This year it was indeed a roller-coaster ride. Artificial intelligence has changed up everything, everywhere we see and hear the news that A.I is going to replace your job. Honestly until yesterday I was under the same fear and impression as you guys where. Without beating any further around the bush, let’s get straight into the topic, but before that it’s important for every one of us to actually understand how chat-gpt works.
How chat-gpt actually works?
Short answer, LLM. The long answer is chat-gpt [generative pre-trained transformer ] is an implementation of LLM which stands for ‘Large Language Model’, which refers to a class of trained models on enormous amounts of text data, to learn patterns and relation within a language.These models strive their best to generate a human-like text based on the trained techniques which are well known for the A.I developers like unsupervised learning.
Now, here is where things get interesting, GPT or generative pre-trained transformer is a specific architecture within the LLM. GPT uses the TRANSFORMER architecture to understand and generate the text. Here is the fun part, does the GPT actually understand and feel what it says? Turns out the answer is a big NO.
These things don’t generally think so much. They just think of the next most probable word to answer and that’s all the large language model does. Here is the context and here is the next word in the context. And this is the same reason why we get different response each time we regenerate the text, the probabilities related to what the next word could be changes. And technically they are not called as words, they are referred to as tokens which could be any character that is humanly understandable like a new line or a alphabet, or a question mark. Note that they keep on changing depending on the context of the scenario.
If you are confused about the similiarties and the differences between LLM and GPT have a look at this table below.
The good news !
Chat-gpt actually don’t understand. Although they can generate coherent and contextually relevant text, it lacks true understanding of the concepts it is talking about. It generates and operates text based on statistical pattern and probabilities rather than the model actually understanding the stuff it was talking about. So all the you-tubers who made video on chat-gpt creating a website or debugging a program or creating a malware was purely a statistical pattern and the program which the chat-gpt generated worked because it had high probability of working. Not because it was smart enough to learn and understanding programming.
Still not convinced? Try asking chatgpt any hard topics like ‘operating system and how it works’. Ask chatgpt on how different things works like ‘how does a translation look aside buffer works? ’ what is ‘direct memory access? , classify and explain in detail’ or ask ‘types of hazard in an operating system’ or ask chat-gpt to code a difficult program related to DSA or memory allocation algorithms. For sure chat-gpt will keep on regenerating vague text instead of giving an correct answer.
Here is an another example. Use of LLM is not enything new, we all have been using it for years. Everyone must be aware of the following sentence which contains all the 26 alphabets.
the quick brown fox jumps over the lazy dog
What what happens when you type in half way on a search engine like google.
Notice how quickly it guesses the next appropriate words based on what we have typed. We were all using LLMs long before there got popular and yes these are just advanced or powerful text completion algorithms which guess the order of next word nothing new, nothing surprising.
Well what about the stuff they say and talk about A.I in news?
Come on, by this time everyone should have realized that the media was, were, are , and will be using fear, despair , anxiety to control and manipulate people. If you are someone who consumes news and the public media one hour a day guess what, someone already planted their propaganda in your brain.
As told earlier, these models just guess the next appropriate word based on the context[input we have given]. Nothing new and just prompts most appropriate term. So all the dating, friend chat bots you are playing and messing around probably the chat both-they themselves don’t understand what they are saying but they are forced to say in a certain way due to their training models.
Still don’t trust me? watch the interview shows of ‘Michio Kaku’ an famous theoretical physicist whom you guys might have seen on many TV channels, Joe-Rogan podcast, and in many you-tube interviews. He is also an popular science writer, futurologist and a guy who knows something about science and tech. If you watch any or all of his videos where is talks about AI, he refers A.I chatbots them as a teenager who plagiarizes other people’s works and passes on it. It does not even understand or remember any of the stuff, we the humans are the ones assuming that it knows something that we humans lack. This feeling of inferiority is portrayed on media and on news, only God knows that their intentions are.
So does chat-gpt is going to take away our jobs?
Well it depends on the job you are talking about. Are you referring to low quality of work which has little to no requirement of using ones cognitive abilities? the ones that are repetitive and anyone without proper technical skill can do? If you are talking about those kinds of job then yes they will be replaced by A.i.
But if you are talking about technical jobs that require proper validation and error checking like programming, pentesting, or any domain in cyber-security then yes those jobs are not and will not be taken away by A.I. Most of the low quality works and job maybe replaced but yes the person with lot of experience and skill will not be replaced as someone is needed to check correctness of the answers displayed by A.I and to implement the solutions right?
For sure A.I is going to automate a lot of boring non productive task and make our lives easier. One could either leverage this new technology to their advantage or complain, mock and missing this amazing opportunity which will be passed on to those who leverage A.I.
Concerns about A.I
1. Since the text generated by A.I is based on probability, one should note that if not monitored and trained properly it could be disastrous
Again Michio Kaku suggests the to monitor and train the A.I models automatically so that they don’t spread mis-information or cause any harm. Since they don’t know what they are talking about they can spit out non-sense which could be very dangerous. And can also impersonate things. He suggest quantum robots to act like ‘fact checkers’ so that these kinds of harms does not occur. He also jokingly mention to make us understand the consequences of impersonation by saying ‘What if an chat-bot impersonates Putin and declares war on NATO nations?’
Funny but indeed scary. We already have an A.I called ‘CarynAI’ which impersonates an social media influencer ‘Caryn Marjorie’ and people pay the chatbot one dollar to talk to an A.I model that talks exactly like her. And on the first week of beta release she makes over $70,000 for just creating an A.I version of herself.
If A.I has the potential to replicate words and facial expression of a person then maybe we need to do something to regulate these A.I so that they don’t impersonate important or normal people and cause havoc. This also leaves an door open for cyber-criminals to impersonate people and open doors for cyber-criminals to make money online through unique and easy ways.
2. A.I is not cyber-proorf.
Although this term is cliche, you heard me right. A.I is **** I mean just look the the picture posted by VX-underground on December 5 2022. Open-AI was hacked using netcat LOL, I mean just look at these images.
Also these models are vulnerable to ‘PROMPT INJECTION ATTACKS’.
Few months ago an researcher who found an exploit won the game '‘ALPHA-GO’', and this A.I was defeated 14/15 times in a game by someone who does not know how to play this game but knew how to exploit it.
Just because user input was not sanitized, a lot of things can go wrong for an A.I and that is was exactly have happened. And here you have an slide talking on how the attack/exploit have worked.
To know and to learn more about these cyber-attacks on A.I I highly recommend to looking at sans-webcast from SANS Offensive Operations '‘VulnerabilityGPT: Cybersecurity in the Age of LLM and AI’' from 50.10 to 59.00. And I am convinced that no one can better explain these attacks apart from Matthew Toussain [author of sans sec460 course] where is talks about all the injection and prompt hi-jacking in detail.
Conclusion
Although I agree that this is just a summarized version of the original sans webcast by sans instructor Matthew Toussain, after watching this webcast and after little bit of digging into the topic, I felt alive and the fear of A.I taking out our jobs was reduced and I literally felt the light at the end of the tunnel, a ray of hope. For those into cyber-security they also have a huge bright future or a large scope as the people who build software, programs, chat bots and A.I need people and skilled individuals who can safeguard their products from cyber-attacks.
And if you are still in despair that A.I will replace you, worry not my friend, you always have an option B and that is to become a farmer like Bill Gates or at least own some farm land. [sarcasm]
Special thanks to Matthew Toussain, without him I would have never gotten grip of what’s the reality and could have never even bothered to write a blog about this topic at first place.
