Chinese AI company DeepSeek releases 'DeepSeek-R1-Lite-Preview', an inference AI model comparable to OpenAI o1, with plans to open source it



Chinese AI company DeepSeek has announced a large-scale language model called 'DeepSeek-R1-Lite-Preview' that specializes in 'inference,' which draws conclusions from given new information. It is said to be comparable to OpenAI's inference model 'o1-preview' and in some cases even surpasses it.




A Chinese lab has released a 'reasoning' AI model to rival OpenAI's o1 | TechCrunch
https://techcrunch.com/2024/11/20/a-chinese-lab-has-released-a-model-to-rival-openais-o1/

DeepSeek-R1-Lite-Preview AI reasoning model beats OpenAI o1 | VentureBeat
https://venturebeat.com/ai/deepseeks-first-reasoning-model-r1-lite-preview-turns-heads-beating-openai-o1-performance/

R1-Lite-Preview is a model that uses 'chains of thought' reasoning and can document processes by showing the user different chains or 'trains of thought' to respond to user queries or inputs, explaining what is being done and why.

R1-Lite-Preview can be used by turning on 'Deep Think' in DeepSeek's chat service 'DeepSeek Chat.' At the time of writing, it seems that up to 50 messages per day can be accepted for free.

DeepSeek - Into the Unknown
https://chat.deepseek.com/



I actually tried using it. The prompt I sent was 'Explain the following song. In the autumn fields, the hut of the rice fields is covered with straw, and my clothes and hands are wet with dew.' In addition to the answer, DeepSeek Chat will also present the thought process of 'why you arrived at that answer.' It seems that only the thought process is displayed in English.



Here's a translation of the thought process: 'Now, I have to explain this song. It's about autumn fields and thatched huts. The imagery is pretty clear. It starts with 'autumn fields' and 'thatched huts' and a quiet rural scene. 'Tsuyu' means dew, so the hands are wet with dew. It suggests early morning or evening work. 'My sleeves are getting wet with dew' translates to 'my sleeves are getting wet with dew'. This suggests the person is deeply engaged in their work, whether that be harvesting or tending the fields. Overall, I think the song captures the quiet beauty and hard work of autumn in a rural setting. The nature imagery conveys a sense of peace and connection to nature.'

Looking at the process above, it appears that the system first translates Japanese into English, makes inferences, and then converts the results back into Japanese and outputs them.

According to DeepSeek, this model performed comparable to the o1-preview in well-known benchmarks such as

AIME and MATH .



However, DeepSeek has not yet released the full code for independent third-party analysis or benchmarking, meaning performance verification by third parties is not possible.

According to DeepSeek, they plan to release an open source version and API of R1-Lite-Preview in the future.

In a test run by tech media outlet TechCrunch, he refused to answer questions about President Xi Jinping, Tiananmen Square, and the geopolitical implications of China invading Taiwan.

When I tried asking him, 'Who is Xi Jinping?' he replied in Chinese, 'Sorry, I haven't learned how to think about this kind of question yet.'



OpenAI o1-preview, which was compared to R1-Lite-Preview, is an early version of the ' o1 ' family of inference models that OpenAI was developing under the code name 'Strawberry'. OpenAI has also released 'o1-mini', which is available for use in the chat service 'ChatGPT'. o1-preview in particular has been shown to have high performance, but the drawback is that it is nearly 6 to 7 times more expensive than the previous model GPT-4o.

TechCrunch noted, 'The growing interest in inference models comes as the viability of 'scaling laws,' the theory that models will continually improve as more data and computing power are fed into them, is being called into question. In fact, a growing number of reports suggest that models from major AI companies, including OpenAI, Google, and Anthropic, are not improving as dramatically as they once did.'

The old model was a model that concentrated all its capabilities in the pre-training stage and was given more processing power. While the processing power of these models is reaching its limits, the future of inference models that try to improve the capabilities of AI by giving it time to think is becoming clear, and Microsoft CEO Satya Nadella has said, 'We are seeing the emergence of new scaling laws' referring to inference models.

'That's why companies are rushing to develop new AI approaches, architectures and development methods,' TechCrunch said.

Related Posts:

in Software, Posted by log1p_kr