Why GPT-4 is such a big deal ?

I read a trilogy of science fiction novels where an alien attack is imminent on earth after 200 years.  To face them humanities make major scientific advances. They build nuclear fusion and propulsion. But they fail to build Artificial General Intelligence (AGI). 

Many scientists consider general intelligence as an outcome of quantum phenomena 

https://bigthink.com/hard-science/brain-consciousness-quantum-entanglement/

It is not clear how our ability to reason emerges from our brains. Scientists know how our neurons are connected in our brains.  They know the structure of neurons. But it is not known how the connection between 100 billion neurons in our brain allows us to appreciate art, go to the moon, display emotions, and do so many other things. 

Scientists don't know the recipe for intelligence. So they did the next best thing i.e copying our brain architecture. Neural Networks were proposed way back in 1958 by Rosenblatt. But scientists quickly realized that our brain is much more complicated. Our brain contains around 100 billion neurons. 

Our CPU chips still don't contain as many transistors as the human brain. But GPUs have allowed scientists to train increasingly larger machine learning models. 

GPT-3 contains 175 billion parameters. It was one of the first models to contain parameters comparable to the number of neurons in the human brain. GPT-4 contains an even greater number of neurons. 

GPT-3 displayed impressive writing abilities. But it was not able to make a distinction between truth and fiction. It did not display logical reasoning abilities. 

GPT-3 and GPT-4 are trained only to predict the next word in the corpus of text in their training data. That GPT-3 did not display logical reasoning abilities was only entirely expected. 

ChatGPT (GPT 3.5) and GPT-4 are able to show an understanding of the world. GPT-4 shows the ability to do logical reasoning. It passed several difficult exams with a top percentile.



I tested GPT-3 , Chatgpt (GPT 3.5) and GPT4 on several puzzles. We can see GPT 4 can handle difficult puzzles

GPT 3


ChatGPT


GPT 4



Another puzzle

GPT 3


ChatGPT


GPT-4


Can you see why GPT 4 is correct in above puzzles. 

I have taken Chatgpt Plus subscription. I  have also tested GPT-4 on CAT verbal reasoning questions. It was able to get several of them right. In the office, it is helping me with coding. GPT-4 shows the proficiency level of an experienced coder. Unless you are writing a very long and complicated code, GPT-4 will be able to code it for you. 

Further in the next phase Open AI is going to enable plugins for chatgpt. Like apps for androids or IOS, thousands of plugins would be created for chatgpt. Plugins that allow chatgpt access to web search and code interpretation are currently under testing. 

With its proficiency in English and logical reasoning, GPT would be able to disrupt work in many industries. Poor drafting skills would not be an acceptable excuse anymore for bad written communication. 

Already GPT-5 is under training. Open AI is training  GPT 5 on hardware worth 250 million dollars. If it achieves further improvement it would come closer to the goal of creating a neural network that can write its own code.

Already GPT4 can write code for most basic machine-learning problems. Once GPT future versions are able to write code for their own training, progress towards AGI will be very rapid. 



Comments

Post a Comment

Popular posts from this blog

Your Inconsistency is a Feature, Not a Bug

Image
We are living in the Golden Age of Guilt. You know the feeling. You wake up, check your phone, and are immediately bombarded by the "High-Performance" brigade. The 5 AM Club. The Atomic Habit builders. The Tech Bros meditating in ice baths. They all have the same message: Consistency is King. If you aren't grinding 24/7, if you abandoned that hobby you started last month, or if you can’t focus on a spreadsheet for four hours straight, you are told you are broken. You lack "grit." You need to buy another book to fix your "laziness." But what if I told you that the self-help industry is gaslighting you? What if your inconsistency, your fickleness, and your refusal to stick to the plan aren't moral failings—but the very biological features that keep our species alive? 1. The "Death Spiral" of Perfect Discipline We tend to think of discipline as the ultimate virtue. But in nature, perfect discipline is often a suicide pact. Take the Army Ant....
Read more

LLMs have already taken over the world. Dont be left behind.

Image
You’ve probably already read about how Large Language Models (LLMs) like ChatGPT, Gemini, and others are growing in popularity. Many say they are “taking over the world,” but the truth is even more striking—their impact is already massive, though in ways we often cannot easily quantify. Understanding Tokens At the core of this impact is the concept of the token . A token is a unit of text measurement used by companies to calculate usage and costs. Roughly speaking, one token equals about three-fourths of a word. For example: A page of text contains about 225 words , which is nearly 300 tokens . Across the world, there are an estimated 130–190 million books , which equals a total size of 15–20 trillion tokens of text content. Now compare that to social media: every day, around 500 million tweets are posted. In one year, that’s roughly 200 billion tweets , amounting to about 5 trillion tokens . Beyond Books and Tweets: Other Human Text Reservoirs Books and tweets are only two exampl...
Read more

Common theme across important essays

 So I asked chatgpt to identify common theme across important essays in life, economics and social science. This is what it said  Below is a formal, analytical synthesis of all the essays you requested summaries for. By examining arguments across sociology, psychology, leadership, economics, AI, creativity, and science , we can identify the single most repeated theme and the supporting sub-themes that appear again and again. The Most Repeated Theme Across All Essays Excellence, progress, and human achievement emerge from small, structured, consistent actions rather than from dramatic talent, inspiration, or centralized control. This theme appears everywhere — whether the subject is elite performance (Chambliss, Ericsson, Csikszentmihalyi), leadership (Greenleaf), learning (Dweck), creativity (Graham), complex systems (Hayek, Read), sociology (Putnam, Granovetter), or even scientific progress (Asimov, Feynman). Across fields, the message is consistent: Big outcomes ar...
Read more