DeepSeek 再被懷疑用 Google Gemini 訓練新版 R1 模型
DeepSeek 以低成本訓練出足夠強效的推理 AI 模型，曾經震驚業界，甚至是政界。DeepSeek 最新推出的 R1-0528 模型主打更強數理和編程表現，不過他們的訓練數據卻未曾公開，AI 業界又再一次懷疑 DeepSeek 是透過蒸餾其他 AI 模型而開發新版本。
其中一個支持這論點的是澳洲開發者 Sam Paech，他在 X 上發文指出R1-0528 模型的語言風格與 Google Gemini 2.5 Pro 極為相似。他認為 DeepSeek 已經從以往基於 OpenAI 的數據切換至 Gemini 的合成數據。另一位開發者 SpeechMap 則發現，R1 模型生成的'推理痕跡'（AI 在得出結論時的思維過程）也與 Gemini 模型極為相似。
If you're wondering why new deepseek r1 sounds a bit different, I think they probably switched from training on synthetic openai to synthetic gemini outputs. pic.twitter.com/Oex9roapNv
— Sam Paech (@sam_paech) May 29, 2025
另一邊廂非牟利 AI 研究機構 AI2 的 AI 專家 Nathan Lambert 更發文指 DeepSeek 在缺乏 GPU 和鉅額資金的支持下，也一定會透過市場最佳的模型 API 來蒸餾數據，這次就是 Gemini。
2024 年時，OpenAI 透過金融時報發聲，指他們獲得證據指 DeepSeek V3 是透過蒸餾 ChatGPT 的數據來訓練而成，後來 Bloomberg 也報道指主要金主 Microsoft 偵測到在 2024 年年底，有大量資料經過 OpenAI 開發者帳戶外洩，他們相信是與 DeepSeek 有關。
為防止競爭對手利用其模型數據，AI 公司正加強安全措施。例如，OpenAI 現在要求用戶完成身份驗證才能訪問高級模型，而 Google 則開始對 Gemini 模型生成的'推理痕跡'進行摘要處理，讓競爭對手更難以利用其數據。
更多內容：
DeepSeek may have used Google's Gemini to train its latest model
DeepSeek 懶人包｜中國AI新創如何影響美國AI巨企？一文整理歷史、最新影響及未來
中國 DeepSeek AI 模型自稱 GPT-4，「AI 天材」是抄襲還是幻想？
DeepSeek 反客為主！連百度搜尋都已確定引入
緊貼最新科技資訊、網購優惠，追隨 Yahoo Tech 各大社交平台！
🎉📱 Tech Facebook：https://www.facebook.com/yahootechhk
🎉📱 Tech Instagram：https://www.instagram.com/yahootechhk/
🎉📱 Tech WhatsApp 社群：https://chat.whatsapp.com/Dg3fiiyYf3yG2mgts4Mii8
🎉📱 Tech WhatsApp 頻道：https://whatsapp.com/channel/0029Va91dmR545urVCpQwq2D
🎉📱 Tech Telegram 頻道：https://t.me/yahootechhk
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
CNBC
2 hours ago
- CNBC
As nations build 'sovereign AI,' open-source models and cloud computing can help, experts say
As artificial intelligence becomes more democratized, it is important for emerging economies to build their own "sovereign AI," panelists told CNBC's East Tech West conference in Bangkok, Thailand, on Friday. In general, sovereign AI refers to a nation's ability to control its own AI technologies, data and related infrastructure, ensuring strategic autonomy while meeting its unique priorities and security needs. However, this sovereignty has been lacking, according to panelist Kasima Tharnpipitchai, head of AI strategy at SCB 10X, the technology investment arm of Thailand-based SCBX Group. He noted that many of the world's most prominent large language models, operated by companies such as Anthropic and OpenAI, are based on the English language. "The way you think, the way you interact with the world, the way you are when you speak another language can be very different," Tharnpipitchai said. It is, therefore, important for countries to take ownership of their AI systems, developing technology for specific languages, cultures, and countries, rather than just translating over English-based models. Panelists agreed that the digitally savvy ASEAN region, with a total population of nearly 700 million people, is particularly well positioned to build its sovereign AI. People under the age of 35 make up around 61% of the population, and about 125,000 new users gain access to the internet daily. Given this context, Jeff Johnson, managing director of ASEAN at Amazon Web Services, said, "I think it's really important, and we're really focused on how we can really democratize access to cloud and AI." According to panelists, one key way that countries can build up their sovereign AI environments is through the use of open-source AI models. "There is plenty of amazing talent here in Southeast Asia and in Thailand, especially. To have that captured in a way that isn't publicly accessible or ecosystem developing would feel like a shame," said SCB 10X's Tharnpipitchai. Doing open-source is a way to create a "collective energy" to help Thailand better compete in AI and push sovereignty in a way that is beneficial for the entire country, he added. Open-source generally refers to software in which the source code is made freely available, allowing anyone to view, modify and redistribute it. LLM players, such as China's DeepSeek and Meta's Llama, advertise their models as open-source, albeit with some restrictions. The emergence of more open-source models offers companies and governments more options compared to relying on a few closed models, according to Cecily Ng, vice president and general manager of ASEAN & Greater China at software vendor Databricks. AI experts have previously told CNBC that open-source AI has helped China boost AI adoption, better develop its AI ecosystem and compete with the U.S. Prem Pavan, vice president and general manager of Southeast Asia and Korea at Red Hat, said that the localization of AI had been focused on language until recently. Having sovereign access to AI models powered by local hardware and computing is more important today, he added. Panelists said that for emerging countries like Thailand, AI localization can be offered by cloud computing companies with domestic operations. These include global hyperscalers such as AWS, Microsoft Azure and Tencent Cloud, and sovereign players like AIS Cloud and True IDC. "We're here in Thailand and across Southeast Asia to support all industries, all businesses of all shapes and sizes, from the smallest startup to the largest enterprise," said AWS's Johnson. He added that the economic model of the company's cloud services makes it easy to "pay for what you use," thus lowering the barriers to entry and making it very easy to build models and applications. In April, the U.N. Trade and Development Agency said in a report that AI was projected to reach $4.8 trillion in market value by 2033. However, it warned that the technology's benefits remain highly concentrated, with nations at risk of lagging behind. Among UNCTAD's recommendations to the international community for driving inclusive growth was shared AI infrastructure, the use of open-source AI models and initiatives to share AI knowledge and resources.
Android Authority
3 hours ago
- Android Authority
Google Wallet gets a fresh coat of Material 3 Expressive with latest Play System update
Edgar Cervantes / Android Authority TL;DR Google Wallet is getting a new design that adopts Google's Material 3 Expressive aesthetic. The update is now rolling out widely, and Google has also made it official in its latest Play System update. With the update, Google Wallet has a more modern layout, with new icons, rounded rectangle containers, fresh buttons, and more. Google is officially rolling out a new Google Wallet experience with the latest Play System update (version 25.25). The redesign brings elements from Google's Material 3 Expressive design language to Google Wallet, freshening up its look with a more modern aesthetic. Some folks spotted Google Wallet's Material 3 Expressive makeover last week, but Google is now rolling it out more widely on version 25.24.772650276 of the app. With the update, Google Wallet now has new icons and rounded rectangle containers that are a trademark of the Material 3 Expressive look. You'll also notice a new Google Wallet logo in the top left corner of the app instead of the 'Wallet' text. Buttons are also getting a bit of an update, and you'll now notice a more minimalistic '+' FAB (floating action button) instead of the 'Add to Wallet' button. Nothing much is changing in terms of functionality, so you don't have to relearn how to use the app. The new design may make it more intuitive and pleasurable to use. Google Wallet isn't the only app the company is updating with the new design. We recently spotted Material 3 Expressive changes in Gmail, while Google's Phone app also just received its expressive update. In fact, all Google apps are due to get an expressive refresh sooner or later.
CNBC
3 hours ago
- CNBC
Apple weighs using Anthropic or OpenAI to power Siri in major reversal, Bloomberg News reports
Apple is weighing using artificial intelligence technology from Anthropic or OpenAI to power a new version of Siri, instead of its own in-house models, Bloomberg News reported on Monday. Shares of the iPhone maker, which had traded down earlier in the session, closed 2% higher on Monday. Apple has had discussions with both companies about using their large language models for Siri, asking them to train versions of their LLMs that could run on Apple's cloud infrastructure for testing, the report said, citing people familiar with the discussions. Apple's investigation into third-party models is at an early stage and the company has not made a final decision on using them, the report said. Amazon-backed Anthropic declined to comment, while Apple and OpenAI did not respond to Reuters requests. The company had in March said AI improvements to its voice assistant Siri will be delayed until 2026, without giving a reason for the setback. Apple shook up its executive ranks to get its AI efforts back on track after months of delays, resulting in Mike Rockwell taking charge of Siri, as CEO Tim Cook lost confidence in AI head John Giannandrea's ability to execute on product development, Bloomberg had reported in March. At its annual Worldwide Developers Conference earlier this month, Apple focused more on incremental developments that improve everyday life — including live translations for phone calls — rather than the sweeping ambitions for AI that Apple's rivals are capitalizing. Apple software chief Craig Federighi had then said it is opening up the foundational AI model that the iPhone maker uses for some of its own features to third-party developers, and that the company will offer both its own and OpenAI's code completion tools in its key Apple developer software.
