A new player has entered the AI villa, and it’s creating significant disruption.
Chinese AI startup DeepSeek made waves last week when it released the full version of R1, the company’s open-source reasoning model that can outperform OpenAI’s o1. On Monday, App Store downloads of DeepSeek’s AI assistant — which runs V3, a model DeepSeek released in December — I also tested DeepSeek’s R1 and V3 coding skills – and we’re not all doomed (yet).
I also tested DeepSeek’s R1 and R3 coding abilities – and we’re not all doomed yet.
DeepSeek has already climbed up to the third position overall on HuggingFace’s Chatbot Arenais available under several Gemini models, as well as ChatGPT-4o. DeepSeek started almost as soon as OpenAI was dethroned. Limiting signupsbecause of a alleged cyberattack and then following up with a promising image model.
DeepSeek’s chat page at the time of writing.
Screenshot by Radhika Rajkumar/ZDNET
What is DeepSeek?
Founded by Liang Wenfeng in May 2023 (and thus not even two years old), the Chinese startup has challenged established AI companies with its open-source approach. According to Forbes: DeepSeek may have an edge because it is funded solely by High-Flyer – a hedge fund run by Wenfeng. This funding model supports rapid growth and research.
DeepSeek R1: What is it?
DeepSeek released R1 in full last Thursday. It is the company’s flagship reasoning engine, and it performs at or higher than OpenAI’s acclaimed o1 model across several math, coding and reasoning benchmarks. R1 is unique in that it’s free to download and use, unlike many other top models from tech companies. DeepSeek, however, has not revealed the training dataset for R1. All other models that DeepSeek has released so far are also open-source. DeepSeek costs less than comparable US models
. For example, OpenAI charges $7.50 for an equivalent tier of R1 API access. The price for R1 is $0.14 per million tokens.
Enterprises are hitting a “speed limit” in deploying Gen AI. Here’s why.
DeepSeek’s claims in a company Research paperstates that the V3 model costs $5.6 million in training. This number is circulated as the total development cost for the model. As the AP Some lab experts believe that the paperis referring only to the final training run of V3, and not its entire development costs (which would be just a fraction what tech giants spent to build competing models). Someclaim DeepSeek costs do not include infrastructure, R&D costs, data, or personnel costs.
A drawback of the o1 could be its long-term competitiveness with other US-made models. censorshipChinese models may have blocks on certain subjects, so they may not be able to answer all queries. Here(). Tiernan Ray, a ZDNET employee in December, compared the ability of R1-Lite to explain its thought process to that of o1, with mixed results.
Enterprises are hitting a “speed limit” in deploying Gen AI. Here’s why.
Ofcourse, all popular models have their own red-teaming backgrounds, community guidelines, content guardrails, but at this stage, American made chatbots will not refrain from answering questions about historical events.
Privacy issues
Data security concerns that have been raised about TikTok, the Chinese-owned app that has been banned in the US, are also being addressed. DeepSeek’semergence.
“The personal information we collect from you may be stored on a server located outside of the country where you live,” DeepSeek’s Privacy policy is stated. “We store the information we collect in secure servers located in the People’s Republic of China.”
‘Humanity’s Last Exam’ benchmark is stumping top AI model – can you do better?
According to the policy, DeepSeek collects a lot of information. This includes but is not limited to
-
“IP address, unique device identifiers, and cookies”
-
“date of birth (where applicable), username, email address and/or telephone number, and password”
- “your text or audio input, prompt, uploaded files, feedback, chat history, or other content that you provide to our model and Services”
- “proof of identity or age, feedback or inquiries about your use of the Service,” when you contact DeepSeek.
Further in the policy: “Where we transfer any personal information out of the country where you live, including for one or more of the purposes as set out in this Policy, we will do so in accordance with the requirements of applicable data protection laws.” GDPR compliance is not mentioned. Adrianus Warmenhoven is a member of NordVPN’s security advisory board. He told ZDNET by email that the policy was
“Users need to be aware that any data shared with the platform could be subject to government access under China’s cybersecurity laws, which mandate that companies provide access to data upon request by authorities,” . He added. “This has sparked concerns about potential biases and external influence on the platform’s content moderation policies.”
See also: How to protect privacy on Facebook – and how it doesn’t work.
Some observers believe that the fact that the R1 model is open-source increases transparency by allowing users to examine the source code to look for privacy-related activities. DeepSeek has released smaller versions of R1 that can be downloaded locally and run to avoid any concerns over data being sent to the company. All chatbots including ChatGPT collect some user data when accessed via the browser. What this means for AI in general
R1’s successes highlight a sea-change in AI, which could empower smaller labs to create competitive models, and diversify the available options. OpenAI has a staff and funding, but organizations can download R1 to fine-tune the model so that it can compete with models such as o1. Researchers at UC Berkeley, just before R1 was released, created an open-source version of o1-preview (an early version of o1) in 19 hours for $450.
Many are speculating about the AI bubble bursting due to the exhorbitant investment in AI. Multiple Reports suggest that the stock market has already begun to panic.
DeepSeek is gaining momentum just days after a long-fought TikTok banning was partially implemented. It only takes $450 and 19-hours to rival OpenAI’s o1 preview. Ironically, DeepSeek presents in plain language what the US struggled with to prove regarding TikTok during its long-fought ban.
