Decoding AI: Understanding GPT-3, GPT-4, LLMs, and Custom GPTs

GPT-3, GPT-4, LLMs, custom GPTs, and OpenAI are familiar terms, but understanding what each one does, what they mean, and where their capabilities lie can be challenging due to all the conflicting and sometimes complicated information out there. I'm Declan Cavanagh, Avec's Technical Consultant – AI and I'm here to break it down in simple terms to demystify AI and its components.

What is the difference between GPT3, GPT 4, LLMs, Custom built GPTs, and Open AI? 
GPT-3 is a language model developed by Open AI. It is known for its large-scale capabilities in natural language processing and generation. It is designed to understand and generate human-like text based on the input it receives. 

GPT-4 is the successor to GPT-3, offering advancements in natural language understanding and generation. It builds upon the strengths of its predecessor and has new features and improvements. 

LLMs (Large Language Models) refer to the category of language models, such as GPT-3 and GPT-4, that are characterised by their extensive size and capacity to process and generate human language. These models are trained on vast amounts of text data to understand and produce human-like responses.

Custom-built GPTs are variations or adaptations of existing language models tailored to specific use cases or domains. These models are fine-tuned or modified to address requirements, such as industry-specific language processing or enhanced performance on niche tasks.

OpenAI is the organisation behind the development of GPT-3 & 4 and other AI technologies.

Can GPT-3 perform web scraping? 
GPT-3 does not have the capability to perform web scraping directly. It cannot interact with websites, extract data, or handle HTTP requests. Instead, GPT-3 can assist in generating code for web scraping tasks based on provided prompts. For instance, you can prompt GPT-3 to write a Python script using libraries like BeautifulSoup or Scrapy to scrape data from websites. 

Does GPT-3 have built-in web searching capabilities? 
GPT-3 does not inherently have built-in web search functions. To perform web searches, one would typically need to integrate GPT-3 with external tools or APIs that can handle web search queries and retrieve information from the internet. While GPT-4 and some custom implementations of GPT-3 may include web browsing or search functionalities, these are not native to the standard GPT-3 model and require additional integration and setup. 

Can GPT-4 perform web searching within the consumer app? 
Yes, GPT-4, through the ChatGPT Plus subscription, has web browsing capabilities. This feature allows it to access and retrieve information from the web, making it more capable of providing up-to-date answers compared to GPT-3, which lacks this ability.

How does GPT -4's web search functionality compare to custom-built GPTs or agents with web scraping tools? 
While GPT -4's web search functionality within the consumer app offers convenience, real-time information access, and ease of use, it may lack the depth and customisation potential of custom- built GPT's, or agents equipped with web scraping tools. Custom solutions provide greater control and specificity but require significant resources to develop and maintain. Each approach has its strengths and limitations, and the choice between them depends on the specific needs and constraints of the user organisation. 

How dependable are LLMs (like GPT-3 and GPT-4) in providing accurate and hallucination-free information, especially when asked to cite sources? 
LLMs like GPT-3 and GPT-4 can be highly effective in generating coherent and contextually relevant test, but their ability to provide accurate, hallucination-free information, especially with proper citation, is limited. They perform better with well-documented, widely discussed topics but struggle with lesser-known entities and specific, up-to-date queries. Integrating external verification tools and using LLMs in conjunction with databases of verified information can help mitigate some of these limitations.

How do LLMs handle less known entities or specific investment-related queries? 
When dealing with less-known entities or specific investment-related queries, LLMs may produce less reliable results. Their training data may not cover niche topics comprehensively, and without access to real-time data or specific databases, they may fabricate information to provide a seemingly complete answer. This limitation underscores the importance of using additional data sources or tools that can access up-to-date and specific information relevant to the query.

What are the costs and challenges associated with using custom GPTs for web searching at scale? 
Custom GPTs offer greater control, specialisation, and the ability to integrate specific datasets and tools, making them highly effective for tailored applications. However, they come with the need to consider costs and challenges, including development, infrastructure, and legal considerations. Standard GPT models provide ease of use and flexibility but lack the precision and customisation capabilities of custom solutions. The choice between custom and standard models depends on the specific needs, resources, and scale of the application.

Can you direct the output of custom GPTs effectively compared to standard GPT models? 
Custom GPTs can be directed more effectively compared to standard models by tailoring them to specific tasks and integrating them with domain-specific knowledge and tools. This allows for more accurate and relevant outputs. However, this customisation requires significant expertise and resources, making it a more complex and potentially costly solution compared to using general-purpose models like GPT-4 with built-in web search capabilities.

What functionalities are available in the OpenAI API for web searching and scraping? 
While the OpenAI API itself does not natively support web searching or scraping, these functionalities can be achieved through integration with third-party tools and APIs. Enhanced browsing capabilities are available in specific plans, and continuous updates and new plugins extend the model's functionality. For specialised tasks, developers can build custom solutions using web scraping libraries and search engine APIs, often leveraging frameworks like LangChain to streamline the integration with LLMs.

Are there any new features or updates in the API that might enable web search or scraping? 
The OpenAI API itself does not provide direct web searching or scraping functionalities. For tasks that require live data fetching from the web, integrating with other APIs, such as Bing Search API, or using additional tools like web scraping libraries is necessary. Recent updates and features, such as the WebGPT project, enhance the ability of language models to access and cite web data, but these are typically built on top of the base models and require additional setup and configuration.