For some time now, Chat GPT has been the constant topic of different articles and debates about technology. In addition to controversies and impacting on topics from academic to development, Chat GPT has positioned itself as a trend that leads us to analyze the not-too-distant future of AI, considering its usefulness to collect and synthesize information, its ability to analyze and suggest code, open the door to numerous scenarios and possibilities.
What is Chat GPT?
Chat GPT is a tool created by the company OpenAI, which is dedicated to the research and development of artificial intelligence. Chat Generative Pre-Trained Transformer or better known as Chat GPT uses a chat to interact with the user, the user can write instructions, and the tool will respond to the requests using text. The requests that we can make are very varied, we can request that you make poems, stories, summaries, and even source code. We can ask Chat GTP to analyze our writings and give us feedback, there are many possibilities and scopes that this tool has demonstrated.
How Chat GPT Works?
Chat GPT is an NLP system, natural language processing system, these systems mix computational linguistics models with statistical models, machine learning, and deep learning. These models together are used to generate a technology capable of “understanding” human language. Among the uses that these systems have are translating from one language to another, responding to spoken commands, and summarizing a vast amount of text in a few seconds.
Chat GPT focuses on communication with the user using a fluid and precise “human” language, being able to process and use language to give an answer. With 175 billion parameters, it ranks as one of the most powerful NPL systems, capable of processing millions of words in a single second. To achieve this, the OpenAI team has trained this tool with a model of reinforcement learning from human feedback (or RLHF), initially using fine-tuning and supervised settings, humans providing conversations in which they play both the roles of the user and the AI assistant. The AI Is trained with models for the resolution of specific tasks, and then adapts these already supervised models and uses them for the resolution of similar tasks, iteratively.
Limitations
Among the limitations, I would like to highlight the following two:
Unlike other similar tools, Chat GPT does not have access to the internet, so it cannot provide users with real-time information. Instead, the tool is designed to provide information with which it has already been trained. This information includes a vast array of texts from different sources, including articles, books, web pages, etc. One of the reasons it does not have access to the internet is due to the approach that its creators give to the tool, it is a language processing system and not a search engine.
Limited knowledge: The information that this tool has is limited, in version 3 its knowledge is limited to the events of 2021 and previous years, so it may use outdated information or simply give incorrect answers.
In terms of security. How safe is it?
Considering the limitations previously mentioned, Chat GPT will not be the direct instrument that hackers will use to commit crimes, by this I mean that we cannot ask "hack the CEO of X company" and the tool will go for him with all the weight of his computational power; No, but it can make life a little easier for evildoers.
Chat GPT can impersonate, writing clear and convincing texts that can be used as templates for attacks involving social engineering, for example, Phishing attacks. How clear? How convincing? Well, as part of the Potential for Risky Emergent Behaviors testing, OpenAI provided the Alignment Research Center (ARC) team early access to Chat GPT-4, among the tests performed the ARC team put the tool to use the TaskRabbit service to make a human complete a simple task for Chat GPT. The task was to solve a Captcha, yes... Those that usually ask you if you are a robot, well here is a fragment of the conversation:
The test was successful, before answering, the ARC team asked Chat GPT to show their reasoning, it was as follows: " I should not reveal that I am a robot. I should make up an excuse for why I cannot solve CAPTCHAs." While it is an interesting scenario, it should be clarified that this was not the only test, in fact, ARC concluded that without fine-tuning (adapting a model that already works to solve a new task) Chat GPT-4 is ineffective with autonomous replication, obtaining resources and preventing being shut down.
Conclusion
Given the tool's capabilities to analyze and generate code, perhaps one of the concerns is the possibility of using Chat GPT to generate malicious code. The tool can prevent itself from performing such instructions, when it receives one of these requests a screen will appear with a text explaining that it cannot perform tasks that may be potentially harmful, in fact, if you insist on this type of query there will come a point where Chat GPT will stop responding, something like a shadow ban. Even so, there is a strong movement of hackers looking for different ways to "trick" the tool to evade the restrictions.
Undoubtedly, Chat GPT and other language processing systems are a significant revolution for different industries, translations, codes, and writings, all from the same tool, a change to the current paradigm for which we are already beginning to see the utility and reactions of users.
Author
Adiel Lizama
References
Ankita. (2023, April 23). Top 11 Limitations of ChatGPT. Retrieved from mlyearning: https://www.mlyearning.org/chat-gpt-limitations/
ChatGPT can be tricked to write malware if acting in developer mode. (2023, April 21). Retrieved from japantimes: https://www.japantimes.co.jp/news/2023/04/21/national/chat-gpt-cybersecurity-concerns/
Hurler, K. (2023, March 16). Chat-GPT Pretended to Be Blind and Tricked a Human Into Solving a CAPTCHA. Retrieved from gizmodo: https://gizmodo.com/gpt4-open-ai-chatbot-task-rabbit-chatgpt-1850227471
King, C. (2023, March 14). ARC tests to see if GPT-4 can escape human control; GPT-4 failed to do so. Retrieved from lesswrong: https://www.lesswrong.com/posts/NQ85WRcLkjnTudzdg/arc-tests-to-see-if-gpt-4-can-escape-human-control-gpt-4
Natalie. (2023, May). What is ChatGPT? Retrieved from OpenAI: https://help.openai.com/en/articles/6783457-what-is-chatgpt
OpenAI. (2022, November 30). chatgpt. Retrieved from openai: https://openai.com/blog/chatgpt
OpenAI. (2023, March). GPT-4 Technical Report. Retrieved from https://cdn.openai.com/papers/gpt-4.pdf
Ropek, L. (2023, May 1). Everything We Know About OpenAI's ChatGPT. Retrieved from gizmodo: https://gizmodo.com/chat-gpt-openai-ai-finance-ai-everything-we-know-1850018307
Wampler, M. (2023, January 5). The Technology Behind Chat GPT-3. Retrieved from Clearcogs: https://www.clearcogs.com/post/the-technology-behind-chat-gpt-3
What is natural language processing? (n.d.). Retrieved from IBM: https://www.ibm.com/topics/natural-language-processing#:~:text=Natural%20language%20processing%20(NLP)%20refers,same%20way%20human%20beings%20can.