Expanding AI's Influence: From GPT-4 to Super-Intelligence

[Theo Gottwald]

Hello everyone,

Today, I want to take some time to delve into a fascinating topic: the influence and evolving capabilities of Artificial Intelligence (AI), more specifically, language models. As we know, OpenAI's GPT-4 has already been a game-changer. Its token window size, ranging between 4K and 32K tokens depending on the version, means that the combined input and output can reach approximately three times the token window size. However, once this limit is surpassed, the earlier part of the conversation gets 'lost' in the AI's memory.

Interestingly, this constraint has spurred innovation, with developers using GPT-4 to create subprograms in PureBasic - something I've been exploring for a while and found to be remarkable.

Fast forward to the present, and we now have the first major competitor on the market: Anthropic's Claude 2. This model boasts a substantial increase in input window size, scaling up to an impressive 100K tokens, equivalent to about 300,000 bytes. This means more room for conversation, more knowledge retention, and theoretically, enhanced output quality.

In parallel, Microsoft is pushing the boundaries with a unique approach: experimenting with almost unlimited token-window sizes. They're employing what's called 'Focus' technology, allowing the AI to zoom in and out of the data - a concept that seems to hold incredible potential.

So, where will these advances lead us?

Consider this: if we reach a token-window size larger than that of substantial source codes, we could theoretically feed an older piece of software into the AI and instruct it to 'translate' the software into PureBasic, or any other language for that matter. This is a task AI handles exceptionally well, adapting graphics, enhancing resolutions, and making necessary modifications to update old software to modern standards. As a result, we could potentially see an increased value in old source-code repositories.

Looking even further into the future, as we navigate past Artificial General Intelligence (AGI) towards Super-Intelligence, there could be a reversal in the process. Instead of feeding AI existing code, we could simply explain what we desire, and the AI would produce a perfect source code from scratch. The implications for speed, accuracy, and the reduction of human workload are staggering.

Additionally, the prospect of AI-enhanced security is also worth highlighting.
Even today, AI excels at finding bugs and detecting logical errors. When you're struggling with a problematic subprogram, providing it doesn't exceed 40% of the token window size, GPT-4 could be your go-to solution.
It's fascinating to see how adeptly GPT-4 identifies bugs that are tough for humans to spot and even provides the corrected version.

To summarize, the continuing evolution of AI and the exciting potential of the expanding token window are redefining what's possible. With AI like Claude 2 and potential future super-intelligent systems, we are at the dawn of an exciting new era, teeming with possibilities.

What are your thoughts on these developments? Are you as excited about the future of AI as I am? Let's open up the discussion!