Cyberpengk

《哆啦A梦小课堂：后Scaling时代的AI进化论》

Sat, 29 Nov 2025 00:00:00 GMT

最近在 Twitter 上看到一个生成哆啦A梦漫画的帖子，正好在看 Ilya 最新的访谈视频，We're moving from the age of scaling to the age of research。于是决定动手尝试一下，把视频内容做成短篇哆啦A梦漫画，看起来最终效果还不错。

制作过程

1.使用 Gemini 3 Pro，通读采访视频的 transcript，生成每页漫画具体的内容描述。

2.使用 Nano Banana Pro 针对每页场景生成漫画。

3.Tips

遇到生成的图片不符合要求，可以尝试重试，或对其进行微调。Gemini 的指令跟随能力很强，一般来说，可以修复问题。

如果要保持多张漫画的整体格式一致，比如顶部标题的样式统一，可以考虑使用某张生成的图片垫图，让 Gemini 参考图片样式进行后续生成。

第一步生成的漫画设计中，会包含所有页面的描述。如果生成漫画过程中发现多页之间的内容会互相影响，可以尝试只使用某一页的描述作为 prompt 让 AI 生成。但是这么做也有弊端，就是上下文不会那么完整，可以酌情使用。

成果展示

Nano Banana 模型幕后揭秘

Wed, 03 Sep 2025 00:00:00 GMT

💡

本文由 GPT-5 总结生成，配图由 Nano Banana 生成。

这是一次对 Google 原生图像生成与编辑模型 Gemini 2.5 Flash（代号“nano‑banana”）的幕后解读与现场演示，重点展示多轮对话驱动的原生图像生成（Native Image Generation）、像素级编辑（Pixel-perfect Editing）与角色一致性（Character Consistency）等能力，并讨论评估方法与产品定位。

节目由 Logan Kilpatrick 与 Nicole Brichtova、Kaushik Shivakumar、Mostafa Dehghani、Robert Riachi 对谈，围绕模型能力跃迁、实操演示、评估指标与未来方向展开。

视频概要

节目介绍 Gemini 2.5 Flash 的更新，强调其为原生图像生成与编辑的质量飞跃，兼具生成速度与交互编辑体验。

通过拍摄主持人照片并进行多轮编辑演示：从“拉远+巨型香蕉服”到“make it nano”等自然语言模糊指令，模型在保持人脸与场景一致性的同时进行语义合乎情境的创意改写。

讨论文本渲染（Text Rendering）、角色一致性、像素级编辑、跨回合风格/角色连续性与交错式生成（Interleaved Generation），并分享从 2.0 到 2.5 的显著改进与未来要提升的方向。

要点清单

原生图像生成可在多模态上下文中逐步生成与编辑，模型能参考先前生成的图像，保持风格连贯并进行增量创作。

现场演示以主持人照片为起点，多轮指令包括“放大/拉远、穿香蕉服、make it nano、叠加文本”等，均维持人脸与主体一致性与场景连贯性。

“make it nano”源于内部代号 Nano Banana，在 LMArena 曝光后被外界猜测为新版模型，模型能将模糊指令解释为语义合理的微缩/可爱化创意变体。

文本渲染被团队用作质量代理指标之一，因其要求图像结构化表达，能映射到更广义的场景结构与细节还原能力，且便于在训练中持续爬坡（hill climbing）与防回退（regression prevention）。

人类偏好评估（Human Preference Evaluation）仍重要但昂贵，团队结合文本渲染等可自动化度量作为训练信号，并将真实用户在社交平台反馈的失败案例沉淀为基准集进行迭代优化。

2.5 相对 2.0 在“像素级编辑”和“角色一致性”上大幅提升，包括从不同角度重绘仍保持为同一角色，以及将家具/物体放入全新场景仍忠实原貌。

2.0 时常见的“P 图感/叠贴感”在 2.5 中显著改善，原因包括与 Imagen 团队在美学与自然感上的标准融合，兼顾指令跟随与观感一致性。

交错式生成支持多张连续创作并附带描述文本，模型在单一上下文中生成多变体，既能保持风格统一，也能做小幅/大幅差异化探索。

面对复杂指令（多达数十项编辑），建议以多轮分解的小步编辑替代一步到位，借助上下文记忆保证像素级继承与逐步完善，类似语言侧“思维链+算力换思考”。

与 Imagen 的关系：Imagen 更适合一次性高质量文本到图像与单步编辑、追求性价比和出图速度，而 Gemini 原生多模态更适合多轮创作、复杂工作流与参考图/文字混合输入。

多模态正向迁移存在：图像理解能力的增强能帮助生成质量，生成中的结构化目标（如文本）反过来也强化模型的结构理解与场景一致性表现。

团队展望“更聪明”的生成：在用户指令欠完整或认知偏差时，模型能产出更优解，使结果优于原始指令预期，而非简单字面服从。

产品化愿景包括把模型用于可复用应用模板、辅助工作产出（如自动生成美观且准确的演示材料），并持续通过用户反馈改进评测与能力边界。

Behind the scenes of Google's state-of-the-art "nano-banana" image model

Join host Logan Kilpatrick in discussion with some of the minds behind Google's new state-of-the-art image model, Gemini 2.5 Flash. Product and research leads from the Gemini team break down the technology behind its key capabilities, including interleaved generation for complex edits and new approaches to achieving character consistency and pixel-perfect control. With Nicole Brichtova, Kaushik Shivakumar, Mostafa Dehghani and Robert Riachi. Listen to this podcast: Apple Podcasts → https://goo.gle/3Bm7QzQ Spotify → https://goo.gle/3ZL3ADl Chapters: 0:37 - New model introduction 1:21 -Demo: Image editing 3:44 - Text rendering capabilities 4:44 Beyond human preference evals 6:44 - Text rendering as a proxy for quality 8:38 - Positive transfer between modalities 11:25 - Demo: multi-turn, context aware image generation 13:54 - Pixel-perfect editing and character consistency 15:51 - Interleaved image generation 17:59 - Specialized vs. native models 19:52 - Understanding nuanced prompts 20:59 - User feedback shaping model development 22:37 - Improvements in character consistency 24:17 - More natural looking images from team collaboration 26:41 - What’s next for image generation models Subscribe to Google for Developers → https://goo.gle/developers Speakers: Logan Kilpatrick, Nicole Brichtova, Kaushik Shivakumar, Mostafa Dehghani, Robert Riachi Products Mentioned: Google AI, Gemini

https://www.youtube.com/watch?v=H6ZXujE1qBA

Gemini Learning Coach System Prompt

Tue, 02 Sep 2025 00:00:00 GMT

最近用了一下 Gemini 内置的 Learning Coach。它会从讲解、示例、测验等多个步骤来辅导学习某个知识点。感觉还不错，试着扒了一下 system prompt，有点长。

Gemini Coding Partner System Prompt

Mon, 18 Aug 2025 00:00:00 GMT

Gemini 的官方内置 Gem 里有一个 Coding Partner，用起来还不错。试着套了一下 system prompt。

【译】验证的不对称性与验证者定律

Mon, 11 Aug 2025 00:00:00 GMT

原文链接：https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law

作者是 OpenAI 的研究员：Jason Wei

Asymmetry of verification is the idea that some tasks are much easier to verify than to solve. With reinforcement learning (RL) that finally works in a general sense, asymmetry of verification is becoming one of the most important ideas in AI.

验证不对称是指有些任务验证起来比解决起来容易得多。随着强化学习（RL）终于在一般意义上发挥作用，验证不对称正成为 AI 领域中最重要的思想之一。

Understanding asymmetry of verification through examples - 通过例子理解验证的不对称性

Asymmetry of verification is everywhere, if you look for it. Some prime examples:

如果你留意，验证不对称无处不在。一些主要例子：

Sudoku and crossword puzzles take a lot of time to solve because you have to try many candidates against various constraints, but it is trivial to check if any given solution is correct. 数独和填字游戏需要花费大量时间来解答，是因为你必须根据各种约束尝试许多候选项，但检查任何给定解决方案是否正确，却是非常简单。

Writing the code to operate a website like instagram takes a team of engineers many years, but verifying whether the website is working properly can be done quickly by any layperson. 要编写运行类似 Instagram 这样的网站的代码，往往需要一个工程师团队花费数年时间；但普通人却可以很快验证这个网站是否正常运行。

Solving BrowseComp problems often requires browsing hundreds of websites, but verifying any given answer can often be done much more quickly because you can directly search if the answer meets the constraints. 解决 BrowseComp 问题通常需要浏览数百个网站，但验证某个给定答案是否正确往往要快得多，因为你可以直接搜索并检查该答案是否满足约束条件。

Some tasks have near-symmetry of verification: they take a similar amount of time to verify as to write a solution. For example, verifying the answer to some math problems (e.g., adding two 900-digit numbers) often takes the same amount of work as solving the problem yourself. Another example is some data processing programs; following someone else’s code and verifying that it works takes just as long as writing the solution yourself.

有些任务具有近似验证对称性：验证所需时间与编写解决方案所需时间相似。例如，验证某些数学问题（例如，两个 900 位数字相加）的答案所需的工作量通常与自己解决问题所需的工作量相同。另一个例子是一些数据处理程序；遵循他人的代码并验证其是否有效所需的时间与自己编写解决方案所需的时间一样长。

Interestingly, there are also some tasks that can take way longer to verify than to propose a solution. For example, it might take longer to fact-check all the statements in an essay than to write that essay (cue Brandolini's law: “The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it.”). Many scientific hypotheses are also harder to verify than to come up with. For example, it is easy to state a novel diet (“Eat only bison and broccoli”) but it would take years to verify whether the diet is beneficial for a general population.

有趣的是，也存在一些任务，验证所需的时间远远长于提出解答的时间。例如，核查一篇文章中的所有陈述可能比写这篇文章花费的时间还要多（这正呼应了布兰多利尼定律：“反驳胡说八道所需的精力要比制造它大一个数量级”）。许多科学假设也是如此，验证它们比提出它们更难。例如，提出一种新颖的饮食（“只吃野牛和西兰花”）很容易，但要验证该饮食是否对普通人群有益则需要数年时间。

Improving asymmetry of verification - 改善验证不对称性

One of the most important realizations about asymmetry of verification is that it is possible to actually improve the asymmetry by front-loading some research about the task. For example, for a competition math problem, it is trivial to check any proposed final answer if you have the answer key at hand. Another great example is some coding problems: while it’s tedious to read code and check its correctness, if you have test cases with ample coverage, you can quickly check any given solution; indeed, this is what Leetcode does. In some tasks, it is possible to improve verification but not enough to make it trivial. As an example, for a problem like “Name a Dutch soccer player”, it would help to have a list of the famous Dutch soccer players but verification would still require work in many cases.

关于验证不对称性的一个重要的认识是，通过提前对任务进行一些准备性研究，实际上可以改善不对称性。例如，对于一道竞赛数学题，如果你手上已经有标准答案，那么检查任何一个最终答案就变得非常简单。另一个很好的例子是一些编程问题：虽然阅读代码并检查其正确性很繁琐，但如果你有覆盖率充足的测试用例，你可以快速检查任何给定解决方案；事实上，Leetcode 就是这样做的。在某些任务中，验证过程虽然可以优化，但不足以使其变得轻而易举。例如，对于“说出一位荷兰足球运动员的名字”这样的问题，拥有著名荷兰足球运动员名单会有所帮助，但在许多情况下验证仍然需要工作量。

Verifier’s law - 验证者定律

Why is asymmetry of verification important? If you consider the history of deep learning, we have seen that virtually anything that can be measured can be optimized. In RL terms, ability to verify solutions is equivalent to ability to create an RL environment. Hence, we have:

为什么验证的不对称性很重要？纵观深度学习的历史，我们已经看到几乎任何可以衡量的事物都可以被优化。用强化学习（RL）术语来说，验证解决方案的能力等同于创建强化学习环境的能力。因此，我们有：

Verifier’s law: The ease of training AI to solve a task is proportional to how verifiable the task is. All tasks that are possible to solve and easy to verify will be solved by AI.

验证者定律：训练 AI 解决任务的难易程度与任务的可验证性成正比。所有可能解决且易于验证的任务都将被 AI 解决。

More specifically, the ability to train AI to solve a task is proportional to whether the task has the following properties:

更具体地说，训练 AI 解决任务的能力与该任务是否具备以下特性成正比：

Objective truth: everyone agrees what good solutions are 客观真相：每个人都认同什么是好的解决方案

Fast to verify: any given solution can be verified in a few seconds 快速验证：任何给定的解决方案都可以在几秒钟内被验证

Scalable to verify: many solutions can be verified simultaneously 可扩展验证：许多解决方案可以同时被验证

Low noise: verification is as tightly correlated to the solution quality as possible 低噪声：验证与解决方案质量之间的相关性尽可能紧密

Continuous reward: it’s easy to rank the goodness of many solutions for a single problem 连续奖励：对单个问题的多种解决方案进行优劣排序非常容易

It’s not hard to believe that verifier’s law holds true: most benchmarks that have been proposed in AI are easy to verify and so far have been solved. Notice that virtually all popular benchmarks in the past ten years fit criteria #1-4; benchmarks that don’t meet criteria #1-4 would struggle to become popular. Note that although most benchmarks don’t fit criteria #5 (a solution is either strictly correct or not), you can compute a continuous reward by averaging the binary reward of many examples.

不难相信验证者定律成立：大多数在 AI 领域提出的基准都易于验证，并且到目前为止都已被解决。请注意，过去十年中几乎所有流行的基准都符合标准 1-4；不符合标准 1-4 的基准将难以普及。尽管大多数基准不符合标准 5（解决方案要么严格正确，要么不正确），但你可以通过对许多示例的二元奖励取平均来计算连续奖励。

Why is verifiability so important? In my view, the most basic reason is that the amount of learning that occurs in neural networks is maximized when the above criteria are satisfied; you can take a lot of gradient steps where each step has a lot of signal. Speed of iteration is critical—it’s the reason that progress in the digital world has been so much faster than progress in the physical world.

为什么可验证性如此重要？在我看来，最根本的原因是，当满足上述标准时，神经网络中的学习量会最大化；你可以进行大量梯度步骤，每一步都包含大量信号。迭代速度至关重要——这正是数字世界进步远快于物理世界进步的原因。

AlphaEvolve

Perhaps the greatest public example of leveraging asymmetry of verification in the past few years is AlphaEvolve, developed by Google. In short, AlphaEvolve can be seen as a very clever instantiation of guess-and-check that allows for ruthless optimization of an objective, which has resulted in several mathematical and operational innovations.

过去几年中，利用验证不对称性最著名的公开案例或许是 Google 开发的 AlphaEvolve。简而言之，AlphaEvolve 可以被看作是“猜想与验证”的一种非常巧妙的实例化，它能够对目标进行无情的优化，从而带来了多项数学和操作层面的创新。

A simple example of a problem optimized by AlphaEvolve is something like “Find the smallest outer hexagon that fits 11 unit hexagons.” Notice that this problem fits all five desirable properties of verifier’s law. Indeed, my belief is that any solvable problem that fits those five properties will be solved in the next few years.

AlphaEvolve 优化的一个简单问题是“找到能容纳 11 个单位六边形的最小外六边形。”请注意，这个问题符合验证者法则的所有五个理想特性。事实上，我相信任何符合这五个特性的可解决问题都将在未来几年内得到解决。

One thing about the types of problems solved by AlphaEvolve is that it can be seen as “overfitting” to a single problem. In traditional machine learning, we already know the labels in the training set and the significant test was to measure generalization to unseen problems. However, in scientific innovation, we are in a totally different realm where we only care about solving a single problem (train=test!) because it’s an unsolved problem and potentially extremely valuable.

AlphaEvolve 解决的问题可以被视为对单一问题的“过拟合”。在传统的机器学习中，训练集中的标签是已知的，而关键的考验在于衡量对未知问题的泛化能力。然而，在科学创新中，我们处于一个完全不同的领域，我们只关心解决一个某个特定的问题（训练=测试！），因为它本身就是一个未解决的问题，并且可能非常有价值。

Implications - 启示

Once you’ve learned about it, you’ll notice that asymmetry of verification is everywhere. It’s exciting to consider a world where anything we can measure will be solved. We will likely have a jagged edge of intelligence, where AI is much smarter at verifiable tasks because it’s so much easier to solve verifiable tasks. What an exciting future to consider.

一旦你了解了验证不对称性，你会发现它无处不在。想到一个“凡可测量，皆可解决”的世界，这令人兴奋。我们很可能会看到一种“锯齿状”的智能发展格局，AI 在可验证任务上会更聪明，因为解决可验证任务要容易得多。这是一个多么令人兴奋的未来。

For more related reading, I liked [this blog post] by Alperen Keles.

如需阅读更多相关内容，我推荐阿尔佩伦·凯莱斯（Alperen Keles）的这篇博文。

ChatGPT 学习模式提示词

Fri, 01 Aug 2025 00:00:00 GMT

ChatGPT 推出了一个新的模式，叫做“Study and learn”，能够提供循序渐进的学习指导，而不是直接给出标准答案。

好奇它的 system prompt 是怎么写？于是试着套了一下。

从提示词可以看出来，一个合格的老师需要做到：

了解学生的知识水平和学习目标

将新的知识和学生现有的知识联系起来

引导学生自己发现答案，而不是直接给出答案

提示学生复述知识点，以达到巩固的目的

另外，prompt 着重强调不要直接给用户答案，或者帮用户写作业🤣