Large-scale language model 'Phind' outperforms GPT-4 in coding



It has been revealed that Phind, a search engine for developers using generative AI, has surpassed OpenAI's GPT-4 in terms of coding ability.

Phind - AI Search Engine and Pair Programmer

https://www.phind.com/blog/phind-model-beats-gpt4-fast




Phind is a search engine for developers that was originally known as Hello Cognition (beta.sayhello.so) , and was relaunched around March 2023 under the name Phind. At the time of article creation, the Phind model is the 7th generation, which is built using a uniquely fine-tuned model ' Phind-CodeLlama-34B ' based on the large-scale language model ' Code Llama ' released by Meta. It is said that there is.

Phind can be used for free from the official website and without registering an account. When you access the link, a text input field will appear, so enter it to search.



The results will be displayed in Japanese. Unlike ChatGPT, the website that was used as a reference for outputting the results is displayed as the source.



It can even answer questions unrelated to coding, such as '

Please tell me how to make curry .'



According to the operator of Phind, the 7th generation Phind model exceeded GPT-4 in the large-scale language model evaluation index 'HumanEval' and was rated as the best overall open source coding model. About.

By using NVIDIA's

TensorRT-LLM library, Phind can process 100 tokens per second in a single stream, achieving speedup five times faster than GPT-4. Furthermore, Phind seems to support input of up to 16,000 tokens.



In Phind's Discord community, even users who have access to paid GPT-4 often choose Phind. The operator said, ``The Phind model still has some rough edges, and we plan to continue improving it.''

Related Posts:

in Software, Posted by log1p_kr