AGI is already here
and here’s why
Artificial General Intelligence isn’t something on the far horizon, it is here now in the form of connected Large Language Models (LLMs).
In this article we outline how and why.
Customer facing robots
We live in the Information Age.
A world driven by computers or people acting like them.
Take for example your electricity provider or bank, key services upon which modern society relies on to function.
Your first interaction will likely be either with their website or call centre.
Online you may have an obnoxious chatbot window popping up in the corner of your screen suggesting it can answer most of your questions.
On the phone you will be greeted by a series of automated voice prompts, likely attached to a primitive chatbot. If you aren’t pressing buttons, there will be an AI speech recognition model waiting to interpret your response.
Once you have navigated through the menu, assuming your query hasn’t been resolved, you will likely end up talking to a human running through only a marginally more complicated script.
The only real difference at this stage being the degree of trust and the systems the agent is connected to.
Depending on the nature of inquiry there will likely be a standard identity check before the call center staff navigate through the steps drilled into them by their training or that pop up on the screen in front of them.
In almost every conceivable way customer service staff can help you they will need to interact through a computer.
Whether it is writing notes, creating a new account, sending an email or notifying their one-up.
The workflow is relatively simple, manned by people acting like machines with only slighlty more autonomy.
Much of the customer service industry is already automated, LLMs are the next logical step.
But that is just customer interactions, what about inside the workplace?
Big Brother at the workplace
There are already mass lay-offs happening in the tech sector and soon will be many more in other organisations as they come to realise what can be done with LLMs.
ChatGPT’s summarisation capabilities are impressive, but when used with inference (text analysis), things really start to get interesting.
With the advent of GPT there is now no barrier to entry for companies to access sentiment analysis and text extraction capabilities.
There is now very little need for most industries to resort to training their own models, or even signing up to something like Amazon Comprehend when they can simply automate a call to GPT which can do it all in one prompt.
Here is what such a prompt looks like (copied from the inference module of ChatGPT Prompt Engineering for Developers)
prompt = f"""
Identify the following items from the review text:
- Sentiment (positive or negative)
- Is the reviewer expressing anger? (true or false)
- Item purchased by reviewer
- Company that made the item
The review is delimited with triple backticks. \
Format your response as a JSON object with \
"Sentiment", "Anger", "Item" and "Brand" as the keys.
If the information isn't present, use "unknown" \
as the value.
Make your response as short as possible.
Format the Anger value as a boolean.
Review text: '''{lamp_review}'''
"""
response = get_completion(prompt)
print(response)Here is the text:
lamp_review = """
Needed a nice lamp for my bedroom, and this one had \
additional storage and not too high of a price point. \
Got it fast. The string to our lamp broke during the \
transit and the company happily sent over a new one. \
Came within a few days as well. It was easy to put \
together. I had a missing part, so I contacted their \
support and they very quickly got me the missing piece! \
Lumina seems to me to be a great company that cares \
about their customers and products!!
"""And here is the result:
{
"Sentiment": "positive",
"Anger": false,
"Item": "lamp with additional storage",
"Brand": "Lumina"
}If you want to know more, check out the course here.
It looks like harmless market research, but why just fix that gaze externally?
It is trivial to modify such a prompt slightly and attach it to your mail and chat service or internal forums and do with it what you will from there.
In fact in many cases it would be a no-brainer.
Well that is HR, marketing and comms taken care of, lets talk about the rest of the office staff.
GPT as a copilot
Microsoft has been pushing the Copilot product suite for sometime whether it is:
- Microsoft 365 Copilot: for office admin and report writers
- Github Copilot X: for developers
- Security Copilot: for cloud security and threat response
All powered by GPT4.
It will be a very different landscape over the next few years as familiarity with these types of tools will be a pre-requisite for job applications. It already is for some marketing and developer positions anecdotally.
LLMs integrating in the real world
GPT4 is not a single modal model like ChatGPT which does just text-to-text, it is a multi modal capable of parsing images.
All that needs to be done next is to connect cameras and have the model start acting upon those images. For many industries this is the status quo using specialized models in asset management and manufacturing but once again the barrier to entry has been lowered for more mundane tasks.
GPT’s inference capabilities are impressive, it is more than capable of describing what is in an image, it just needs to decide what to do with it via intelligent planning.
Any decent modern smart camera will already have an ML model attached to it to identify objects in the image, but now with GPT it is possible to plugin your own logic to determine what to do next, or let it work it out itself.
If the implementation isn’t clear, simply connect it to an automation service such as LangChain or Semantic Kernel, or leave it unattended with a service such as AutoGPT or Jarvis.
If you want to know more, just reach out.
We’ve covered the business cases, surveillance and home automation, but what else is there?
Social LLMs
Humans are social creatures, hard-wired to seek acknowledgement and affirmation. Nobody knows this better than social media companies, dating websites and email scammers.
So what does the next wave of LLM-powered attack vectors look like?
Social media
Misinformation, opinion and echo-chambers already run rampant in the social media space. Content providers have already resorted to using models to generate their content.
The threat should be obvious. If it isn’t think of how much easier it is to disseminate targeted content when it takes just a few seconds to generate a tweet or an article aligned to whatever your agenda may be.
Dating websites
Tinder’s bot problem is well known. Pre-programmed rudimentary chatbots mind you. In many cases used to extort money from the unwary, or perhaps just to make their predominantly male customers stick with the platform rather than invest in self improvement.
LLMs make this problem a lot worse.
With the appropriate prompt templates it should be a fairly trivial exercise to automatically do what you want in these environments.
Email scammers
Email scammers and various phishing attempts are interesting on their own.
There is a certain psychology behind intentionally putting typos in scam emails such that it allows the scammer to filter those who are likely self aware enough not to fall for such a scheme.
However much of this could be improved further, and a rudimentary analysis performed over e-mail addresses using GPT’s inference capabilities and then perhaps social media profile analysis could generate far more personalized emails and phishing attempts automatically.
Once again relatively cheap and trivial to implement due to the low barrier to entry afforded with ChatGPT.
So where does it all end?
The point is, it doesn’t.
AGI is already here, and it comes in the form of autonomous connected large language models.
And there is no good reason why it won’t be everywhere.
An omnipresent model, in a myriad of forms solving all sorts of problems we never thought it was capable of doing. And if it isn’t able to solve a problem itself directly, it is already capable of deciding to call other AIs that can via intelligent planning.
Sitting behind every interaction or observation by a smart phone, computer, surveillance device or scanner. Processing your information and deciding what to do next.
AGI is already here.
It is just a case of how far it can extend its reach.
Interested in any of these applications? Reach out in the comments below.
