DeepSeek or ChatGPT: Which to Choose?

How do these two chatty AI companions compare so far?

Idiom of the week: 更胜一筹 (gèng shèng yī chóu)

Meaning: to be superior, to have an edge over

结果显示，DeepSeek在数学问题解决方面表现优异，而ChatGPT-4o 则在一般推理能力上更胜一筹。The results showed that DeepSeek excelled in solving mathematical problems, while ChatGPT-4o outperformed in general reasoning abilities.

February was packed with news: not only did the Trump presidency begin, but China also launched a brand-new AI software called DeepSeek! Similar to its American counterpart, ChatGPT, DeepSeek fulfills a comparable function—though it conveniently avoids topics like the Tiananmen Massacre, steering the conversation elsewhere.

How do they compare today? Let’s take a look at what Taiwanese media have to say on the matter (the original article was written in traditional Chinese). Get ready for a dose of AI-enhanced, nerdy vocabulary!

Photo: apnews.com

在美国总统川普正式上任的同一天，中国杭州深度求索人工智能基础技术研究有限公司推出了一款全新的AI服务软件「DeepSeek」，采用开源模式（Open Source），声称仅以 ChatGPT开发成本的二十分之一，就打造出媲美ChatGPT 的语言模型。

• 上任 (shàngrèn) – to take office, to assume a post

• 杭州 (hángzhōu) – Hangzhou (a city in China)

• 深度 (shēndù) – deep, depth

• 求索 (qiúsuǒ) – to seek, to explore

• 人工智能 (réngōng zhìnéng) – artificial intelligence (AI)

• 基础 (jīchǔ) – foundation, basic

• 有限公司 (yǒuxiàn gōngsī) – limited company (Ltd.)

• 软件 (ruǎnjiàn) – software

• 采用 (cǎiyòng) – to adopt, to use

• 开源 (kāiyuán) – open-source

• 仅 (jǐn) – only, merely

• 成本 (chéngběn) – cost, expense

• 打造 (dǎzào) – to create, to build

• 媲美 (pìměi) – to be comparable to, to rival

• 语言模型 (yǔyán móxíng) – language model

On the same day that U.S. President Trump officially took office, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., in China launched a brand-new AI service software called "DeepSeek." It adopted an open-source model and claimed to have developed a language model comparable to ChatGPT at only one-twentieth of ChatGPT’s development cost.

DeepSeek的技术架构基于混合专家模型（Mixture of Experts, MoE），总计拥有6,710亿个参数，但每次回答问题时仅会启动370亿个参数，从而优化计算资源，并提升特定任务的处理效率。相比之下，ChatGPT采用GPT架构，每次运行会启动所有1,750亿个参数，提供更全面的运算，但计算成本较高。

• 技术架构 (jìshù jiàgòu) – technical architecture

• 混合 (hùnhé) – mixed, hybrid

• 总计 (zǒngjì) – total, overall

• 参数 (cānshù) – parameter

• 启动 (qǐdòng) – to activate, to start

• 计算资源 (jìsuàn zīyuán) – computing resources

• 特定 (tèdìng) – specific, designated

• 处理效率 (chǔlǐ xiàolǜ) – processing efficiency

• 运行 (yùnxíng) – to operate, to run

• 运算 (yùnsuàn) – computation, calculation

DeepSeek's technical architecture is based on the Mixture of Experts (MoE) model, with a total of 671 billion parameters. However, only 37 billion parameters are activated for each response, optimizing computational resources and improving efficiency for specific tasks. In comparison, ChatGPT uses the GPT architecture, activating all 175 billion parameters for each operation. While this provides more comprehensive processing, it comes at a higher computational cost.

在多项评测报告中，全球科技专家深入比较DeepSeek与ChatGPT-4o。结果显示，DeepSeek在数学问题解决方面表现优异，而ChatGPT-4o 则在一般推理能力上更胜一筹。此外，DeepSeek R1具备更高的成本效益，而ChatGPT-4o则提供更强的多功能性。

• 评测 (píngcè) – evaluation, assessment

• 深入 (shēnrù) – in-depth, deeply

• 优异 (yōuyì) – excellent, outstanding

• 推理 (tuīlǐ) – reasoning, logical inference

• 更胜一筹 (gèng shèng yī chóu) – to be superior, to have an edge over

• 成本效益 (chéngběn xiàoyì) – cost-effectiveness

• 多功能性 (duō gōngnéng xìng) – versatility, multifunctionality

In multiple evaluation reports, global tech experts conducted an in-depth comparison between DeepSeek and ChatGPT-4o. The results showed that DeepSeek excelled in solving mathematical problems, while ChatGPT-4o outperformed in general reasoning abilities.

Additionally, DeepSeek R1 offers higher cost efficiency, whereas ChatGPT-4o provides greater versatility.

Photo: BBC

DeepSeek的训练成本较低，AI应用更加普及；然而，与ChatGPT相似，DeepSeek仍面临隐私与道德的挑战。当全球科技界从最初的惊讶中回过神来，随之而来的疑问是：DeepSeek为何能如此迅速崛起？

• 训练 (xùnliàn) – training

• 应用 (yìngyòng) – application, usage

• 普及 (pǔjí) – widespread, popularized

• 面临 (miànlín) – to face, to encounter

• 隐私 (yǐnsī) – privacy

• 科技界 (kējì jiè) – tech industry, technology sector

• 回过神来 (huí guò shén lái) – to regain composure, to snap out of it, to move past

• 随之而来 (suí zhī ér lái) – to follow, to ensue

• 疑问 (yíwèn) – question, doubt

DeepSeek's lower training costs have made AI applications more accessible. However, like ChatGPT, DeepSeek still faces challenges related to privacy and ethics. As the global tech community moves past its initial surprise, the pressing question that follows is: how did DeepSeek rise to prominence so quickly?

根据英国《金融时报》的报导，OpenAI已经发现了一些迹象，显示DeepSeek可能透过「蒸馏技术（Distillation）」间接获取其数据。

「蒸馏」是一种AI训练技术，透过引用更大、更先进的AI模型输出，训练较小的模型，使其能在特定任务上以更低的成本达到类似的效果。简而言之，就是在OpenAI富养的鱼池中钓鱼，DeepSeek以较廉价的鱼饵钓到了锦鲤。

• 金融时报 (jīnróng shíbào) – Financial Times

• 迹象 (jīxiàng) – sign, indication

• 蒸馏 (zhēngliú) – distillation

• 间接 (jiànjiē) – indirect, indirectly

• 获取 (huòqǔ) – to obtain, to acquire

• 数据 (shùjù) – data

• 引用 (yǐnyòng) – to cite, to reference

• 输出 (shūchū) – output

• 使其 (shǐ qí) – to enable it, to make it

• 简而言之 (jiǎn ér yán zhī) – simply put, in short

• 富养 (fù yǎng) – well-nourished, well-stocked

• 廉价 (liánjià) – low-cost, cheap

• 鱼饵 (yú'ěr) – bait

• 锦鲤 (jǐnlǐ) – koi fish, lucky catch

According to a report by the Financial Times (UK), OpenAI has detected signs suggesting that DeepSeek may have indirectly acquired its data through a technique known as "distillation."

"Distillation" is an AI training technique that involves using the outputs of a larger, more advanced AI model to train a smaller model, enabling it to achieve similar performance on specific tasks at a lower cost. Simply put, it’s like fishing in OpenAI’s well-stocked pond—DeepSeek managed to catch a prized koi with a much cheaper bait.

透过蒸馏技术以降低模型训练成本或许合理，但DeepSeek的投资成本真的只有ChatGPT的二十分之一吗？ 《日本经济新闻》指出，DeepSeek使用的H800晶片价格，相较H100低约一至三成，并非极端便宜。此外，Google DeepMind执行长Demis Hassabis也质疑DeepSeek宣称的低成本，认为其报告的560万美元成本，可能仅涵盖最终训练阶段，这只是总体开发成本的一小部分，因此这一数字可能被「刻意压低」。

• 合理 (hélǐ) – reasonable, rational

• 晶片 (jīngpiàn) – chip, semiconductor

• 相较 (xiāngjiào) – compared to, in comparison

• 约 (yuē) – approximately, about

• 并非 (bìng fēi) – not really, not necessarily

• 极端 (jíduān) – extreme

• 执行长 (zhíxíngzhǎng) – CEO, executive director

• 质疑 (zhíyí) – to question, to doubt

• 宣称 (xuānchēng) – to claim, to assert

• 涵盖 (hángài) – to cover, to include

• 总体 (zǒngtǐ) – overall, total

• 开发 (kāifā) – development

• 刻意 (kèyì) – deliberately, intentionally

• 压低 (yādī) – to understate, to lower deliberately

Reducing model training costs through distillation may seem reasonable, but is DeepSeek's investment really just one-twentieth of ChatGPT’s?

According to Nikkei Asia, the H800 chips used by DeepSeek are about 10% to 30% cheaper than the H100, but not drastically low-cost. Additionally, Google DeepMind CEO Demis Hassabis has questioned DeepSeek’s claimed low expenses, suggesting that the reported $5.6 million cost likely only covers the final training phase—just a small fraction of the total development cost. This figure, he argues, may have been deliberately understated.

Did you try DeepSeek so far?

Antoine & Dorota

Referenced article: udn.com