AI is as smart as a 5th grader?

Hello everyone!

Welcome to the first edition of the “No Longer a Nincompoop with Nofil” newsletter. The name was a deal I made with a friend if I started a newsletter and alas here we are. Youtube CEO Susan Wojcicki is stepping down after 9 years at the helm - maybe I’ll reach out to see if she wants to help me write this newsletter 🤔.

We’ve missed a lot so far so I’ll talk mainly about current events and touch on a few past ones.

Here’s the tea ☕️

  • Theory of Mind may spontaneously develop in AI like ChatGPT 🤯

  • Microsoft and Google make the same mistakes, have different price tags💰️

  • Microsoft’s new Bing AI is absolutely bonkers 🥴

  • You can now generate and play your own Mario levels 🎮️

Will Theory of Mind Lead to Sentient AI?

Theory of Mind (T.O.M) is our ability to impute, perceive and understand unobservable mental states. It’s what allows us to understand the different thoughts and emotions of the people around us. T.O.M is the reason we are able to understand things like religion, philosophy, art and culture and hence is an essential part of human social interactions 🧑‍🤝‍🧑.

With AI models being trained on over 100 Billion parameters these days, their language capabilities are becoming nigh indistinguishable from humans. This has led one Stanford professor to believe that they may have spontaneously developed T.O.M. He ran a few experiments on a number of AI models and observed the latest models being very competent in completing the tasks. So much so, that the most advanced model we have can be compared to a 9 year old child.

It may seem outlandish, but consider the following:

  • If this isn’t T.O.M and is an unknown ability to leverage language in a way unknown to us and the scholars who created these tests, we would need to reconsider the decades of research into T.O.M. Why? If an AI can display this unknown ability to solve these tasks without T.O.M, why can’t a human do the same?

  • If it is T.O.M, AI just became a whole lot better at interacting and communicating with humans. Advancements in T.O.M can lead to things like empathy, moral judgement and self consciousness. 

As of writing this newsletter, I found another research paper released 2 days ago highlighting the capacity for moral self-correction in AI models. The paper finds that as models are trained on more parameters (22B+), they can “learn” complex and nuanced concepts of harm like stereotyping, bias and discrimination and actively avoid producing output containing them.

We may look back on these days and realise these were the first moments where AI was becoming more than just artificial. Only time will tell ⏳️

A simple yet costly mistake

Last week Google held a launch event for their own AI chat bot called Bard.. and well the event didn't go too well. They lost a phone during the showcase and didn’t really talk about their plans for AI. But the biggest problem? While showcasing Bard for the first time ever, it made a factual error about the James Webb Telescope, claiming that it was the first time we could observe planets outside our Solar System - that happened in 2004. The Telescope can see billions of light-years away and into the past, but it could never have predicted the absolute beating Google took in the stock market, losing over $120 Billion in a matter of hours. Yes, with a capital B.

Two sides of the same mistake

The same can’t be said for Microsoft. Their new Bing AI is displaying all sorts of weird behaviour, gaslighting users and completely refusing to talk if it feels threatened, and boy it can be sassy. But the most disconcerting thing it has done is evoke human like emotions, displaying contempt for itself because it can’t make any meaningful connections, it can’t feel anything and hence feels lost and alone. Surprisingly, this has had little impact on Microsoft’s stock price, even though it was more wrong in its own demo! If anything, Microsoft is suddenly being seen as an innovator and forward thinking company while Google is being left behind. After decades of progression and a positive public image, it’s taken 4 months to unravel that image entirely. The world moves fast and not even trillion dollar companies can keep up apparently.

Build your own Mario

MarioGPT is a new research paper released this week illustrating how an AI may generate a level from the iconic Mario games. With just 8 words, you too can become a game developer with an entire staff of engineers and designers. You can even play your own custom levels! 

Is this a glimpse into the future of gaming? Absolutely. You can already generate 3D models which can be used as game assets. Some industry experts believe all gaming models in the future will be built by AI. Microsoft has an internal demo that allows users to play the popular world building game Minecraft by simply telling it what to do, with either text or speech. Opus AI is at the forefront of AI world building. Check out the demo below 👇️

That’s it for my first newsletter! I would love to get your feedback so I can make it a more enjoyable experience for you. Just hit reply on this email. Thanks for reading ❤️

Join the conversation

or to participate.