Like everyone else, I’ve been interested in current Generative AI technology since it first came out. I’ve written quite a lot about Generative AI in sales, including a fair number of rants, and some pieces about how best to apply it in a complex sales environment.
A few weeks ago, I had an opportunity to speak with someone who knows a great deal more about the reality of AI than I do. Lars Albertsson is a data engineer in Stockholm and the founder of Scling, a company that helps clients get more value from their data. He filled us in on the practical history of AI, through the course of three big waves leading to the current wave, and provided some expert insight into what we can (and can’t expect) from this current wave.
Before we had actual Artificial Intelligence, we had dreams of Artificial Intelligence. Leonardo da Vinci envisioned a “calculating machine” for which he drew up complicated sketches. Philosophers and fantasy writers imagined small humanoid contraptions called “homunculuses” that could do basic housework or other mundane tasks.
Prior to the 1940s, however, we had no technology capable of actually producing anything like what authors, philosophers, and artists could imagine.
Early research into computational technology began roughly in the 1940s, and instantly people’s dreams turned to AI.
“When the first computers began to be built,” says Albertsson, “the idea immediately turned up that we might use this to do things that humans can do.”
Researchers, scientists, and thinkers started imagining all the ways computers could imitate humans. In 1950, Alan Turing published a paper that began with the words, “I propose to consider the question, ‘Can machines think?’”
Out of this question grew the famous Turing Test, which suggests that in order to consider a computer “intelligent,” it must be able to fool a human into thinking that it is human.
With computing technology still in its infancy, we had no way to put any of these questions or hypotheses to the test, so the idea of artificial intelligence remained firmly in the realm of imagination for another forty years.
In the 1980s, computer technology advanced enough that it was possible to begin applying theory in practice and develop the first wave of practical AI.
“It was then called expert systems,” says Albertsson. “The idea was that you take an expert in a field, and you translate his knowledge, and map his brain into a computer program to try and express all that he knows and the reasoning that he uses.”
This was the seed of the rules-based systems that were the first practical AI engines. Rule-based systems use knowledge and processes that humans create, and programmers program them into a computer that then executes on those rules.
A useful example is in financial fraud detection. Until the mid 2010s, most financial institutions used rules-based systems to determine when to freeze or notify an account of potential fraudulent activity. This was a major improvement over having individual humans monitor accounts. It took the knowledge of fraud experts, and told the programs to flag accounts that shared certain traits that human experts knew were characteristic of fraud.
One thing rules-based AI was not useful for was languages. You can only go so far with rules-based AI when trying to translate or understand language.
Key point: In rule-based systems, every rule in the program comes from the mind of the programmer.
In the mid-2000s, Google figured out an economically viable way to store every bit of data that was interesting to them in a structure that was possible to search, interact with, and catalog. This was called “Big Data” and was big news at the time. Google published what it was doing, and then people created open source platforms like Hadoop to play with the idea.
“With Hadoop,” says Albertsson, “we went industrial in terms of data processing. The humans stopped working with the data, and started programming the processes to process the data. We weren’t touching the data ourselves, only the processes. If something was wrong, we added a process to measure things and added a third process to correct the things.”
Before big data, it was very expensive to store large amounts of data, which made it impossible to learn from large sets of data. With this new ability to store data, there came a new ability to analyze it.
This was called “machine learning.”
Now, for instance, computers could analyze the data around people who buy a particular item, and see what other people who buy that particular item also tend to buy. This enabled “you might also like” engines to pop up in online shopping, for instance.
It also enabled the next wave of fraud detection software in the finance industry. Now, instead of simply programming rules about how to guess if something is fraud, the big data engines could analyze large amounts of data around actual fraud and create its own rules based on “people who commit fraud also do these things.”
Machine learning is also better at language translation but still not as good as the current wave. It’s important to note, however, that however “smart” machine learning appears to be, it is still based in what humans program and define.
“In popular science it’s sometimes expressed as the system magically learns itself,” says Albertsson. “But you cannot just throw a pool of data at something in the script and magic happens. There’s always a framework set up by a programmer.”
Key point: In a machine-learning system, there’s a combination of decisions made from the mind of a programmer and parameters created in the algorithm learned by the machine from the data.
Prior to 2015, a number of people had discussed the possibility of modeling computers on the neural networks of the human brain. These ideas were largely considered useless and outlandish until around 2015. Google, again, got interested and started building computer systems on this basis.
The way Deep Learning worked that was different from previous versions of practical AI was that it processed in layers. Consider the task of orienting photographs the right direction. The first layer of processing knows, from massive amounts of data showing photos that are already oriented “correctly,” that a large patch of blue in a photo probably belongs at the top. This has the end result of orienting the sky in many photos so that it is “up.” A single layer of processing might only be able to do this much, leaving a large number of photos unsorted or incorrectly sorted. A second layer, however, may understand the rough shape of a human, again, based on large amounts of photos that are correctly oriented. It can then turn photos right side up based on that.
If you have twenty, a hundred, or several hundred layers of processing, you can process incredibly complex patterns. That’s deep learning.
Deep learning is excellent for image detection, object analysis, and video analysis.
Key point: Humans are providing data and datasets showing “correct” and “incorrect” options, and the computers are learning to complete the task without direct human processes or oversight.
Today’s wave of AI technology, like the other waves, does some pretty impressive things. Because it is so powerful at generating language, it seems very smart to us. This may explain why, in the public consciousness, it is the first wave to be referred to almost universally as AI.
Another thing that is different about this wave is that it is readily available and usable by the public, not just by technical experts.
Current “AI” technology originated from a leap-forward realization that graphics cards could be used to process all kinds of data, not just graphic data. They were powerful enough to map a version of neural networks on them.
This led to more effective use of a technology called Embedding, which was invented in the 2010s. Embedding, explained Albertsson, is a way to express the meaning of something mathematically. It works by grouping things that are similar together physically in space. You can do this with language, by grouping words and phrases that are similar in meaning with others that are similar in meaning.
This enabled the Large Language Models of today, the generative language “AI” that has been the hottest tech topic of the year for several years in a row. It looks impressive because it sounds like a real human talking. In fact, AI chat bots have come to easily pass the Turing Test imagined in the 1950s to define whether a technology is intelligent or not.
But it’s not.
“If you look at them technically,” explains Albertsson, “a Large Language Model is doing a very, very simple task. It’s simply asking, what do you think the next word should be.”
In other words, it’s “guessing” what word should come next in any given conversation or sentence. It’s just very good at it because the data it’s fed is organized in a manner that groups things according to meaning.
According to Albertsson, LLMs represent the first wave of “AI” that has very little practical, useful, beneficial application.
Not that it can’t do anything. It is good at summarizing, and good at helping people generate ideas. But it can’t understand context, doesn’t know truth from fiction, and isn’t intelligently analyzing or creating meaning. It’s just guessing, which is why so many AI engines give false information or “hallucinations.”
Here are a few things Large Language Models won’t do, according to Albertsson:
Unfortunately, according to Albertsson, it is useful for “nefarious” purposes: Spreading misinformation, causing confusion, and undermining people’s sense of reality.
Albertsson, paraphrasing Bill Gates, said that people tend to overestimate the short term benefits of new technology, and underestimate the long term benefits. Have we overestimated the short term benefits of Generative AI? Albertsson thinks so. But have we underestimated the long term benefits?
Albertsson thinks that we actually have not. He says there’s a limit that every new technology hits, a ceiling of what’s possible. In the age of da Vinci, the ceiling was any practical technology that could execute on the idea. Today, we don’t yet know where the ceiling is, but we know that what we’re calling AI isn’t so much intelligent as very good at guessing what will make it seem intelligent.
So, what’s next? I think we’ll see basic LLM capabilities, such as summarizing and writing text, become part of operating systems from Microsoft, Apple, and Google (they’ve already been fast movers on this.) Hopefully, we’ll see more tailored AI applications from niched business application vendors as well, instead of duplicate ways to summarize and create content. The CEO of Microsoft, Satya Nadella, thinks that we’ll soon see user interfaces disappear altogether in favor of AI agents that respond to conversational prompts.
What about you? What do you predict for the next wave of AI?
George is the founder & CEO of Membrain, the Sales Enablement CRM that makes it easy to execute your sales strategy. A life-long entrepreneur with 20 years of experience in the software space and a passion for sales and marketing. With the life motto "Don't settle for mainstream", he is always looking for new ways to achieve improved business results using innovative software, skills, and processes. George is also the author of the book Stop Killing Deals and the host of the Stop Killing Deals webinar and podcast series.
Find out more about George Brontén on LinkedIn
From north to south, east to west, Membrain has thousands of happy clients all over the world.