In this issue of the #InfocusAI digest, we’ll discuss the successes of Chinese developers from DeepSeek and Alibaba, the failure of a French LLM, share the challenges of recognizing people from a kilometer high, and take a look at the future of sustainable agriculture. Finally, we’ll cover OpenAI’s advancements in longevity research.
AI-focused digest – News from the AI world
Issues No. 58, 17 — 30 January 2025
A Boom in Chinese LLMs
They are feared, hated, but widely used—this is all about Chinese LLMs, which have literally exploded onto the market over the past two weeks. First, developers from the Middle Kingdom released the DeepSeek V3 model (generates text and solves creative tasks) and DeepSeek R1 (stands out for its ability to reason logically and solve complex problems). Then Alibaba joined the race. The company gifted users Qwen 2.5-Max for the Chinese New Year, capable of processing text and visual data, including videos longer than 1 hour. The startup DeepSeek once again didn’t stay on the sidelines and released the Janus-Pro update—a multimodal model for image generation. According to the developers, it outperforms DALL-E 3 and Stable Diffusion 3 Medium in quality. We’re waiting to see how competitors will respond and how much more NVIDIA’s stock will lose.
France Releases the Worst LLM in History
Against the backdrop of the success of Chinese developers, the results of their European counterparts look more than modest, and sometimes even questionable. In France, users ridiculed the LLM Lucie, calling it “the worst of all existing ones.” Lucie was created by the company Linagora. The French analogue of ChatGPT was launched on January 23 for a trial period that was supposed to last one month. However, users quickly started sharing screenshots of absurd, and sometimes even illegal, responses from the model. “Lucie can’t solve a simple math problem, claims that cow eggs exist, denies women’s rights, and provides a detailed recipe for banned substances,” write the media. As a result, Lucie’s work was suspended, and a lengthy explanatory and apologetic note appeared on the project’s website.
Americans Train CV Algorithms to Recognize People from a Kilometer High
Researchers from Oak Ridge National Laboratory published a paper on biometric recognition of people from drones and cameras at distances of up to a kilometer. They are working to create systems capable of identifying a person from a height, in bad weather, or through atmospheric distortions. Oak Ridge tested five algorithms combining data on faces, bodies, and gaits, identifying key challenges for recognition. These turned out to be low resolution (when a person’s head occupies less than 30 pixels) and distances over 550 meters. For example, at a distance of 1000 meters, the system makes mistakes 10,000 times more often. Interestingly, wind, sun, and temperature have almost no effect on the results. The authors attribute this to the algorithms “catching” good frames from video, ignoring interference. Notably, Oak Ridge National Laboratory researchers are backed by IARPA—a structure linked to U.S. intelligence.
Scientists Develop an AI-Powered Tractor to Reduce Soil Erosion
Let’s move on to more peaceful matters. A group of researchers led by Sajiv Magesh from the Dublin Institute of Advanced Studies (USA) has created an autonomous tractor and an AI system that optimizes plowing and fertilizer application. The results were published in the journal Nature and have already been called a “revolution in sustainable agriculture worldwide.” As part of this project, the scientists considered issues of reducing soil erosion, decreasing CO₂ emissions, and overuse of fertilizers. The researchers combined a convolutional neural network (CNN) that analyzes field photos and determines plowing intensity, with an algorithm that takes into account 10 parameters: soil data (e.g., moisture, temperature, slope, nitrogen, phosphorus, potassium levels), weather forecasts, crop type, and soil type. They then developed an autonomous tractor equipped with cameras, sensors, and modules for plowing and fertilizing. The system is being tested on farms in the USA, India, Belgium, the Netherlands, and France. According to preliminary estimates, the AI-powered tractor can optimize soil treatment intensity, significantly increasing yield, preventing topsoil erosion, and reducing carbon emissions by 57%.
OpenAI Creates a Model for Longevity Research
OpenAI researchers have developed a language model for designing proteins that can turn ordinary cells into stem cells. The GPT-4b micro model was trained on biological data. It can suggest ways to modify Yamanaka factors—proteins that reprogram skin cells into stem cells. During joint experiments with the California-based biotech startup Retro Biosciences, it was found that the modified proteins increase the efficiency of the cell transformation process by more than 50 times. Moreover, the AI-designed proteins turned out to be of higher quality than those created by scientists themselves. This is OpenAI’s first attempt to use artificial intelligence for biomedical research. However, the final results of the experiments have not yet been published, and the model is not available to the general public.
