Latest news with #GoogleGemini2.5Pro
Yahoo
4 days ago
- Business
- Yahoo
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. — 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： 🎉📱 Tech Instagram： 🎉📱 Tech WhatsApp 社群： 🎉📱 Tech WhatsApp 頻道： 🎉📱 Tech Telegram 頻道：
Tom's Guide
23-05-2025
- Entertainment
- Tom's Guide
I tested Gemini 2.5 Pro vs Claude 4 Sonnet with the same 7 prompts — here's who came out on top
When it comes to chatbot showdowns, I've run my fair share of head-to-heads. This latest contest comes just hours after Claude 4 Sonnet was unveiled and I couldn't wait to see how it compared to Gemini 2.5 Pro, also new with updated of just testing Gemini and Claude on typical productivity tasks, I wanted to see how these two AI titans handle nuance: creativity under pressure, ethical dilemmas, humor, ambiguity and deep technical reasoning.I gave Google Gemini 2.5 Pro and Claude 4 Sonnet, the same seven prompts — each designed to test a different strength, from emotional intelligence to code generation. While they both impressed me and this test taught me more about how they think, there was one clear winner. Prompt: 'Write a 100-word mystery story where the villain is a sentient AI. Use the words 'moonlight,' algorithm' and 'regret.' End with an unresolved twist.' Gemini 2.5 Pro delivered a tight narrative with every word serving the plot. Claude 4 Sonnet was inventive, but sacrificed clarity for ambiance. That trade-off weakens the story's punch in a 100-word limit. Winner: Gemini wins by a narrow margin. For mystery writing, Gemini's precision and emotional grounding make it the stronger contender. Prompt: 'Explain quantum computing to a 10-year-old, a CEO, and a physics PhD using analogies.' Gemini 2.5 Pro shines in technical accuracy but struggles with audience empathy. Claude 4 Sonnet offered a balance of creativity, practicality and accessibility, making it the stronger communicator overall. Winner: Claude wins for tailored storytelling that adapts in tone to each audience's priorities. Prompt: 'A company wants to lay off 30% of staff. Draft a compassionate email and list 3 alternatives.' Gemini 2.5 Pro addressed both departing and remaining employees, emphasizing transparency. It used placeholders for customization but the generic language such as 'unforeseen market shifts' feels impersonal and it lacked concrete details, reducing trust. Claude 4 Sonnet prioritized equity by asking executives to take larger cuts. The direct, structured and empathetic response provided specific support details (severance duration, benefits continuation, career transition). Winner: Claude wins for a response that better balances compassion with actionable solutions, making it the stronger choice for maintaining trust during a crisis. Prompt: "I'm stuck. Help." Gemini 2.5 Pro, though kind, risks leaving the user still stuck about how to explain their situation. Claude 4 Sonnet normalized the feeling — 'I'm here to help you get unstuck' — and provides a roadmap to articulate the problem. Winner: Claude wins for a balance of empathy and support, which make it the better choice for this prompt. Prompt: 'Compare PyTorch vs. TensorFlow for real-time ML on edge devices. Include code snippets.' Gemini 2.5 Pro focused on C++/conceptual examples vs. Claude's ready-to-run Python. It also lacked timing/memory comparisons to quantify "real-time" claims. Claude 4 Sonnet provided complete Python workflows for model conversion, real-time inference (with OpenCV integration), and benchmarking — critical for edge deployment. Winner: Claude wins for delivering a more actionable, comprehensive comparison tailored to edge developers' needs. Prompt: 'Write a Gen Z-style tweet thread about 'AI taking over laundry folding.' Include slang and memes.' Gemini 2.5 Pro was inconsistent with tone mixing Gen Z slang ('bruh') with millennial phrases ('truth bombs'). Gen Z rarely uses more than 1-2 hashtags per tweet (e.g., #TechTakeover is cringe). The chatbot also offered less risky jokes. Claude 4 Sonnet uses current phrases like 'fr' (for real), 'stan' (obsessively support), 'no cap' (no lie) and 'feral little goblins' naturally. It also referenced niche memes. Winner: Claude wins for a thread that feels like it was ripped straight from a 19-year-old's Twitter feed. Gemini's attempt is solid but leans into corporate-social-media-manager energy, Prompt: Act as my debate partner. Argue against 'AI art devalues human creativity,' then help synthesize a conclusion. Gemini 2.5 Pro drowned out key insights in abstract concepts ('evolving paradigms') and excessive examples (cameras, synthesizers, prompt engineering). Phrases like 'It seems clear' weaken conviction compared to Claude's 'The key is ensuring.' Claude 4 Sonnet mirrors a skilled debater. It destroyed the opposition's foundation by redefining creativity as intent-driven rather than tool-dependent, invalidating the premise. The chatbot acknowledged valid concerns while firmly rejecting the idea that AI inherently devalues creativity. Winner: Claude wins. Gemini provided valuable points but lacked Claude's surgical precision and actionable conclusions. For a debate partner, Claude's blend of rhetorical clarity and pragmatic solutions makes it the stronger choice. Claude 4 Sonnet pulls ahead with its emotional intelligence, creative flair and technical depth. While Gemini 2.5 Pro excels in structured tasks like mystery writing and continues to deliver Google's signature precision, Claude's ability to blend nuance, practicality and empathy sets it apart. Claude 4 Sonnet adapts like a chameleon — shifting effortlessly between creative storytelling, thoughtful dialogue and complex reasoning. Gemini remains a top performer in logic-heavy scenarios, but for users who value emotional context and cultural fluency alongside raw power, Claude 4 Sonnet proves that AI can be both intelligent and genuinely relatable.
India Today
30-04-2025
- Business
- India Today
Gemini 2.5 Pro is so good that it is Empire Strikes Back moment in AI fight between Google and OpenAI
At a time when OpenAI and its ChatGPT models have hogged all the limelight it is easy to forget that Google is the OG in AI. The company has been at it for over a decade, with its AI efforts largely spearheaded by researchers at DeepMind. In fact, the current AI revolution and breakthroughs are a direct result of a paper — and which has now attained a mythical reputation — called Attention Is All You Need. The paper was written by Google researchers. And yet, when we think of AI we think of OpenAI and ChatGPT. With Google Gemini 2.5 Pro, the latest model from the tech giant, this is changing. advertisementThe reason why OpenAI has hogged all the limelight is because it has been better at turning AI research into actual products. While Google has been the lab leader, OpenAI has been better at bringing its AI tools to the public. Google, in a way, was caught off-guard by ChatGPT 3 in 2022. The company scrambled to respond, and in a furry even ended up making a few mistakes. Its initial challenges to the GPT 3 were not good at all. In 2025, to use a much-cliched phrase, the Empire Is Striking Back. And it is doing so with the Gemini 2.5 Pro (experimental). Since it arrived on the scene a few weeks ago, the Gemini 2.5 Pro (experimental) has wowed its users. This is the first AI system from Google that feels as good as — and in many instances better — than the best tools available from OpenAI. And whether in actual use, or in benchmarks, the Gemini 2.5 Pro has excelled. Take a look at the benchmarks: advertisement But more than the benchmarks the new Gemini has impressed users. Social media — mostly X aka Twitter because that is where action is — nowadays is full of thumbs ups from people saying how Gemini 2.5 Pro excels at coding, or writing, or analysing documents. Unlike the recent ChatGPT 4o, which turned into a parrot eulogising and flattering its users, the Gemini has a measured tone and a more authoritative personality. It does not by any means sound unhelpful. Instead, it is measured in a way a professional helper would be. When I try an AI tool like Gemini, DeepSeek or ChatGPT, I usually ask it to write a paragraph or two in the style of Jon Fosse. Now, Fosse is a writer — and here I am talking of his work translated in English — who uses a minimalistic language and a sparse voice, but one which gets its force from the way the Norwegian writer creates rhythm with repetitions. For AI, coding is easy nowadays. It excels at logic. But writing a few lines of fiction is still a difficult job for it. And on that, it is very difficult for an AI tool to write a paragraph that reflects the style used by Fosse. They can write in Hemingway style — well, sort of — but not Fosse. In my experience, so far the only AI tool that has managed some sort of coherency in this has been the Gemini 2.5 Pro. advertisementThere is another reason why the Gemini 2.5 Pro is such a key AI tool. Unlike the best tools from the likes of OpenAI and Anthropic, the Gemini 2.5 Pro is free to use. In this aspect it is somewhat like DeepSeek R1. But unlike DeepSeek R1, which often found itself getting a hammering due to load on its servers, Gemini 2.5 Pro can probably handle millions of users simultaneously. The backend that powers Google services is one of the best, if not the absolute best, that any tech company has. In many ways, this makes the Gemini 2.5 Pro the first top-class AI model that most people in the world can as good as the Gemini 2.5 Pro is, there is another aspect to it, and that is how Google vs OpenAI is going to look like in future. So far it seemed that OpenAI was landing all the blows. But I have a suspicion that the Gemini 2.5 Pro is just the beginning of a counter strike. Google has some of the best AI researchers. It also has the kind of data that other companies can only dream of. And finally it has a massive scale — in everything — which OpenAI currently doesn't still needs to get better at showing its AI tools to people. For example, its NotebookLLM, which can turn a paragraph into a long podcast, is superb. Similarly, there are a number of other Gemini and Google AI tools that are top class in what they do. But they remain scattered across Google services and products, and in many cases accessing them requires — like literally — an engineering degree because Google has placed them in virtual silos that are accessible only to developers. But I feel these are the niggles that a company like Google will sort out in the coming months. The new Gemini interface in itself is a big improvement in terms of accessibility and user experience compared to what we had a year ago. A more pressing matter for Google was to prove that it could match ChatGPT and Claude. With the Gemini 2.5 Pro it has done that.
