A Chinese lab has apparently unveiled one in every of the primary “logical” AI models that may compete with OpenAI’s o1.
On Wednesday, DeepSeekan AI research company funded by quantitative traders, released a preview of DeepSeek-R1, which the corporate says is an argumentation model competitive with o1.
Unlike most models, argumentation models self-check their facts by spending more time considering a matter or query. This helps them avoid among the pitfalls that typically trip up models.
Similar to o1, DeepSeek-R1 reasons through tasks, forward planning, and performing a series of actions that help the model find a solution. This may take some time. Like o1, DeepSeek-R1 could “think” for tens of seconds before answering, depending on the complexity of the query.
DeepSeek claims that DeepSeek-R1 (or more specifically, DeepSeek-R1-Lite-Preview) is reminiscent of OpenAI's o1 preview model on two popular AI benchmarks, AIME and MATH. AIME uses other AI models to judge a model's performance, while MATH is a set of word problems. But the model is just not perfect. Some commenters on X noted that DeepSeek-R1 fights with tic-tac-toe and others Logic problems (as does o1).
DeepSeek may also be easily jailbroken – that’s, it could possibly be made to disregard security measures. An X user made the model give an in depth description Meth recipe.
And DeepSeek-R1 appears to dam requests which might be deemed too politically sensitive. In our tests, the model refused to reply questions on Chinese leader Xi Jinping, Tiananmen Square, and the geopolitical impact of China's invasion of Taiwan.
The behavior is probably going the results of Chinese government pressure on AI projects within the region. Models in China must undergo Benchmarking by China's Internet regulator to make sure their responses “embody fundamental socialist values.” According to reportsthe federal government even went to this point as to propose a blacklist of sources that can not be used to coach models – and that is the result many Chinese AI systems refuse to answer issues that would raise the ire of regulators.
The increasing attention to argumentation models comes because the viability of “scaling laws” comes under scrutiny. These are long-held theories that providing a model with more data and computing power would constantly increase its capabilities. A excitement of press reports suggest that models from major AI labs like OpenAI, Google, and Anthropic are not any longer improving as dramatically as they once did.
This has led to a race for brand spanking new AI approaches, architectures and development techniques. One of those is the test time calculation, which underlies models akin to o1 and DeepSeek-R1. Test time computation, also called inference computation, essentially gives models additional processing time to finish tasks.
“We are seeing the emergence of a brand new law of scaling,” Microsoft CEO Satya Nadella said this week during a keynote at Microsoft’s Ignite conference, referring to check time calculations.
DeepSeek, which says it plans to open source the DeepSeek-R1 solution and release an API, is a wierd operation. It is backed by High-Flyer Capital Management, a Chinese quantitative hedge fund that uses AI to drive its trading decisions.
One of DeepSeek's first models, a general-purpose text and image evaluation model called DeepSeek-V2, forced competitors like ByteDance, Baidu and Alibaba to lower usage prices for a few of their models – and make others completely free.
High-Flyer is constructing its own server clusters for model training, the youngest of which allegedly features 10,000 Nvidia A100 GPUs and costs 1 billion yen (~$138 million). Founded by Liang Wenfeng, a pc science graduate, High-Flyer goals to realize “superintelligent” AI through its DeepSeek organization.
