Over the past few months, our team has been extensively working on several projects involving large language models. It is impressive to realise what we could achieve with GPT-3.5. Now that GPT-4 is out, we want to share our thoughts on why it will further transform our lives in the coming months.
Undoubtedly, GPT-4 is a significant step forward in natural language processing and AI. It demonstrates fantastic results and will certainly lead to more sophisticated and personalised interactions between humans and machines. Moreover, it lays a foundation for creating new, exciting applications, making us much more productive in many everyday tasks.
We believe there are four reasons why GPT-4 is a significant milestone.
1. Multimodal Model
GPT-4 is not just a large language model like previous generations of generative pre-trained transformers. It is a large multimodal model that can seamlessly process both text and image inputs. As a result, GPT-4 can analyse documents with text and photographs, diagrams, and screenshots, similar to what it does with text-only inputs. For example, given a picture of a mock-up (e.g., drawing on a piece of paper) and a text instruction (e.g., asking to generate a website with an interface defined by the mock-up), GPT-4 can generate source code implementing a functional website based on the mock-up. Impressive! It takes AI capabilities to new heights.
2. Enhanced Performance
While the distinction between GPT-3.5 and GPT-4 can be subtle in casual conversation, the difference becomes apparent for more complex tasks, where GPT-4 can provide more reliable and creative results. For instance, in contrast to GPT-3.5, which passed a simulated bar exam with a score around the bottom 10% of test-takers, GPT-4 passed the exam with a score around the top 10%. During the developer livestream, GPT-4 demonstrated its ability to calculate standard deductions and total liability based on the tax code.
Does this mean we will ask GPT-4 to help review legal documents, automate filling out tax declarations, and diagnose diseases? Of course, it is crucial to consider the ethical implications and ensure that AI-generated insights complement human expertise. But it is already happening, and it will only accelerate from here, becoming the new norm.
3. Multilingual Mastery
GPT-4 performs better in 24 languages compared to the English-level performance of GPT-3.5. Among the list of languages are Latvian, Welsh, and Swahili – languages considered challenging for AI, as there are no substantial amounts of data in these languages for model training. As a result, GPT-4 unlocks a world of possibilities for global applications.
4. Steerability
One of the most exciting developments of GPT-4 is the addition of steerability. Developers, and very soon, ChatGPT users, will be able to direct AI to generate answers following a specified tone of voice and style through a "system" message. It allows for significant customisation within bounds and personalisation, empowering users to tailor AI-generated content to their specific needs.
While AI advancements like GPT-4 can lead to concerns about job displacement and potential misuse, it is essential to focus on harnessing these technologies responsibly and ethically, complementing human expertise and fostering collaboration between humans and machines.
Our team is genuinely impressed by the advancements in developing GPT-4, and we are excited to see how it will set the stage for further improvements. It makes us even more enthusiastic, realising that we have all the skills and expertise to help our clients leverage GPT-4 to transform their businesses.
If you are curious about how large language models like GPT-4 can benefit your business, we would be happy to connect and discuss potential applications and strategies. Get in touch with us: contact@samuylov.ai.
