ChatGPT, OpenAI’s most famous generative AI revelation, has taken the tech world by storm. Many users pointed out how helpful the tool had been in their daily work and for a while, it seemed like there’s nothing that the tool cannot do.
However, while it’s in fact very powerful, more and more people point out that it also comes with its set of limitations.
What kind of limitations, exactly? To find that out and set up a coherent list of what ChatGPT and GPT-4 are missing, I’ve spoken to Monterail’s biggest AI enthusiasts, who work with generative AI almost every day.
And as a bonus, I will also look beyond ChatPGS’s current shortcomings – and analyze the recent information on how ChatGPT and GPT will likely be developed in the near future.
Without further ado – let’s go ⬇
If you stumbled upon this text, but don’t know much about Generative AI yet (for example, about tools like MidJourney, Jasper, and more) make sure to read the first blog post from our AI series first:
GPT stands for Generative Pre-Trained Transformer which signifies the model’s ability to generate text, the fact that it’s been trained using existing data, and its architecture that’s based on Transformer, a neural network model used for natural language processing tasks.
When it comes to the limitations of GPT language models and ChatGPT, they typically fall under two categories.
Firstly, there are challenges related to general AI and AI as a domain that touch upon the ethics of this technology (as described in this guide to ethical AI). Specifically, this includes certain biases, a phenomenon known asAI ‘hallucinations’, and legal issues that may undermine the progress of this technology. We’ve discussed these issues in more detail in the first article from our AI series, so we won’t discuss them in this text.
However, what we’re going to discuss is everything that falls under the second category of AI shortcomings – which typically includes the limited functionality of ChatGPT and similar tools.
So, let’s now take a look at specific issues.
GPT-4 has a cut-off date of September 2021, so any resource or website created after this date won’t be included in the responses to your prompts.
Luckily, with GPT-4, your prompts can be longer than in the case of the earlier versions, so you can supplement them with additional information or context that will improve the final output.
Additionally, GPT-4 doesn’t have access to the latest data nor does it have access to your company’s internal information and subject matter experts. What it can do is create generic content based on its dataset. As mentioned above, developing more in-depth studies and articles based on your experience and domain knowledge will require a bit of prompt engineering empowered by additional details and context.
Although It’s not stated in the official ChatGPT documentation, users have spotted that the tool may stop responding, or take a very long time and then ‘hallucinate’ if the prompt is very complex and longer than around 500 words or 4,000 characters. The same goes for the response the ChatGPT can produce – it will usually be around 500 words or 4,000 characters.
For now, ChatGPT can only generate text-based replies. In the basic version of the product, your prompts have to be text-only as well.
With the introduction of the developer mode of GPT-4, you can use both text and images in your prompts, and the tool can correctly assess and describe what’s in the images you’ve provided and produce outputs based on that.
As mentioned, ChatGPT was pre-trained using the dataset that was last updated in 2021 and as a result, it cannot provide information based on your location.
However, this may change following recent news and releases from the OpenAI team. You need to sign up for the waitlist to use their latest feature, but the latest ChatGPT plugins allow the tool to access online information or use third-party applications. The list for the latter is limited to a few solutions for now, including Zapier, Klarna, Expedia, Shopify, KAYAK, Slack, Speak, Wolfram, FiscalNote, and Instacart.
Many people voice their reasonable concerns regarding the security of AI tools, but there’s also the topic of copyright. While the texts, and images generated by artificial intelligence are for now not deemed copyrightable, we need a discussion around intellectual property and how to protect human-authored works of art, books, and scientific studies and ensure that companies such as OpenAI will have to disclose where their datasets are coming from and incentivize the authors of the materials used to build those similarly to how it’s been proposed in the early version of the recently released EU’s AI Act.Krzysztof Kaiser Head of Product Design
I’d appreciate it if there was more transparency on the sources of generated insights and the reasoning behind them. I’d also like to see the ability to add specific domain knowledge and the customization of where the outputs may come from i.e. only backed up by specific scientific sources. Although ChatGPT may refuse to answer when asked about sensitive topics, it would be great if it could have in-built biases and misinformation detection feature to further secure the tool and ensure that it won’t be used for malicious purposes.Darya Demidenko Former Business Analyst
AI tools are already being used to assist in completing work-related tasks, so the natural thing here is using them to build an assistant that would know everything about a certain company or a product, and would help i.e. support these teams to perform more efficiently. One thing I'd really like to see, and something the AI community is also pushing towards, is the ability to self-host tools like ChatGPT and use them locally without the need for internet access. This would allow us to use the model for sensitive internal data as well and would address the security concerns that people have about using AI and uploading their data to external servers.Artur Rosa Frontend Architect
To try to predict the future of ChatGPT and similar tools, let’s first take a look at the timeline of OpenAI GPT releases. This may help us understand when the upcoming launches from their team will occur and what type of models and products they may be focusing on in the next few months and years.
While GPT-4 has certainly skyrocketed artificial intelligence as a concept into everyone’s attention, it’s important to highlight here that the product itself is not new and has been preceded by a long line of previous models from OpenAI.
What’s more, the organization’s CEO, Sam Altman, is a person you may be familiar with already. Before focusing on bringing AI into the public eye and making it more widely used, he was a partner at Y Combinator, the largest startup incubator in the US that brought us Airbnb, Stripe, and Dropbox. He’s a well-established figure in the tech world and not a newbie to successfully growing a product.
Just to recap how the OpenAI’s tool has been progressing over the years, let’s have a look at the timeline of the GPT releases:
December 2015 - OpenAI is established in San Francisco, with the sole purpose of finding a way to build General AI. According to Sam Altman, this goal was quite publicly frowned upon, and as he mentioned in this interview, the company was mocked by the rest of the scientific community.
June 2018 - The OpenAI team released GPT-1 – known back then as simply, GPT – a large language model (LLM) that was trained using a dataset of 8 million websites and had 117 million parameters.
November 2019 - An even mightier version – GPT-2 – was made available. This time, it was trained on a dataset consisting of 40 GB of text with 1.5 billion parameters.
May 2020 - GPT-3 was launched and yes, it was more powerful than its predecessors. The dataset it was trained on included 175 billion parameters and 570 GB of text.
November 2022 - Here’s where almost everyone started paying attention: ChatGPT was released to the public with its abilities and similarities to natural language models seen by many as revolutionary.
March 2023 - With the interest in ChatGPT and General AI still on the rise, the Open AI team launched its latest LLM, GPT-4.
With the timeline of the previous launches from the OpenAI team, the question of when GPT-5 will be released becomes valid - I will discuss it in the section below.
The short response to the question ‘when will GPT-5 be released?’ is – ‘not very soon’.
As you can see on the timeline, a new version of OpenAI’s neural language model is out every 2 - 3 years, so if they want to make the next one as impressive as GPT-4, it still needs to be properly trained. This simply takes time.
At a recent event at the Massachusetts Institute of Technology, Sam Altman referred to the open letter that has been signed by the number of people in the tech industry, including Elon Musk and Steve Wozniak, that was requesting the organizations currently ‘training the models more powerful than GPT-4’ to stop their research for at least six months. Altman mentioned that the letter inaccurately claimed that OpenAI is currently working on the GPT-5 model. ‘We are not and won’t for some time’, was his response to this claim.
What he did stress though was that the current GPT-4 model will be expanded and that the new features will be added on top of it, including the ones that will be addressing the security concerns listed in the open letter.
There are sources that mention the planned release of GPT-4.5 for September or October 2023, but this hasn’t been confirmed by OpenAI as of May 2023.
With all that being said, even with the limitations and missing features, ChatGPT and GPT-4 as a neural language model are the most impressive and bold applications of artificial intelligence to date.
Open AI’s competitors, including Bard and Claude, are also taking steps in this direction, but they are not there just yet. It may change very soon though, especially with the update to Google Search and Google’s PaLM announced at the latest Google I/O presentation on 11/May 2023.
There are a lot of predictions and speculations about what will happen with the ChatGPT and OpenAI’s LLMs in general, but the consensus here is that the real milestone will be the upcoming release of GPT plugins to the general public — as mentioned, they’re currently available in ChatGPT Plus plan only.
As of May 2022, the OpenAI API allows you to connect to and build tools based on the company’s existing language models or integrate the ready-to-use applications with them.
It’s important to note here that while ChatGPT may be the perfect off-the-shelf solution, it won’t cover all of your product needs and unless you’re using OpenAI API or plugins, you can’t integrate it with your tools.
That’s why it may be so beneficial to consider developing your own generative AI solution, fully tailored to your specific needs.