On Tuesday, March 14th, OpenAI presented a keynote for developers, showcasing its latest generation large language model (LLM), GPT-4. An abbreviation for Generative Pre-trained Transformer, GPTs represent the underlying algorithms and artificial intelligence dataset, on which OpenAI builds chatbots. The most popular example so far was ChatGPT, which was launched in November and quickly amassed over 100 million users worldwide. Now, while ChatGPT was previously based on an older LLM, GPT-3.5, GPT-4 presents whole new abilities for the text-generating tool — and what’s more, businesses and users alike can take advantage of its abilities, too.
On this note, here’s taking a look at what exactly is GPT-4, and five key ways in which the new LLM improves the popular AI text generator, ChatGPT.
Table of Contents
What is GPT-4?
GPT-4 is the latest LLM that will power a host of applications, including OpenAI’s paid-only chatbot, ChatGPT Plus, and Microsoft’s new AI-powered Bing chatbot-plus-search engine. In fact, shortly after the unveiling of GPT-4, Microsoft said in a blog post that Bing had been running on GPT-4 for a while already. To be clear, it is not a tool, but the underlying base of data and algorithms that will power various chatbots, search platforms, and other tools that businesses create on it.
Is GPT-4 a replacement for the older ChatGPT?
Well, not really. ChatGPT is the AI chatbot, the tool, that OpenAI launched and made public in November. The tool is the vehicle through which users can experience various iterations of GPTs, meaning that while ChatGPT will remain constant (and perhaps evolve in its interface over time), the underlying logic will change. Back when it was launched, ChatGPT was powered by GPT-3.5, the then-latest language model that used billions of data points to understand what a user wrote, and offer a human-like response based on the knowledge it had.
Theoretically, GPT-4 is pretty much the same, but just better. This means that the new version of ChatGPT will be powered by GPT-4. However, do note that OpenAI has only offered paid access to users for trying out GPT-4 — to use this, you can pay $20 (about Rs 1,650) per month to get access to ChatGPT Plus, which is now powered by GPT-4. The free version of ChatGPT, which you can still access here, is powered by GPT-3.5 and will continue to remain so for the time being.
Five key differences between the two
With this out of the way, here’s looking at five key differences between GPT-4 and the older version of ChatGPT that was powered by GPT-3.5.
GPT-4 can process images
The first and biggest change that GPT-4 brings is its ability to understand images. In simpler words, while ChatGPT could so far only understand text inputs, with GPT-4, you can now simply add an image to the query field on GPT-4-powered ChatGPT, and ask it to decipher the image in various ways. For example, one of the demos shown by OpenAI used an image of a VGA cable being connected to an iPhone and asked the new GPT-4-powered chatbot to explain why the image was funny.
The GPT-4 chatbot responded that using a VGA cable with a modern smartphone was impractical, and therefore, humorous.
Once the feature is opened up for more users, this can be used to translate photos of bills, or illustrations of a webpage, into text or code. To be clear, GPT-4 will still produce only text outputs, but users can use images as part of their queries, and GPT-4 can understand them.
It was also trained in Indian languages
On the OpenAI research page, the company said that of the 26 global languages that GPT-4 was tested on, there were five Indian languages — Bangla, Urdu, Punjabi, Marathi, and Telugu. This could technically mean that users may soon be able to ask a GPT-4-powered bot, such as ChatGPT Plus, questions in local languages, and get an answer. However, it is not yet clear if the responses would be offered in the local languages too, or in English only.
A Mint report said that Indian companies are likely to leverage GPT-4 to build more advanced local language chatbots, a factor that has so far been a roadblock for AI to solve.
OpenAI also said that it uses Azure Translate for the feature, which suggests that GPT-4 likely understands only English, but queries may be translated on either end to process local language inputs. At the moment, it is unclear how this would work, but the feature promises a future where more regional languages are supported by ChatGPT.
It is more powerful and accurate than before
Following the release, there were two key data points that OpenAI shared on its landing page for GPT-4. These state that the new LLM is:
- 82% less likely to respond to requests for disallowed content
- 40% more likely to produce factual responses than GPT-3.5
This is based on the fact that GPT-4 was trained on a significantly larger volume of data than what ChatGPT’s GPT-3.5 was based on. While OpenAI did not specify how much, a TechCircle report said that GPT-4 is likely to be using nearly 3 trillion data points — over 17x more amount of data than the 175 billion parameters that GPT-3.5 used.
As a result, you can now expect more accurate, faster, and more in-depth responses to your queries. It should also become more difficult to ‘trick’ GPT-4 into giving wrong or hateful answers — a problem that ChatGPT has already faced before.
To give an example of how smart it is, OpenAI said that GPT-4 could clear the US law exam in the 90th percentile (i.e. by scoring among the top 10 percent students), and score in the 99th percentile in a US Biology olympiad (i.e. by scoring among the top 1 percent). In comparison, ChatGPT with GPT-3.5 could only score 10th percentile and 31st percentile in the two exams, respectively.
It can write long-form content
One key ability for GPT-4 is the ability to store more information within one query. During its developer keynote, OpenAI said that GPT-4 will have 32,768 memory tokens for answering one user query — up from 4,096 tokens in one ChatGPT (GPT-3.5) query. This corresponds to 8x more amount of information that GPT-4 can process in every user query — a factor that OpenAI said will allow the new LLM to understand much longer and more complex user queries, and generate long-form written content based on such inputs.
Interestingly, the feature comes a day after Google announced that it will be integrating its own LLM, called PaLM (Pathways Language Model), into Google Docs and Gmail. This means that a select number of early access users can already use Google’s own ChatGPT-like chatbot within Docs and Gmail to ask it to generate written content — such as job descriptions and employee information.
Plenty of thought toward safety
OpenAI also detailed that it has focused on safety with the new GPT-4 model. As part of its release, the company acknowledged that while there could still be issues around AI bias, safety, and accuracy of information, they have worked with over 50 AI safety experts to vet the responses that GPT-4 will produce. However, whether this could actually work or not will remain to be seen as more users gain access to GPT-4.
