DeepSeek pioneers a new way for AI to ‘reason’
Have you tried out DeepSeek? Released on January 20, this new artificial intelligence tool quickly became the world’s most downloaded free app. It held that spot for several days. More than 22 million people were using it to chat, assist with writing or coding and more.
Such bots are tools built on top of a type of AI known as a large language model. New AI bots pop up all the time. Some of the best known — ChatGPT, Grok, Claude and Gemini — can all perform similar tasks.
So what made DeepSeek special? It matches the performance of the best of those. But it needs less energy to do so, explains Vijay Gadepally. A lot less. That makes it “better for the environment per unit of work,” he explains.
Gadepally is a computer scientist at the Massachusetts Institute of Technology’s Lincoln Laboratory in Lexington. There, he works on boosting the efficiency of AI models and the computers that run them. He did not take part in DeepSeek’s development.
AI models have a reputation for being energy hogs. But the team that made the DeepSeek bot found clever innovations to slim down the processes of both developing and running AI models, says Gadepally. This “was very novel,” he says.
That team was part of a company based in Hangzhou, China. It announced the energy-saving approach last December 27 on arXiv. A month later, this company — which is also, confusingly, named DeepSeek — released a new bot. It had been built on top of a slimmer AI model.
People call this new bot DeepSeek. Its technical name, however, is DeepSeek-R1.
It can answer questions like an ordinary chatbot. But it’s actually something new. This new type of bot is known as a reasoning agent.
Less chat and more action
A chatbot answers one question at a time. A reasoning agent doesn’t. It can “take a complex task and break it into smaller pieces,” explains Gadepally.
Your initial question or request goes into the underlying AI model, which comes up with an answer.
But the agent doesn’t stop there. It uses the first answer to ask itself more questions. This process is somewhat similar to what people do when they think over a task.
Researchers call this a chain-of-thought. Depending on the difficulty of the task, this process can take a few seconds or as much as a half hour. During this time, questions and answers are “bouncing around inside the model,” says Gadepally.
Reasoning agents are most useful when paired with search engines, robots and other technology. When this happens, questions and answers bounce in and out of the model, too. The reasoning agent acts sort of like a brain to direct other tools. This entire system can then “act on your behalf and actually perform some activity,” Gadepally explains.
OpenAI, the company behind ChatGPT, has also developed AI models that support this sort of system. Its first one was called o1.
This model (or newer versions of it) support several agents. One agent is called Operator. It has access to a web browser, so it can do such things as book appointments for you. Another agent, called Deep Research, uses search engines. It can gather information and create a custom report.
The newsletter Understanding AI had 19 people try out Deep Research. Seven of them said the tool accomplished work similar to what a professional in their field could do. But OpenAI charges $200 per month for a subscription to this agent. And to use o1 directly, you must pay $60 for every million words of output.
When the DeepSeek-R1 agent came out, it could do the same sorts of tasks as agents built on top of o1. But DeepSeek’s model is far more efficient. It cost just $2.19 for every million words.
This shocked many experts. “I can get a lot more intelligence [for the same amount of computing power],” Jeremie Harris said on his podcast, Last Week in AI. “That is the DeepSeek story.”
“It’s absolutely impressive,” he adds. “This is legit.” (In addition to co-hosting this podcast, Harris runs Gladstone AI, a company based in Washington, D.C.)
Another benefit to DeepSeek: Its maker has shared how it built the reasoning agent and the model that runs it. Techies call this open-source tech. ChatGPT’s developer, OpenAI, has taken a different approach. It’s kept technical details of its bot and platform secret.
“Open-source tools give you the benefit of peeking under the hood and seeing what’s going on,” says Avijit Ghosh. Based in Boston, Mass., he’s an expert in AI ethics at the company Hugging Face. This company runs a platform for open-source AI development.
Anyone can now build their own version of DeepSeek. Some have already done so. The AI company Perplexity launched a free Deep Research tool in February. It’s built using a modified version of DeepSeek-R1.
.cheat-sheet-cta
border: 1px solid #ffffff;
margin-top: 20px;
background-image: url(“https://www.snexplores.org/wp-content/uploads/2022/12/cta-module@2x-2048×239-1.png”);
padding: 10px;
clear: both;
Do you have a science question? We can help!
Submit your question here, and we might answer it an upcoming issue of Science News Explores
The future of ‘thinking’ AI
AI companies often describe their reasoning agents as “thinking” machines. That can be misleading. These agents do perform impressive feats. But they don’t work like our brains do. They are built on the same sorts of AI models as chatbots. That means they can suffer some of the same drawbacks.
They may produce biased content, for instance. Or people can jailbreak them to get around safety guardrails. And these bots may sometimes confidently provide wrong information, a problem called hallucinating.
When an agent uses a model to perform a complex task on its own, a bunch of small problems can also snowball into bigger ones, notes Ghosh. And that worries him.
He co-authored a paper on arXiv in February that sums it up: “Fully autonomous AI agents should not be developed.” Being “fully autonomous,” explains Ghosh, means being able to act entirely on its own.
Science News Explores reached out to DeepSeek for comment but the company didn’t respond.
DeepSeek-R1 and similar bots aren’t fully autonomous yet. They can do more complex tasks than chatbots. Yet a real person still guides the process. And Ghosh thinks that’s essential. “I don’t think [these systems] are nearly as reliable and as accountable as humans,” he says.
But things are moving quickly. In late February, the robot company Figure AI announced a new AI model called Helix. Robots that use this model are better at reasoning.
In early March, a different Chinese company, named Monica, announced a brand new agent named Manus. It’s built atop a handful of different AI models. It’s also close to fully autonomous. When you give it a task, it decides on its own how it should solve it. Then it starts taking actions, such as searching online, writing code, creating charts and more. This all happens (in the cloud) while you wait. Manus then alerts you when its task is done.
And consider this: Companies like DeepSeek, Monica and OpenAI can now use their agents to come up with ideas on how to build better reasoning models.
Harris, of Last Week in AI, is confident that this is already happening. How well they will work remains to be seen. Clearly, it’s an exciting time for AI innovation.
This is one in a series presenting news on technology and innovation, made possible with generous support from the Lemelson Foundation.
