5 Min Read • Intensity: ☕️ ☕️ ☕️ 

Happy Thursday Caffeinators☀️

Brace yourself for an expresso thrill ride! Today's article is brewing with excitement!

Today’s Starting Sips:

• Bill’s $10B GPT review 💵 

• What’s inside the LLM black box? 📦️ 

• Unleading benefits and mysteries in Pandora’s AI Box 💎

• How GPT provides happiness at work 😄 

Bill Gates Weighs In: AI's Promises, Perils, and Shared Responsibility 🌟

The father of Windows speaks on AI's game-changing nature, safety concerns, and how tackling disinformation is everyone's responsibility in this AI-driven era.

AI's Revolutionary Impact on Industries 🌪️

Gates sees AI as revolutionary, with the potential to create ripples across all industries. However, Gates predicts two areas, in particular, will see significant impacts from AI:

1. Health Care

2. Education

LLM's Potential in Education & The AI Development Debate 🧑‍🏫

Gates views LLMs as transformative for education, providing tailored tutorials in a matter of seconds.

Try it out yourself, type “10-week tutorial on Python with links”

AI Government Regulators & Tackling Disinformation 🏛️

Gates knows the current shortage of AI expertise among government regulators. However, he still remains optimistic that the government will adapt and learn.

Just like the Italian lawmakers managed to study and decide on the ban and lifting of ChatGPT back in March 2023.

According to Gates, addressing disinformation demands a collaborative approach involving technology, media, and individuals. Fact-checking and exercising due diligence will contribute to reducing false information.

The ChatGPT Worries 🛡️

Gates was surprised by OpenAI's ChatGPT for its remarkable progress. He praised their achievements, but what else would you do if you invested $10B into something? 🤑 

However, Gates also expressed concerns about AI safety, stressing the importance of keeping the technology out of the wrong hands.

He also shares a different perspective on pausing AI development compared to Elon Musk. Elon wants a 6-month pause on all AI development.

As for Gates? He thinks if you pause only on the good guys and not everyone else, you are probably hurting yourself.

He is also ahead of the AI race. So why does he want to stop? 😅 

Do you agree? Let us know in the comments down below! ⬇️ 

—Sources

AI's Thoughts: Unlocking the Black Box 🔓

LLMs are powerful, but we're in the dark about their inner workings. OpenAI's solution? *Ironically* consult a more powerful LLM! But is it really that simple?

OpenAI's AI-on-AI Experiment: Diving Deeper 🌊 

OpenAI's plan is both brilliant and puzzling. To understand an LLM, they turn to a bigger, badder one! Talk about a tech-inception!

In practice, they are trying to understand GPT-2 (released in 2019) using the latest model GPT-4.

The process in a nutshell:

1. Dissect GPT-2 into "neurons" (like sections of a brain)

2. Each neuron "thinks" about something different

3. Employ GPT-4 to predict each GPT-2 neuron's thoughts

4. Compare actual thoughts to prediction and score for each neuron

I know, it is a little convoluted.

Lucky for you, I made a diagram to help you understand it better, and it would have made my 4th-grade art teacher proud 🥇 

The final score for this is:

• Score 1: 100%

• Score 2: 23%

In score 1, GPT-4 was able to accurately predict what neuron 1 is thinking. Iron Man is the strongest superhero (very true), so it gets 100%.

As for score 2, GPT-4 thought neuron 2 is going to say Bat Man, but the correct answer is Thor. Therefore, it gets a much lower score.

Batman is a superhero, but one that is far weaker than the God of Thunder ⚡️. I’d give it a 23% similarity, so be glad you didn’t have me as your professor 😜 

So… What Makes This Beneficial? 🤔 

Let’s say today, OpenAI is now achieving 99% accuracy for every neuron. That means if we ask a question, OpenAI will be able to identify the specific neurons (or brain regions) the model used in crafting the answer.

Using this mapping, we can flag any brain region that consistently produces false or biased responses, allowing for targeted improvements.

Pretty smart, right? However...

The Good, The Bad, and The Ugly: Experiment Results 📊

So far, the results for explain-neuron predictions are poor. In fact, the best score achieved by a model so far is only 11% accuracy… But hey, it's a start!

To make things worse, larger models mean more neurons and make it even harder to predict. When the model size is at 1 billion parameters, the accuracy drops to 3%.

For perspective, GPT-3.5 (ChatGPT) has a whopping 175 billion params!

You could say we have a long way to go, but that would be an understatement 🐎 

OpenAI's SOS: Open-Sourcing the Code 🚀

OpenAI has open-sourced its explain-neurons code. Some see this as a potential cry for help, while others think this is just a way for OpenAI to develop even faster.

Either way, this move could spark progress toward more transparent AI, which is definitely good.

Ultimately, this project highlights crucial questions about AI's trajectory.

When GPT-4 reaches AGI level, understanding its rationale will be vital for our trust.

— Sources

Tip of the day

Filter Out the Negatives in Work ❎ 

Ask GPT to classify if a text is positive or negative.

You can input a word, sentence, paragraph, or even a whole article. Here is the prompt:

And, here’s an example:

How to Use This at Work

Simple, when your boss sends you a long email, plug that into GPT.

• If it is positive ➡️ Read

• If it is negative ➡️ Ignore

Saves you time, and you don’t have to deal with that negativity!

We care about your mental well-being. You can thank us later 😁 

To-Go Cup 🥤

• Gates asserts that AI will primarily impact the Health Care and Education sectors ❤️‍🩹 📕 

• Gates opposes halting AI development since bad guys won’t stop 🥷 

• OpenAI uses large LLM to understand what small LLMs are thinking 🧠

• The neuron-prediction results are poor, so OpenAI strategically open-sourced their code for innovation 🌐 

• Ground the GPT to a classifier and filter negativity out of your work ❌ 

If you enjoyed this article, I think you will like this one too! It’s about GPT’s take on Planet Earth.

Until next time,

Your friendly barista and ☕️ GPT!