Google quietly crushed every OpenAI model on the market with the release of Gemini 2.5 Pro.

隨著 Gemini 2.5 Pro 的發佈，谷歌悄然碾壓了市場上所有的 OpenAI 模型。

Meanwhile, DeepSeek, Tencent, and Qwen all released banger Chinese models, but nobody's talking about it.

與此同時，DeepSeek、騰訊和 Qwen 都發布了更酷的中國模型，但卻無人問津。

Instead, every eyeball in the tech world right now is hypnotized by OpenAI's brand new GPT-4o image generator, which is transforming the internet into a Ghibli anime cartoon nightmare.

相反，現在科技界的每一隻眼球都被 OpenAI 全新的 GPT-4o 影像基因催眠，這讓網路變成了吉卜力動畫的噩夢。

This is exactly the AI dystopia that Senpai Miyazaki, creator of the Ghibli animation studio tried to warn us about years ago.

這正是吉卜力動畫工作室的創始人宮崎駿前輩多年前試圖警告我們的 AI 烏托邦。

When he said, "I am utterly disgusted. If you really want to make creepy stuff, you can go ahead and do it.

他說：「我非常反感。如果你真的想做令人毛骨悚然的東西，你可以去做。

I would never wish to incorporate this technology into my work at all. I strongly feel that this is an insult to life itself."

我根本不希望將這種技術應用到我的工作中。我強烈認為，這是對生命本身的侮辱。」

That was years ago, but now this nightmare is in full effect.

那是幾年前的事了，但現在噩夢又開始了。

If you make the mistake of going on Twitter today, you'll find that all your favorite memes have been ruined by GPT-4o.

如果你今天誤打誤撞上了 Twitter，你會發現你最喜歡的迷因都被 GPT-4o 毀了。

In today's video, we'll take a closer look at OpenAI's latest toy, along with a few other awesome tools that just put us one step closer to the singularity.

在今天的影片中，我們將近距離了解 OpenAI 的最新玩，還有其他一些很棒的工具，讓我們離奇點更近了一步。

It is March 28th, 2025, and you're watching the Code Report.

現在是 2025 年 3 月 28 日，你正在收看 Code Report。

When I read in the newspaper that GPT-4o added an image generator, my expectations were extremely low.

當我在報紙上看到 GPT-4o 增加了影像生成器時，我的期望值極低。

Especially after the disappointments of Sora and GPT-4.5.

尤其是在 Sora 和 GPT-4.5令人失望之後。

Lowered expectations.

降低期望值。

However, OpenAI just went out and totally redeemed itself by shipping some actual fire.

不過，OpenAI 剛剛推出了一些真正的火種，完全救贖了自己。

This tool has allowed me to throw graphic design software like Canva in the garbage.

這個工具讓我可以把 Canva 等平面設計軟體扔進垃圾桶。

Because now I can create infographics and marketing material with near perfect text rendering.

因為現在我可以用近乎完美的文字渲染來製作資訊圖表和行銷材料。

It even produced an almost good comic strip for me.

它甚至為我創作了一幅幾乎不錯的連環漫畫。

And it can even handle transparency layers unlike most other image generators.

與其他大多數影像生成器不同，它甚至可以處理透明圖層。

But everybody's talking about its ability to transform images into specific art styles.

但大家都在談論它將影像轉化為特定藝術風格的能力。

That's pretty cool, but the significant thing here is that it gives us the ability to render our AI girlfriends in ways that were previously impossible by maintaining character continuity.

這很酷，但重要的是，它讓我們有能力以以前不可能的方式渲染我們的 AI 女友，保持角色的連續性。

Like now I can upgrade her with new poses, new outfits, or even add her to photos of those lonely family dinners.

就像現在，我可以給她換上新姿勢、新衣服，甚至把她加到那些孤獨的家庭聚餐照片中。

That's impressive, but how does it work?

這令人印象深刻，但它是如何運作的呢？

This is ClosedAI we're talking about so there's no paper or code to look at, but apparently, they use an autoregressive approach instead of diffusion.

ChatGPT 很封閉式，所以沒有論文或 code 可查，但顯然他們使用的是自動迴歸法，而不是擴散法。

A diffusion algorithm in models like stable diffusion and mid-journey generate an entire image all at once.

穩定擴散和中程等模型中的擴散算法可以一次性生成整個影像。

But GPT-4o takes a different approach, generating the image pixel by pixel from left to right, top to bottom.

但 GPT-4o 採用了不同的方法，從左到右、從上到下逐個像素生成影像。

It almost doesn't even look artificial, but it actually contains a controversial watermark provided by the coalition for content provenance and authenticity.

它看起來幾乎沒有人工痕跡，但實際上卻包含了聯盟為內容出處和真實性提供的有爭議的水印。

In fact, if you take this image and upload it to the C2PA tool, you can see it was generated by OpenAI along with a history of any modifications to it.

事實上，如果你把這張圖片上傳到 C2PA 工具，就能看到它是由 OpenAI 生成的，同時還能看到對它進行修改的歷史記錄。

Camera manufacturers and software developers like Adobe are implementing this into their software with the goal of tracking every change that happens to a digital asset like an image.

相機制造商和軟體開發商（如 Adobe）正在其軟體中採用這種技術，目的是追蹤影像等數位資產發生的每一次變化。

And that of course is to keep you safe from misinformation at the expense of your privacy and freedom.

這當然是為了保證你的安全，使你免受錯誤訊息的侵害，但卻犧牲了你的隱私和自由。

And platforms like YouTube and Steam are now requiring people to disclose that they use AI assets in their work.

現在，YouTube 和 Steam 等平臺都要求人們披露他們在作品中使用了 AI 資產。

And that brings up a big philosophical question that was first posed by someone with an anime profile picture on the internet.

這就引出了一個重大的哲學問題，這個問題最早是由一個在網上擁有動漫個人照片的人提出的。

With Slop's Razor, you ask, can you tell it's AI generated by looking at it?

利用 Slop's Razor，你問它能通過觀察判斷出它是 AI 生成的嗎？

If the answer is no, then it's indistinguishable from human work, thus no disclosure is needed.

如果答案是否定的，那麼它就與人類工作無異，無需披露。

But if you answer yes, then it's visibly slop. And again, no AI disclosure is needed.

但如果回答是，那麼它顯然就很糟糕。再說一遍，不需要 AI 披露。

C2PA is just deep state bloatware, but I digress.

C2PA 只是深層狀態膨脹軟體，但我離題了。

If you're poor, the other thing you need to know about is Gemini 2.5 Pro.

如果你是窮人，你需要知道的另一個模型就是 Gemini 2.5 Pro。

This new state of the art model from Google actually slaps.

谷歌推出的這一最新技術模型實際上是在打耳光。

Not only is it arguably just as good for programming as Claude 3.7 with a bigger context window, but it's also better than reasoning models like OpenAI o3.

它不僅可以說與具有更大上下文視窗的 Claude 3.7 一樣適合編程，而且比 OpenAI o3 等推理模型更好。

And you can actually use it for free right now instead of paying $200 a month for OpenAI Pro.

實際上，你現在就可以免費使用它，而不用每月支付 200 美元購買 OpenAI Pro。

But the problem for Google is that the Chinese are ruining their plans for world AI dominance.

但谷歌的問題是，中國人正在破壞他們稱霸世界 AI 的計劃。

DeepSeek just released version 3.1 and it too slaps.

DeepSeek 剛剛發佈了 3.1 版本，它也讓人拍案叫絕。

Alibaba released Qwen 2.5 Omni, which can see here talk and write thanks to its new thinker talker architecture.

阿里巴巴發佈了 Qwen 2.5 Omni，由於採用了新的 Thinker Talker 架構，它可以在這裡進行說和寫。

Then Tencent released T1 to compete with DeepSeek.

隨後，騰訊發佈了 T1，與 DeepSeek 展開競爭。

And Bytedance, the company behind TikTok, released DAPO, an open-source reinforcement learning system for building large scale large language models.

TikTok 背後的公司 Bytedance 發佈了開源強化學習系統 DAPO，用於構建大規模的大型語言模型。

We're currently living in a vibe coder's paradise where anyone can pick up an open-source Chinese model,

目前，我們正生活在一個充滿活力的編碼員天堂，任何人都能獲得開源的中國模型，

and generate more code than they could ever possibly need, and that means real programmers are going to have a lot of code to fix and refactor.

並產生比他們可能需要的更多的程式碼，這意味著真正的程式設計師將有大量程式碼需要修復和重構。

And an awesome tool that can help you review all that code is CodeRabbit, the sponsor of today's video.

今天影片的贊助商 CodeRabbit 就是一款能幫你查看所有代碼的超棒工具。

An AI co-pilot for code reviews that gives you instant feedback on every pull request.

用於代碼審查的 AI 輔助駕駛員，可為你提供每個請求的即時反饋。

Unlike basic linters, it understands your entire code base, so it can catch more subtle issues like bad code style or missing test coverage.

與基本襯砌器不同，它能理解整個代碼庫，所以可以捕捉到更多細微的問題，如糟糕的代碼或缺失的測試覆蓋率。

Then it will suggest simple one click fixes to help you get things cleaned up quickly.

然後，它會提出簡單的一鍵修復建議，幫助你快速清理問題。

CodeRabbit keeps learning from your PRs over time, so the more you use it, the smarter it gets.

隨著時間的推移，CodeRabbit 會不斷從你的 PR 中學習，你用得越多，它就越聰明。

It's 100% free for open source projects.

它對開源項目是 100% 免費的。

But you can get one month free for your team using the code fireship with the link below.

但是，使用下面的連結代碼 fireship，你的團隊可以獲得一個月的免費服務。

This has been the Code Report, thanks for watching and I will see you in the next one.

Code Report 到此結束，感謝收看，我們下期再見。