Deepseek's latest V3 update has showcased its remarkable capabilities in coding tasks. The model was able to create a modern landing page with HTML, CSS, and JavaScript in a single file, generating almost 20,000 tokens to achieve this effect. This is a significant improvement compared to the output from Sonnet 3.7.

Furthermore, Deepseek demonstrated its ability to generate an HTML script for a ball bouncing inside a spinning tessellated cube. The model not only accurately rendered the visual interaction but also added the functionality to highlight the specific side of the cube that the ball collides with.

Deepseek's reasoning capabilities have also been enhanced. When presented with modified versions of classic problems, such as the trolley problem and the Monty Hall problem, the model was able to recognize the changes and provide appropriate responses. This suggests that Deepseek has developed a better understanding of the nuances in these types of problems.

Overall, the impressive coding and reasoning capabilities showcased in Deepseek's V3 update highlight the significant progress made in large language models. These advancements suggest that Deepseek's future iterations, such as R2 and V4, could potentially offer even more remarkable performance in a wide range of tasks.

Deepseek V3 has demonstrated impressive reasoning capabilities, particularly in handling modified versions of classic problems. The model is able to carefully read the user input and rewrite it verbatim before providing a thoughtful answer.

For example, when presented with a modified version of the trolley problem where the people on the track are already dead, Deepseek recognized the difference from the classic scenario and refused to pull the lever, explaining that the people were already deceased. Similarly, for the dead cat problem, Deepseek correctly identified that the cat was already dead when placed in the box, and that no subsequent events could change its state.

However, the model still struggles with certain reasoning tasks, such as the modified version of the Monty Hall problem. In this case, Deepseek got confused and provided steps that were not necessary to solve the problem.

Overall, Deepseek V3 shows a notable improvement in its reasoning capabilities compared to the original V3 model. The model's ability to carefully read and comprehend the user's input, and then provide a thoughtful and tailored response, is a significant advancement. While there are still some areas for improvement, Deepseek's reasoning skills are a promising step forward in the development of more capable and reliable language models.

Deepseek, a powerful language model, has recently received a minor update to version 3, which is considered one of the biggest updates to an LLM. The model has demonstrated impressive capabilities, particularly in coding-related tasks.

To access and use Deepseek, there are several options:

1. Deepseek Website: The updated Deepseek model is available on the official Deepseek website. Users can access the model through the website's app or mini-program, with the "Deep Seek Thinking Feature" disabled.

2. Hugging Face: The model weights for Deepseek V3 are available on the Hugging Face platform, but they require a significant amount of storage (around 700 GB) and GPU capacity to run locally.

3. Open Router API: An alternative option is to use the Open Router API, which provides access to a model named "DeepSQ3 0324". This API offers a flexible token limit, with the ability to generate up to 231,000 tokens, making it particularly suitable for tasks like software development. The Open Router API is currently available for free.

It's important to note that while the Open Router API may provide a more accessible option, the performance and capabilities of the model may differ slightly from the official Deepseek version hosted on their website. Users are advised to experiment with both options and compare the results to determine the best fit for their needs.

Overall, the updated Deepseek V3 model has demonstrated impressive capabilities, particularly in coding-related tasks, and offers multiple avenues for users to access and utilize its powerful features.

The updated DeepSeek V3 model appears to be a significant improvement over the previous version, showcasing impressive capabilities in various tasks, particularly in the realm of coding and problem-solving. The model's ability to generate complex HTML, CSS, and JavaScript code from a single prompt, as well as its adeptness in solving modified versions of classic logic problems, demonstrates its enhanced reasoning and language understanding capabilities.

The availability of the model's weights on Hugging Face, along with the option to use the API through platforms like OpenRouter, provides users with multiple avenues to experiment and leverage the model's capabilities. The increased token generation limit on OpenRouter, up to 231,000 tokens, is particularly noteworthy and could be beneficial for tasks requiring extensive output.

While the model still exhibits some limitations in handling certain modified problem scenarios, the overall performance suggests that the DeepSeek team has made substantial progress in enhancing the model's reasoning abilities. The author's excitement for the potential of future upgrades, such as R2 and V4, further underscores the rapid advancements in the field of large language models.

Overall, the updated DeepSeek V3 model appears to be a significant step forward, offering impressive capabilities that could have a meaningful impact on various applications, particularly in the realm of software development and problem-solving.