Our First Look at GPT-4

Today we delved into the world of GPT-4 (Generative Pretrained Transformer). We’ve been waiting for this iteration since sampling the first version, released by OpenAI back in November. 

To give you a glimpse into why we are so excited to give this a try, ChatGPT was trained using 570GB of datasets including the internet, books and articles. For context, this equates to reading around 300 billion words or 175 billion parameters.  

GPT-4 on the other-hand, is another beast entirely. Rumored to have been trained with over 100 trillion parameters, this makes GPT-4 over 500 times as powerful as ChatGPT. Additionally, it will have the ability to process and recognize images. For example, it will be able to process a picture of a chicken laying an egg and give you information on trajectory and the likely impact of that egg – recognizing not only that there is a chicken, but that it has real world interactions and implications. This feature has been delayed however, with OpenAI stating concerns of the system being abused. ChatGPT Plus users will have a limited access to this mode eventually though. 

To say that this is game changing is a drastic understatement. Admittedly, it still has its limitations. Even after consuming that depth of information, it still has issues with reasoning. At its root, it is still a complex database, providing responses based on datasets. 

From its own interface, certain levels of ChatGPT perform differently: 

How GPT-4 Interprets Commands

We wondered how GPT-4 would take a few commands – so we provided it with this prompt:

“Write a poem in the style of Edgar Allen Poe on the subject of humanity and the progression of our use of technology.”

Here is what we received:

Whispers ‘Midst the Gears of Time

In shadows deep, where spirits dwell,
A tale of human strife I tell,
Of mortal coils and worldly gains,
And how our heart’s blood marks the chains.

Through the mists of time, I see,
A race consumed by industry,
With every cog and gear a plea,
To harness powers once set free.

A spark ignites, ignites the mind,
In search of truths we long to find,
Machines and wires, our hands entwine,
The darkened path of fate we bind.

Oh, what cruel cost we pay,
As we forge forth in disarray,
Our souls embroiled in cold dismay,
And shadows cast by our own sway.

In halls of glass and gleaming steel,
We dare not pause, nor ever feel,
For time, relentless, spares no wheel,
And grinds us ‘neath its iron heel.

Yet in this maelstrom, hope remains,
A whispered prayer, a light sustained,
For in the heart of man, untamed,
Lies dormant strength, unchained, unnamed.

The raven’s call, the tolling bell,
Must serve not as our deathly knell,
But as a clarion, a cry to quell,
The darkness borne of our own spell.

Awake, O man, to endless skies,
To break the bonds of our demise,
And through the embers, shall we rise,
To claim our place ‘neath starlit eyes.

Someone’s English teacher is likely rolling in their grave as we speak. We’ll admit though, it’s a compelling poem, no?

We tried a few more prompts and were enthused with our results, but will share an interesting occurrence that took place, much to our own surprise. We found that we had started our experimentation with very specific and direct prompts, but by the end, we were using proper grammar… and saying please. We hadn’t noticed at first but when we caught ourselves being polite to this AI, some self-discovery took place. We stopped thinking about it as just a tool set and started being thankful for its help. We humanized it. Someone in psychology write this down, we foresee an amazing thesis about this in your future.

Now that we stand back and contemplate our actions, we find that we are pondering larger topics. This program would never pass the Turing test but it has certainly made us think. Have you made any discoveries using Chat GPT or other AI generative tools? We’d love to hear your stories – or poetry.