I thought it might be a good idea to start this issue with an amazing talk by Andrej Karpathy, considered one of the elite in AI. Andrej worked at Tesla on their Autopilot Vision and is currently at OpenAI as a computer scientist specialising in deep learning and computer vision. He is also a great educator having taught Deep Learning at Stanford University and releasing several Youtube videos on Neural Networks.

There have been x2 AI technologies that have taken the world by storm in the last year. They are both known as GenerativeAI. Essentially, this means they are able to generate something useful based on a prompt. Their power has been realised through refined machine learning techniques on a vast amount of data.

What are they?

1. Diffusion Models - Generate Images & Video - Examples for images are MidJourney, DALLE3 & Stable Diffusion - For Video RunwayML & Pika

2. LLM’s (Large Language Models) - Generate Text - Examples are ChatGPT4, Pi and Claude

The continued development of LLM’s are proving useful and capable of many tasks, from generating code to potentially working together (as a group of ‘agents’) to organise your next holiday or answer your emails.

If you are curious about what exactly LLM’s are, check out this video. You will feel like an expert by the end of it 🤓

OpenAI the creators of ChatGPT (the most advance AI to date) nearly lost it all last week. Here’s a quick fire blow by blow 🔥🤜🤛

If this whole saga hadn’t been reversed the repercussions would have meant that the advancements OpenAI made with ChatGPT and their future rollout of GPT5 would have had a serious time delay, some saying 1 year and potentially putting the whole AI industry back as the talent in OpenAI are transferred.

Thousands of companies rely on ChatGPT as part of their AI services using the LLM for many AI start-up ventures and may have had to fold or hope to pivot elsewhere. Investors also may have been left without much to show and so on.

There were no real solid reasons for the firing of Sam Altman, however it would appear that there may have been a new powerful model at the centre of it. See Q* story below.

In the meantime, as all the OpenAI leadership and staff have stated…

In the fallout as to why OpenAI nearly destroyed itself the most recent speculation has focused on a leaked paper about a model the company were working on called Q*.

Essentially Q* has been described as a LLM that can solve maths problems. Sounds simple enough, but this advancement is surprisingly huge and with wide ranging capabilities that may have got some previous Board Members concerned.

Therefore the Q* speculation has lead to some great discussions about LLM advancements.

Here’s a snapshot:

These recent events have revealed that LLMs may be getting a phase 2 in training that will allow for reasoning through reinforcement learning.

Even Google has hinted that Gemini has been trained similarly to AlphaGo (see AlphaGo story here), essentially using another AI to test step-by-step reasoning to solve or answer questions.

Instead of merely adding the most likely next word in a series, based on a large dataset of human-written text, the focus is now on 'understanding' the actual composition and relationships of words to provide accurate answers.

For instance, if a sentence mentions a 'dog', the model, through reinforcement learning, comes to 'understand' that a dog is a four-legged, furry animal commonly kept as a pet etc.

When asked if a dog lives in a house or a wallet, even just understanding the scale of these options enables the model to answer correctly.

This approach not only enables LLMs to solve problems through reasoned thinking but also, by having an AI provide the feedback loop instead of a human, allows for rapid iterative growth and expansion.

The potential becomes so expansive that it would hint at intelligence levels beyond human reasoning, like how AlphaGo surpassed human competitive play.

It’s foreseeable to imagine an AI with a complex understanding or model of the entire world 🤯.

As mentioned, this method of training could also be applied to other reasoning, like solving maths problems. Apparently Q* is able to solve grade school level maths.

This may sound innocent enough, but with further reinforcement learning rapidly undertaken with another AI, models like Q* could feasibly reach a level that could break encryption and in rapid time. Encryption being one major technology that keeps many things online safe from hacking, like your passwords for instance.

It is likely that these kinds of concerns will come hand in hand with AI’s as they get more powerful and companies like OpenAI seek to attain AGI level AI’s. It also seems feasible that this could also have caused some debate within OpenAI.

Obviously, it’s important that careful consideration be made to understanding these potentials as AI capabilities increase. Personally, I think as many others do this should not lead to a deceleration of the advancements, rather the opposite, as rapid advancements in AI’s continue so too can rapid solutions.

MatthewBerman has done a great breakdown if you are interested in the recent speculation of this next phase in LLM development.

https://youtube.com/watch?v=Z6E41eXStsU

Personally, I think we underestimate the depth and breath of the systems in the programming that our brain and bodies possess in order to be a human. We can replicate aspects of it, but are we really creating new 'beings'.

So much to unravel over these coming years as AI's replicate our intellectual interactions.

The below article gives some great perspective on the current state of AI ‘Agency’

https://vox.com/future-perfect/23971093/artificial-intelligence-chatgpt-language-mind-understanding

We are nearing the end of 2023. This year will go down as a seminal one in the Techno Era and the future AI Revolution.

What will 2024 bring?

Some initial predictions:

1. LLMs will be more than next token guess machines and start to use step by step reasoning.

2. Platforms will begin offering more accessible ways to use AI for the average consumer with varied success. These platforms will be more about subtle integration rather than interacting with AI through an input field in a text-based user interface.

3. Google will release Gemini and offer a vastly more competitive alternative to ChatGPT[5] if not become a leader.

4. Apple will show some, not all, of it's hand with a subtle rollout of AI generated applications that lead to some very powerful future ones in their ecosystem.

What did you think about this weeks Newsletter?

Did we nail it? Can we do better? Want to know more? Hit reply and let us know!

If you like our content and think it might be worth a shout out, please do!

Right click the below button and share the link.