AI-Detection Tools: The Joke's on Us

Good Morning AI Runners

Here's what we've got for you today:

• Wait, what are neural networks?

• AI-Detection Tools: The Joke's on Us

• Hungry Hungry Hippos

Wait, what are neural networks?

Feeling lost with all the concepts and technical terms in AI? Don't worry, you're not alone. I, too, am learning so much just writing this daily newsletter. So, for today's post I thought we could dive into what neural networks are and how they work because they seem like an important concept (and they are!).

Neural Networks 101

Neural networks are a subset of machine learning that forms the foundation of deep learning algorithms. And, as the name suggests, their structure and names are inspired by the design of the human brain. Basically, neural networks are designed to mimic how biological neurons communicate with each other.

Neural networks are made up of layers of nodes or artificial neurons. Each node has a threshold and weight, and if the output of an individual node exceeds the threshold value, it sends data to the next layer. These networks use training data to improve and learn their accuracy over time. And let me tell you, they can be pretty powerful tools in artificial intelligence and computer science, allowing us to classify and cluster data at lightning speeds.

So, how do these networks learn? Well, there are two ways: supervised and unsupervised learning. Supervised learning is when the neural network is shown input data and a desired output, and it's trying to match the two. It's like having a teacher showing the network what the answer should be. On the other hand, unsupervised learning is when the network is only given input data and it's trying to find patterns on its own, like a detective searching for clues.

Now, what are the components of a typical neural network? There's neurons, synapses, weights, biases, propagation functions, and a learning rule. Neurons receive input from other neurons and have an activation, threshold, activation function, and output function. Connections are made up of connections, weights, and biases that govern how one neuron transfers output to another. The learning process is all about adjusting the free parameters, like weights and biases.

It may sound like a lot of technical jargon, but once you get the hang of it, it's a fascinating concept to explore and expand on.

AI-Detection Tools: The Joke's on Us

I decided to put a few AI-detection tools to the test, and let me tell you, they are not great. These tools can still be easily fooled if the text is slightly altered.

Take for example, GPTZero. This tool was developed by a Princeton student and it's getting a lot of media attention (we even mentioned it in a previous post). It analyses the "perplexity" and "burstiness" of a given text and gives you a numerical value. The lower the value, the more likely it is that the text was AI-generated. But, when I fed it a ChatGPT-generated text and made a few small changes to the text, the tool completely missed it and said that the text was likely human-generated. The changes I made were small and "human" (like a typo), but they were enough to throw the model off.

The point is, detecting AI-generated text is harder than finding a needle in a haystack. These tools don't expect machines to make typos or use obscure words.

Want to outsmart an AI text detector? Easy peasy, just throw in some typos and some words that even spell check doesn't know and you're good to go.

Hungry Hungry Hippos: Towards Language Modeling with State Space Models

H3 is a new generative language model that has been developed by researchers. It is able to perform better than GPT-Neo-2.7B, which is a large language model that uses transformer architecture, with only 2 attention layers. The researchers have replaced the attention mechanism used in transformer models with a new layer that is based on state space models (SSMs). With the right modifications, H3 can outperform transformer models. H3 also does not have a fixed context length, which means it can handle variable-length input sequences.

Link to the paper, here.

Pic of the day:

Midjourney imagining a python developer:

(source)

That's it from RunTheAI for today.

THANK YOU FOR READING AND SEE YOU TOMORROW, SUBSCRIBE TO STAY UPDATED!