The title of this two-part piece suggests we are five years from artificial intelligence (AI) taking over the world, or at least helping us meet the challenges of this century and beyond through its benign magnificence or malevolence. The pace of new releases of AI coming from Silicon Valley and China seems frenetic. Why are the technology giants in such a hurry? What are the consequences of this haste for society?
An AI Lexicon
AI comes with its own terminology, including tokenizers, tokens, algorithms, machine learning (ML), Large Language Models (LLMs), Artificial General Intelligence (AGI), Agentic AI and Artificial Super Intelligence (ASI). So what do all these mean?
Tokenizers convert raw text into tokens. Algorithms are programs that process these tokens. Without tokenizers, AI doesn’t have a clue about what it sees.
Tokens from English can be as small as punctuation or a portion of a word. A phrase containing three to four words in English could consist of three or more tokens, depending on word complexity.
LLMs are ML algorithms. MLs rely on computing infrastructure replicating a facsimile of our brains. These are called neural networks. LLMs train on millions of tokens. The tokens they see, they transform into machine syntax. Words turn into bits and bytes and get mapped and embedded into the LLM’s syntactical map. The more the LLM sees, the more it can begin to transform the mass of tokens upon which it is trained. Exposure leads to pattern detection in syntax, word associations, phrase structure and grammatical rules.
Thus, LLMs become AGIs. Does an AGI understand language like you and me? No. AGIs are trained, complex pattern-recognition systems that generate coherent answers to our queries. Their output is a product of the material upon which they have been trained.
My first exposure to OpenAI’s ChatGPT showed me just how limited it was when asked about current events. When queried to explain something appearing in the news cycle, it replied that my request was beyond its training.
Today’s AGIs, for the most part, still lag behind live news streams and, as a result, require frequent periodic retraining. A few, however, are hybrids, incorporating a live retrieval layer that can mine more current content in near real time.
Some are Agentic AIs, trained on a smaller set of tokens and serving a specific purpose. An Agentic AI could be an office assistant. It could help organize daily routines, plan trips, and book appointments. A simple example of the differences between an AGI like ChatGPT and an Agentic AI follows:
Ask ChatGPT to draft an email, postponing an office meeting, and it confines itself to the specific task and sends it to you.
Ask the Agentic AI office assistant to do the same, and it writes and sends the email to an attendee list, polls the list to find a convenient alternate date for everyone to attend, cross-puts the date in your calendar and notifies you.
For ASI, we will get to that further down in this article.
Why I Use AI and in Particular Perplexity.ai
I readily admit to using AI to help me research articles I write. I have tested several AGIs and work most often with Perplexity.ai. It is a hybrid AGI, an answer engine that resembles a smart search engine rather than an LLM, like Google Search on steroids. Perplexity.ai answers questions, providing live web search, citation-backed summaries, links to source information and other LLMs.
Ask Perplexity.ai to describe itself, and it states it is strongest at fast research summaries and current facts, more useful to journalists and policy wonks, and weakest as an AGI chat companion.
If you seek long conversational answers to queries, Perplexity.ai suggests trying ChatGPT, Gemini or Claude. For technically complex questions and answers, it points to DeepSeek. Here is its simple rule of thumb:
- Perplexity.ai to find and verify;
- ChatGPT/Claude to think and write;
- Gemini as a multimodal producer of text, images, audio, video output and more;
- DeepSeek, when seeking technically detailed information;
- Llama for building custom Agentic AIs.
Why We Are Building So Many Data Centres
It seems that data centres are sprouting everywhere these days. Why?
Data centres provide the computing power and resources LLMs use when training and when answering human queries. The billions and trillions of calculations LLMs do require massive parallel computing, which means thousands of server racks, tons of data storage and networking, enough electricity to run the lights and heating of small towns, and energy to keep the computers and centres cool.
Millions of daily queries keep the data centre hardware very busy. Not all LLMs, however, need data centres. An LLM can train on small amounts of data to serve local application requirements. An LLM can be small enough to run on a smartphone or laptop.
Among the American heavyweight LLMs like OpenAI’s ChatGPT, Anthropic’s Claude, Meta’s Llama, Google’s Gemini, X’s Grok, and Microsoft’s CoPilot and others, there are varying energy requirements. Gemini is more efficient per query, for example, than ChatGPT. China’s DeepSeek, an open-source LLM, is harder to compare against American AI competitors. Initially, it may have a much lower per-query energy requirement, but for more complex tasks, it may turn into an energy hog.
To understand the energy price we pay for current AI LLMs, a good source to go to is the Stanford AI Index. It is an independent organization that tracks the unit cost of AI per token as well as other data points associated with the technology. According to it, token processing costs have been dropping precipitously, from US$ 20 per million tokens to $0.40 in the last 24 months. This enormous cost decline in token processing has Sam Altman of OpenAI stating:
“We are past the event horizon. The takeoff has started. Humanity is close to building digital superintelligence.”
Altman isn’t alone in prophesying a future of data centre geniuses.
Dario Amodei, CEO of Anthropic, sees superintelligent AI happening as early as this year or next. He states:
“After powerful AI, we will make all the progress in biology and medicine in a few years that we would have made in the whole 21st century.”
When AI Will Solve Everything: From AGI to ASI
Elon Musk, CEO of XAI, predicts artificial general intelligence (AGI) will emerge before the end of this year. What makes him think that? Self-improvement is the answer. Current AIs are learning to design better versions of themselves at an increasingly faster pace.
Peter Diamandis of XPrize and Singularity University predicts that AGI will lead to artificial superintelligence (ASI) by 2031.
ASI will combine the intellectual output of all of humanity. The implication is that ASI will have knowledge built on the sum of our digital legacy. It will mean the cost of accessing this knowledge will become dirt cheap. Every founder, researcher, and dreamer will suddenly command the knowledge of a thousand PhDs for the price of a lunch.
Science fiction will become reality with ASI compressing “centuries of discovery into a handful of years.” ASI will help us develop new materials, drugs, save lives, extend lives, and turn last year’s impossible tasks into today’s realities. I leave you with Peter’s ASI prophecy:
“The bottleneck on progress flips. For all of human history, the scarce resource was brainpower (i.e., enough brilliant minds, enough time to chase down a hypothesis). When that becomes infinite and nearly free, the constraint moves to the physical world: how fast can we run the experiments, build the reactors, fabricate the chips? Atoms become the bottleneck, not ideas.”
So ends the promise of AI. But what about the darker, apocalyptic course of this technology? Stay tuned for Part 2.
