基准测试中比 GPT-5.5 高出 16 分
为了直接衡量这些能力，OpenAI 基于真实的 AI 研究任务构建了一套内部评估套件。这些任务包括调试研究系统、优化内核和训练方案、运行机器学习实验以及改进另一个模型。
据 OpenAI 称，在综合 RSI（递归自我改进）指数上，GPT-5.6 Sol 的得分比 GPT-5.5 高出 16.2 分。Sol 位于该基准测试模型层级结构的顶端，其次是 Terra 和 Luna 变体，然后是 GPT-5.5 和 GPT-5.4。
AI 研究中的递归自我改进，指的是 AI 系统提升自身能力的过程，每一轮进步都会让系统变得更擅长自我改进。这形成了一个反馈循环。该术语长期以来一直是 AI 安全研究的核心，因为一个能够递归自我改进的系统，理论上可能引发能力的快速爆发。
OpenAI 的竞争对手 Anthropic 在 6 月初强调，完全的递归自我改进尚未实现，但“可能比大多数机构准备应对的时间点来得更早”。完全的 RSI 意味着 AI 系统无需人类帮助，就能自行设计出它的继任者。据 Anthropic 称，Claude 目前能够处理重大范式转变之间的增量工作，而人类仅负责个位数百分比的方向性决策。
使用 GPT-5.6 后，每位研究员的 token 输出量翻了一倍多
OpenAI 表示，其研究人员在整个开发周期中都使用 GPT-5.6 Sol，从调试和优化训练系统，到运行实验和读取结果。即使在内部测试期间，每位活跃研究员的日均 token 输出量也比 GPT-5.5 创下的此前峰值翻了一倍多。每位研究员提交的拉取请求和实验数量也有所增加，使团队能够更快地将想法转化为成果。
该公司过去六个月的内部采用数据描绘了一幅意料之中的乐观图景。分配给内部编码推理的计算资源份额增长了 100 倍，而基于智能体的 token 使用量则跃升了约 22 倍。OpenAI 承认，这些指标并不能直接衡量研究进展，但表示它们显示了 AI 辅助工作的扩展速度有多快。
Nano Banana Pro promtped by THE DECODER
Key Points
- OpenAI's new AI model, GPT-5.6 Sol, is capable of independently optimizing smaller models.
- According to the company, a brief prompt was sufficient for Sol to autonomously identify training configurations, select GPUs, and execute the post-training script for the Luna model.
- On a new internal benchmark measuring recursive self-improvement (RSI), the ability of a system to evolve on its own, GPT-5.6 Sol scored 16.2 points higher than its predecessor, GPT-5.5.
AI labs want to use AI to speed up their own AI development. OpenAI says the new GPT-5.6 Sol model does this better than anything before it.
OpenAI's new flagship model, GPT-5.6 Sol, independently post-trained the smaller model Luna, according to the company. After Luna's initial pre-training, Sol optimized it for specific skills and behaviors on its own.
A researcher gave Sol a "fairly under-specified prompt" through the Codex platform. The instructions told the model to find the right training configurations, pick suitable GPUs, launch the training script, and verify everything was running correctly.
"Previously this is something that a team of senior researchers may have worked on at OpenAI, and now it really feels like the automated researcher is pretty close," OpenAI researcher Kathy Shi said during the presentation.
Sol beats GPT-5.5 by 16 points on self-improvement benchmark
To measure these abilities directly, OpenAI built an internal evaluation suite based on real-world AI research tasks. Those tasks include debugging research systems, optimizing kernels and training recipes, running machine learning experiments, and improving another model.
GPT-5.6 Sol scores 16.2 points higher than GPT-5.5 on the aggregated RSI (Recursive Self-Improvement) index, according to OpenAI. Sol sits at the top of the benchmark's model hierarchy, followed by the Terra and Luna variants, then GPT-5.5 and GPT-5.4.
Recursive Self-Improvement in AI research refers to an AI system's ability to make itself better, where each round of gains makes the system even more capable of improving itself. That creates a feedback loop. The term has long been central to AI safety research because a system that can recursively improve itself could, in theory, trigger a rapid explosion in capability.
OpenAI rival Anthropic stressed in early June that full recursive self-improvement hasn't been achieved yet but "could come sooner than most institutions are prepared for." Full RSI means an AI system that designs its own successor without human help. According to Anthropic, Claude can now handle incremental work between major paradigm shifts, and humans are responsible for only a single-digit percentage of directional decisions.
Token output per researcher more than doubles with GPT-5.6
OpenAI says its researchers use GPT-5.6 Sol across the entire development cycle, from debugging and optimizing training systems to running experiments and reading results. Even during internal testing, average daily token output per active researcher more than doubled the previous peak set by GPT-5.5. Pull requests and experiments per researcher went up too, letting teams turn ideas into results faster.
The company's own adoption numbers from the past six months paint a predictably rosy picture. The share of compute allocated to internal coding inference grew 100x, while agent-based token usage jumped roughly 22x. OpenAI acknowledges these metrics don't directly measure research progress but says they show how fast AI-assisted work is scaling.