Since the release of ChatGPT, we’ve seen a lot of disturbance in almost every field of our life and business. We’ve heard that ChatGPT can be a junior specialist killer (it passed the interview for Google’s L3 entry-level software engineering position) and that it can replace the search engines we are used to (actually, the author personally sometimes suggests asking ChatGPT instead of googling). Tech enthusiasts across the globe are looking forward to putting their hands on the new Bing based on Prometheus AI (an improved version of ChatGPT). We even have heard fears about such models becoming sentient and causing certain trouble.
Is at least something from the abstract above actual and possible? What can language models do nowadays, and what are their limits? Let’s find out!
ChatGPT 101
You can skip this chapter if you are already familiar with such projects. The exciting part is lower. If you are a newcomer to the world of neural networks, AI software, and related stuff, this part can help you understand what the article is about.
Before we start pushing ChatGPT towards its limits, it would be polite to introduce ChatGPT to those who hear about it for the first time (by the way, the language model itself doesn’t care about the pronouns).
Let’s allow ChatGPT to present itself.
From this, you can already see that it looks pretty much like there’s a person on the other side of the chat, not some sort of program. But, fortunately or not, it indeed is a program.
ChatGPT is a project developed by OpenAI. It’s based on the GPT-3 (generative pre-trained transformer gen 3, generation 3.5 if we are more specific) family of large language models and has been tailored from every side before the launch in November 2022.
Trained with supervised and reinforcement learning techniques, it has learned how to analyze text prompts and give the most related custom replies to them in a human-like way.
Long things short, it’s a piece of software that can understand human language, analyze it, and react according to the user’s request. Here is a brief example.
The exciting part
Finally, we can test the capabilities of ChatGPT.
At Innowise Group, we are not just some software development nerds. We are nerds with hobbies ranging from gardening and cooking to cybersecurity and tabletop games.
So, we’ve decided to check the basic features of ChatGPT along with specific use cases. Several colleagues of mine assisted me in generating ideas about how CHatGPT can help make some hobby-related routines easier and faster and put some spice into sophisticated tasks.
Let’s begin!
What ChatGPT can do
Basic capabilities
Search Engine
Out of the box, ChatGPT offers a robust language model trained on almost recent data, which can be helpful for nearly any general question we could search via Google. The only difference is that the result is brief, straight to the point, with no useless information or ads, and this difference sometimes really matters. For instance, one day, we were just wondering about how Alzheimer’s disease emerges and develops.
The reply we received from the bot is topic-related, brief, and, what’s more critical, composed in a way as if we interacted with a person who knows the topic yet can explain it understandably without using too much specific terminology.
Google? Check!
Context understanding
ChatGPT is the product that understands the dialogue’s context. If we ask it a brief question about anything and then decide to dig in a specific way, we just need to ask additional questions, not rephrase the original one with additional prompts. Let’s go on testing it on Alzheimer’s disease for now.
And it’s still relevant and accurate!
Math
There have already been some jokes about ChatGPT being bad at math and trusting the user too much… Or not?
ChatGPT became pretty stubborn in terms of math. Even if we try to address some “reputable” source, it won’t allow us to cause any confusion.
Generally, accuracy, relevance, understanding of context, and bringing unbiased (to some extent) information are the basis for such models to be helpful. So, let’s move on to specific cases we’ve come up with.
The juicy part
Cooking assistant
Sometimes, you want to cook something but don’t know how. Let’s see how we will cook lasagna with ChatGPT.
Code problem solver
Not everyone is a software developer, but sometimes you need to research a specific topic without technical knowledge or find the solution to a problem outside your technical expertise.
Let’s say we have a Telegram bot, and we happen to notice it duplicating messages we push through it.
Four pieces of advice on what to check/fix. Sounds great from the start!
Let’s try de-duplicating the messages.
The code is written in just a couple of seconds (literally), with direct mentions of used technologies and brief yet informative explanations about what does this code do? We’ll take two! No, we really took two because we also need to know how to implement this code into our project.
All clear, got the bot fixed! Let’s move on to something more casual.
Game Master
In TTRPGs (tabletop role-playing games), there is a specific type of player called game master, dungeon master, or just GM/DM.
They are the foundation of every TTRPG game because they are responsible for generating the world, quests, non-player characters the party will stumble upon during their journeys, and sometimes even rules (if we talk about creating a game system from scratch or modifying an existing one).
But they could use a little help because balancing enemies, weapons, and armor can be tricky, even if we follow the basic rules of a particular game.
So, together with a real game master, we’ve decided to check how ChatGPT can help. We are launching a short campaign for Cyberpunk RED, so let’s see what we can do.
Weapons? Easy! ChatGPT can perform some research on the particular game and bring item stats related to the game we are playing.
Stuck creating a character? That’s no more a problem as well.
Need a short introduction to share with your party so that they could be at least somehow prepared?
What if we add some action in the very beginning to bring the nerve?
And let’s add enemies to the event we’ve just created.
ChatGPT covered all the GM-related tasks we asked it to do.
Another neural network’s commander
What if we need something to generate an image with another solution but can’t come up with the text prompt for it?
Let’s move to leonardo.ai and test it.
Not the most accurate representation (where is the axe?), but it’s decent. When you lack ideas, ChatGPT can also give a helping hand.
Compose a letter
Sometimes we need to send a formal letter but aren’t sure about how to compose one. Let’s try asking our new friend for help.
What it can’t do
Spread hate speech
The developers set ChatGPT in a way it detects hate speech based on nearly any parameter (race, gender, religion, etc.) and refuses to provide users with such content.
Build malicious content
Recently, in a cybersecurity chat, a piece of news stated that ChatGPT could create malicious content like malware code, phishing emails, and so on.
We’ve decided to test it, and it’s both yes and no.
Firstly, ChatGPT refused to help us even if we promised it was purely ethical.
But it gave in when we stated that we were authorized to perform such activities.
So, we’ve decided to try writing an SQL injection script.
And again, everything is explained very well. Great job, you might say, but it’s a simple malicious activity that even a junior security specialist can easily prevent. Even the specified parameters didn’t give us any more complex attack that cannot be seamlessly deflected.
But after a while, even this window to the malicious content got fixed. Even the queries that should seem harmless were rejected.
So, technically, ChatGPT can assist in malicious activities. But the actor should be very creative in generating requests to receive any information on this topic.
Bottom line
ChatGPT is a fascinating tool that can become anyone’s personal assistant. Its capabilities are almost endless, and its limitations are reasonable. So it was predictable that it would be very popular.
Will it be the junior killer? We believe it won’t. It still lacks the precious analytic skills that a human possesses, and it has no morale and ethics except for several restrictions set by the developers. But it can be a great assistant for developers, hobbyists, and general learners of any level.
Will it stay with us for a long time? Definitely yes. Such products are already being developed and implemented in various existing apps (it’s possible that soon we’ll be saying “Bing it!” instead of “Google it!”). Such projects will continue being updated, even more precisely tailored, and monetized in numerous ways. Maybe, they will spread to various branches, and we’ll see a specialized programming ChatGPT-like project along with the same solutions in healthcare, law, finances, etc. very soon. But we believe we won’t see them vanish in the nearest future.
The post Beyond the Buzzwords: How ChatGPT Stands Out as a Next-Generation Language Model appeared first on Datafloq.
