The Future of Conversational AI: ChatGPT4 Takes Center Stage again
People are all aware of the history of chatgpt, how it all started and why it started, and how it came to the state it is today, with a new version of GPT-4 coming this Wednesday. I want to jot down my understanding and predictions for the future in this article.
A brief understanding of GPT
Chatgpt is a language model based on a neural network architecture called a transformer. Since their first appearance as transformers in a research paper, they have been widely used in building language models.
At a very high wage level, the sole purpose of chatgpt is to predict the next word in a sequence of text; this is done by training the model on a large corpus of data (Current ChatGpt3 is trained on 175 billion parameters).
One key feature of ChatGpt, which everyone knows, is its ability to generate coherent and semantically meaningful text, even when the user input is incomplete or impartial. This is achieved using the (attention mechanism), allowing the model to break the entire input text and give more weight to particular words or phrases more relevant to the required prompt.
ChatGPT 4 New Features:
- Outperforms exisitng LLMs, State of the art models
- Outperforming GPT 3.5 in exam results
- Visual Inputs
- Steerability
Limitations
- Hallucinating facts
- Truthful QA
- Bias
- Lack of events on events occured after Sep 2021
- Sensitive prompts further needs improved
Source: https://openai.com/research/gpt-4
So What next …..
As everyone is talking about the current tech Generative AI, it can be seen as both an aid and a bane to human consumption.
When we see from the angle of the aids, we know a lot of creative and innovative use cases and startups coming to light that integrated ChatGpt into their existing products as well as new products. These expand over diverse industries, including commerce, healthcare, and People Analytics, and aid people working in tasks like L1 ticket support, Marketing team, and Content creators, to articulate just a few.
So as an extension, Microsoft also released the Visual ChatGPT, which acts as an interface between ChatGPT and visual feature models (VFMs). The prompt manager helps ChatGPT determine whether it needs to use a VFM to provide the necessary output and takes care of iterative reasoning and housekeeping. The immediate manager is critical to the system as it is responsible for non-language queries and allows ChatGPT to rely on the capabilities of VFMs rather than hallucinations. Visual ChatGPT lowers the hurdle to access text-to-image models and potentially incorporates compatibility across various AI tools.
While there are strict assumptions about how the models can be easily misused, there is a high chance these models can be more biased based on the data on which there are trained; these can easily create offensive outputs.
My Views.
With the new Gpt4, the Hallucinations problem will continue escalating, and it might be challenging to handle the bias. And As evident, it does an excellent job of giving the output with its model being trained on higher data points than GPT3. Also, ethical problem handling might be complex cause handling these models at a large scale still needs time and effort.
Nevertheless, as time advances, we will see much development coming into the light. Even OpenAi opened Eval capabilities to know the developer perspectives and are looking for more contributions from developers.
Thanks for reading!…
Watch this space for more additions to the list of topics. Feel free to shoot me for any discussions in the comments below or connect with me on LinkedIn.
If you thought this was interesting, leave a clap or two and subscribe for future updates.
