Researchers are swarming to test out the latest artificial intelligence (AI) tools, which are published virtually every week and seem to be more remarkable. More generative AI tools are available to researchers than ever before, whether they are trying to build code, edit articles, or generate hypotheses.

DeepSeek-R1, which was released last month, is comparable to O1 in terms of capabilities but is accessible via an API for a far lower price.

Additionally, it differs from OpenAI's models in that it is open weight, which means that anybody can obtain the underlying model and modify it for their own research project even while the training data has not been made public.

R1 Advanced Model

R1 has "just unlocked a new paradigm" that allows communities to create specialized reasoning models, especially ones with limited resources, according to White.

Many academics lack the powerful processing hardware needed to run the entire model. However, researchers are developing versions that can operate or train on a single system, including Benyou Wang, a computer scientist at the Chinese University of Hong Kong, Shenzhen.

DeepSeek Mathematical Powers

Math difficulties and code programming are DeepSeek-R1's strong points. However, White claims that it is also adept at jobs like coming up with hypotheses.

According to him, this is because DeepSeek has chosen to fully expose the model's "thought processes," which enables researchers to better hone their follow-up inquiries and, eventually, enhance its outputs. Medical diagnostics could benefit greatly from this kind of transparency as well.

R1 is being modified by Wang in trials that employ the reasoning-like capabilities of the model to create "a clear and logical pathway from patient assessment to diagnosis and treatment recommendation," according to Wang.

DeepSeek-R1 has certain drawbacks. The model appears to have an unusually lengthy "thought" process, which slows it down and reduces its utility for brainstorming or information searching. Several governments have prohibited employees of national agencies from utilizing the chatbot due to concerns regarding the security of data entered into its API.

Challenge to existing technology

Additionally, compared to its commercial rivals, DeepSeek appears to have taken fewer steps to reduce the negative outputs from its models. It requires time and effort to add filters to stop such outputs, including instructions to produce weapons. "The lack of guardrails is worrisome," Simon argues, even if it is unlikely that this was done intentionally.

Also, according to OpenAI, DeepSeek might have "inappropriately distilled" its models—a reference to a technique that involves training a model using the outputs of another algorithm, which is forbidden by OpenAI's terms of use. Before this story was published, DeepSeek could not be contacted for comment on these criticisms.

While some researchers are comfortable with R1 and consider such distillation to be standard practice, others are cautious about utilizing a technology that may be the target of future legal action.

If utilizing the model was seen to be against the journal's ethical guidelines, scientists using R1 might be compelled to withdraw their publications, according to Ana Catarina De Alencar, an AI law specialist at EIT Manufacturing in Paris.

The usage of models by OpenAI and other companies accused of violating intellectual property rights may be in a similar predicament, according to De Alencar. News outlets say the companies trained their models on journalistic material without authorization.