Origins of AI Agents

Ted Werbel

Ted Werbel

June 20, 2024

•27 min read

AI AI Agents

Before we get started...

Defining AI agents can be a bit tricky since useful implementations of them and relevant design patterns are so new to the engineering world. In this article, I'll try my best to guide you through a brief history of AI, how key intellectuals of the past perceived the future of these capabilities and how it has all evolved into this concept of "agents".

This is the first part of a three part series. In this first segment, I will share with you a brief history and timeline of how AI has evolved since inception. In part two we will discuss cognitive architecture and emerging design patterns from the latest research and finding amongst big tech and academia. Eventually in part three we will discuss some of the most impactful design patterns and walk through practical code implementations of AI agents using a modern technology stack (using Langgraph, Langchain, Autogen and CrewAI).

Slowly but surely we will transition away from theory and take things into practice. By starting this series with a brief history of AI, I hope to give you a holistic view into how this technology has evolved over time such as to help guide our understanding of where it may be heading in the future.

10 Million BCE - The Dawn of Biological Intelligence

400 BCE to 1500 AD - Autonomous Machines

The concept of artificial intelligence has roots stretching back thousands of years, with ancient philosophers pondering about the mysteries of life and death. In those early days, inventors conceived of intricate devices known as “automatons” - mechanical machines capable of moving independently without human intervention. The term “automaton” itself is derived from ancient Greek, meaning “acting of one’s own will.”

One of the earliest documented automatons dates back to 400 BCE, describing a mechanical pigeon created by a friend of the philosopher Plato. Centuries later, around 1495, the legendary Leonardo da Vinci designed one of the most renowned automatons of his time.

Leonardo da Vinci's fascination with automatons is well documented and his personal notebooks were full of ideas around this concept. Amongst his diverse creations, he imagined everything from a hydraulic water clock to an elaborate robotic lion. However, the most extraordinary of his designs was for an artificial man, envisioned as an armored Germanic knight.

^{Reconstruction of Leonardo’s mechanical knight. Photo: Erick Möller. Leonardo da Vinci. Mensch- Erfinder – Genie exhibit, Berlin 2005.}

Da Vinci’s detailed sketches of the knight's key components reveal a sophisticated system. Powered by an external mechanical crank, the knight was designed to use cables and pulleys to sit, stand, turn its head, cross its arms, and even lift its metal visor. Although no complete drawings of this automaton have survived, there is compelling evidence suggesting that Da Vinci may have constructed a prototype in 1495 while under the patronage of the Duke of Milan.

Fast forward to 2002, when NASA roboticist Mark Rosheim decided to bring Da Vinci’s vision to life. Utilizing the Renaissance master’s scattered notes and sketches, Rosheim successfully created a functional version of the 15th-century automaton - further cementing the idea that Da Vinci was indeed a pioneer in the field of robotics.

While the notion of autonomous machines is undeniably ancient, the next segments of this article will concentrate on the 20th century—a pivotal era when engineers and scientists began making significant strides towards the advanced AI we recognize today.

1920 to 1960 - Can Machines Think?

Between 1900 and 1950, the concept of artificial humans captivated the media, sparking the imagination of scientists across various fields. This period saw an increasing curiosity about the possibility of creating an artificial brain and inventors of the time even developed early versions of robots. The term, "Robot", was coined in 1921, emerging from a Czech play. These early robots were relatively simple, often steam-powered, and some had the ability to make facial expressions and even walk.

Key milestones during this era included:

1921: Czech playwright Karel Čapek introduced the world to the term “robot” in his science fiction play “Rossum’s Universal Robots,” which depicted artificial people.
1929: Japanese professor Makoto Nishimura constructed Japan’s first robot, named Gakutensoku.
1949: Computer scientist Edmund Callis Berkley published “Giant Brains, or Machines that Think,” a book that drew comparisons between the latest computers and human brains.

More scientific concepts of modern AI however can be traced back to the mid-20th century, with key contributions from several pioneers in artificial intelligence and computer science. Alan Turing, the famed computer scientist, published an article called "Computing Machinery and Intelligence" where he posed the question:

"Can machines think?"

Rather than defining machines and thinking, Turing proposed an innovative approach to this question, inspired by a Victorian parlor game known as the imitation game. This game involved a man and a woman in separate rooms communicating with a judge through handwritten notes. The judge’s challenge was to identify who was who, complicated by the man’s attempt to impersonate the woman.

Artistic rendition of imitation game ^{Generated with Midjourney: photorealistic image of a victorian era parlor with three seperate rooms, rooms are divided by thick wooden walls, in the left room is man sitting and smoking a cigar, in the center room is victorian era judge examining a slip of paper, in the right room is women smirking and dressed in elegant victorian era dress, playing the classic parlor game called imitation game, walls with doorways seperate each of the three room --ar 7:2 --v 5.0}

Turing adapted this game into a thought experiment where one participant was a computer. If the computer could play the imitation game convincingly enough that the judge couldn’t distinguish it from a human, Turing argued that it would be fair to consider the machine intelligent.

This experiment, known as the Turing test, remains one of the most famous and debated concepts in AI. The appeal of the test lies in its promise of a clear-cut answer to the philosophical question: “Can machines think?” If a computer passes the Turing test, the answer would be affirmative. Philosopher Daniel Dennett noted that Turing’s test aimed to end philosophical debates by suggesting that anything passing this test would undoubtedly possess the essence of thinking.

However, a deeper reading of Turing’s paper reveals nuances that introduce ambiguity, suggesting Turing intended it more as a philosophical challenge than a practical test.

In “Computing Machinery and Intelligence,” Turing simulated a future scenario with an imagined intelligent computer. The dialogue went as follows:

Q: Please write me a sonnet on the subject of the Forth Bridge. A: Count me out on this one. I never could write poetry. Q: Add 34957 to 70764. A: (After about 30 seconds) 105621. Q: Do you play chess? A: Yes. Q: I have K at my K1, and no other pieces. You have only K at K6 and R at R1. It is your move. What do you play? A: (After a 15-second pause) R-R8 mate.

Interestingly, the computer makes an arithmetic error—the correct sum is 105721, not 105621. It’s improbable that Turing, a mathematical genius, left this mistake accidentally. It’s more plausible that this was an intentional “Easter egg” for astute readers.

Turing hinted elsewhere in his article that the error was a programming trick to deceive the judge. He understood that readers noticing the mistake would likely believe they were interacting with a human, since a machine would not be expected to make such a basic error. Turing suggested that machines could be programmed to deliberately introduce mistakes to confuse the interrogator.

Although the notion of using errors to imply human intelligence might have been perplexing in 1950, it has since become a design strategy in natural language processing. For example, in June 2014, a chatbot named Eugene Goostman was claimed to have passed the Turing test. Critics pointed out that Eugene’s success was due to an embedded cheat: Eugene portrayed a 13-year-old non-native English speaker, so his grammatical mistakes and incomplete knowledge were perceived as youthful naivety rather than flaws in language processing.

Similarly, when Google’s voice assistant, Duplex, amazed audiences with its human-like hesitations, many people noted that these weren’t signs of genuine thinking but rather hand-coded pauses designed to mimic human speech. Both instances reflect Turing’s idea that computers can be designed to make simple mistakes to appear more human. The developers of Eugene Goostman and Duplex understood that a superficial appearance of human fallibility can be convincing.

Perhaps the Turing test does not evaluate machine intelligence but rather our willingness to accept it as intelligent. As Turing himself remarked: “The concept of intelligence is emotional rather than mathematical. Our perception of intelligent behavior is influenced by our own mindset and experience as much as by the object’s attributes.”

Turing seemed to suggest that intelligence isn’t a substance to be programmed into a machine, but a quality constructed through social interaction.

The Birth of Artificial Intelligence

Alongside pioneers like Alan Turing, Marvin Minsky, Allen Newell, and Herbert A. Simon - John McCarthy, a famed mathematician and computer scientist, had a vision that helped shape the very foundation of AI. The summer of 1956 marked a pivotal moment when McCarthy, Minsky, Nathaniel Rochester, and Claude E. Shannon introduced the term "artificial intelligence" in their proposal for the Dartmouth Conference. This event wasn't just a meeting of minds... it was the dawn of a new scientific discipline.

AI researchers at Darthmouth in 1956 ^{August 1956. From left to right: Oliver Selfridge, Nathaniel Rochester, Ray Solomonoff, Marvin Minsky, Trenchard More, John McCarthy, Claude Shannon.}

But it was in 1958 that McCarthy truly showcased his visionary prowess with the introduction of the "advice taker." This concept wasn't just ahead of its time - it was a leap into a future where machines could understand and reason about human advice. The idea of the advice taker wasn't merely an academic exercise but rather a proposition that challenged the boundaries of what machines could achieve. McCarthy imagined a system that could accept instructions in natural language, reason about them, and execute actions based on that reasoning.

This groundbreaking concept laid the groundwork for what was eventually recognized as logic programming. Logic programming - a direct descendant of McCarthy's early ideas - became a crucial part of early AI. It uses formal logic to represent facts and rules about a problem domain, allowing computers to infer conclusions and solve complex problems. Languages like Prolog emerged from this foundation, enabling machines to handle intricate reasoning tasks.

Logic Programming

The late 1960s and early 1970s saw fervent debates about the merits of declarative versus procedural representations of knowledge in artificial intelligence. Proponents of declarative representations, such as John McCarthy, Bertram Raphael, and Cordell Green at Stanford, and John Alan Robinson, Pat Hayes, and Robert Kowalski in Edinburgh, championed the idea that knowledge should be expressed in a manner agnostic to the order of operations. Meanwhile, procedural representation advocates, led by Marvin Minsky and Seymour Papert at MIT, focused on the sequencing of operations.

From this proceduralist paradigm, Carl Hewitt at MIT developed Planner, the first language to feature pattern-directed invocation of procedural plans from goals and assertions. Planner’s most influential variant, Micro-Planner, was implemented by Gerry Sussman, Eugene Charniak, and Terry Winograd. Winograd notably used Micro-Planner to create SHRDLU, a pioneering natural-language understanding program. To enhance efficiency, Planner employed a backtracking control structure, ensuring only one possible computation path was stored at a time. This innovation led to the creation of several derivative languages, including QA4, Popler, Conniver, QLISP, and the concurrent language Ether.

In Edinburgh, Hayes and Kowalski sought out to merge the logic-based declarative approach with Planner’s procedural methods. Hayes developed Golux, an equational language enabling different procedures by modifying the theorem prover’s behavior.

Simultaneously, in Marseille, Alain Colmerauer was advancing natural-language understanding by using logic to represent semantics and resolution for question-answering and in the summer of 1971, Colmerauer invited Kowalski to Marseille. Together, they discovered that the clausal form of logic could represent formal grammars and that resolution theorem provers could be utilized for parsing. They observed that hyper-resolution behaved as bottom-up parsers, while SL resolution functioned as top-down parsers.

Let's walk through an example of how to do logic programming today with Prolog and Python. This example demonstrates a simple family relationship problem. We will start by defining some family relationships and then query those relationships using logic programming.

prolog_example.py

from pyswip import Prolog
 
# Create an instance of Prolog
prolog = Prolog()
 
# Define family relationships
prolog.assertz("parent(john, doe)")
prolog.assertz("parent(doe, jane)")
prolog.assertz("parent(jane, bill)")
prolog.assertz("parent(bill, alice)")
 
# Define a rule for grandparent
prolog.assertz("grandparent(X, Y) :- parent(X, Z), parent(Z, Y)")
 
# Define a rule for ancestor
prolog.assertz("ancestor(X, Y) :- parent(X, Y)")
prolog.assertz("ancestor(X, Y) :- parent(X, Z), ancestor(Z, Y)")
 
# Query the Prolog database for grandparents
print("Grandparents:")
for result in prolog.query("grandparent(X, Y)"):
    print(f"{result['X']} is a grandparent of {result['Y']}")
 
# Query the Prolog database for ancestors
print("\nAncestors of alice:")
for result in prolog.query("ancestor(X, alice)"):
    print(f"{result['X']} is an ancestor of alice")

^{Generated with ChatGPT for reference - run at your own risk!}

To set things up, we started by creating an instance of the Prolog interpreter using Prolog(). We then asserted some facts about parent-child relationships using prolog.assertz(). From there, we defined a rule for grandparent/2, which states that X is a grandparent of Y if X is a parent of Z and Z is a parent of Y. And then, we also defined a recursive rule for ancestor/2, which states that X is an ancestor of Y if X is a parent of Y or if X is a parent of Z and Z is an ancestor of Y.

If you run the script, you should see the following output:

Grandparents:
john is a grandparent of jane
doe is a grandparent of bill
jane is a grandparent of alice

Ancestors of alice:
john is an ancestor of alice
doe is an ancestor of alice
jane is an ancestor of alice
bill is an ancestor of alice

This example illustrates the basic principles of logic programming: defining facts and rules, and querying those rules to derive new information.

McCarthy's advice taker and the advent of logic programming was a precursor to the more sophisticated AI systems folks are building today. It envisioned machines capable of understanding nuanced instructions and engaging in intelligent decision-making. This idea didn't just inspire future research... it ignited a FIRE that would fuel decades of innovation in AI.

And so while the story of these early pioneers of AI is truly a testament to the power of visionary thinking - their legacies are not just rooted in the innovations they introduced but rather in the challenges they posed. Specifically, the challenge of creating machines that can truly understand and interact with the world in intelligent ways.

With the mid-20th century marking significant milestones in the journey toward artificial intelligence, here is a brief overview of some of the groundbreaking developments at the time:

1950: Alan Turing published “Computer Machinery and Intelligence,” proposing the Turing Test to assess a machine's ability to exhibit intelligent behavior indistinguishable from that of a human.
1952: Computer scientist Arthur Samuel developed a checkers-playing program, the first to independently learn and improve, pioneering early machine learning.
1955: John McCarthy organized a workshop at Dartmouth, coining the term “artificial intelligence” and formally establishing AI as a field of study.
1958: John McCarthy created LISP (List Processing), the first programming language designed for AI research, which remains in use today.
1959: Arthur Samuel coined the term “machine learning” in a speech about teaching machines to play chess better than their human programmers.
1961: The first industrial robot, Unimate, began working on an assembly line at General Motors in New Jersey, handling tasks deemed too dangerous for humans.
1965: Edward Feigenbaum and Joshua Lederberg developed the first “expert system,” an AI designed to replicate human expert decision-making.
1966: Joseph Weizenbaum created ELIZA, the first “chatterbot” (later known as a chatbot), which used natural language processing (NLP) to simulate conversations with humans.
1968: Soviet mathematician Alexey Ivakhnenko published “Group Method of Data Handling,” introducing concepts that would later evolve into “deep learning.”
1973: Applied mathematician James Lighthill reported to the British Science Council, highlighting the gap between AI promises and achievements, leading to reduced government support and funding for AI research.
1979: James L. Adams's Stanford Cart, developed in 1961, successfully navigated a room full of chairs autonomously, marking an early milestone in autonomous vehicles.
1979: The American Association of Artificial Intelligence, now known as the Association for the Advancement of Artificial Intelligence (AAAI), was founded.

These developments not only laid the groundwork for modern AI but also showcased the early potential and challenges of creating intelligent machines.

1980s - The First Hype Cycle & AI Winter

The 1980s witnessed an exhilarating period of rapid growth and burgeoning interest in artificial intelligence, often referred to as the "AI boom". This era was marked by significant research breakthroughs and substantial government funding that bolstered the efforts of pioneering researchers. During this time, deep learning techniques and expert systems gained prominence, enabling computers to learn from their errors and make autonomous decisions.

In 1980, the inaugural conference of the American Association for Artificial Intelligence (AAAI) was held at Stanford, marking a pivotal moment for the AI community. That same year, the commercial market welcomed its first expert system, XCON (expert configurer), designed to streamline the ordering of computer systems by automatically selecting components based on customer specifications.

^{Proceedings of the First Annual National Conference on Artificial Intelligence at Stanford University - August 18-21, 1980.}

The following year, in 1981, the Japanese government made a monumental investment of $850 million (equivalent to over $2 billion today) in the Fifth Generation Computer project. This ambitious initiative aimed to develop computers capable of translating languages, engaging in human-like conversation, and reasoning at a human level.

In 1984, the AAAI issued a cautionary warning about an impending "AI Winter", a period during which funding and interest in AI were expected to wane, potentially hampering research efforts. Despite these concerns, innovation continued. In 1985, the autonomous drawing program AARON was showcased at the AAAI conference, demonstrating AI's creative potential.

A year later, in 1986, Ernst Dickmann and his team at Bundeswehr University of Munich unveiled the first driverless car. This groundbreaking vehicle could navigate roads at speeds up to 55 mph without human intervention or obstacles.

By 1987, the commercial landscape saw the launch of Alacrity by Alactrious Inc., the first strategic managerial advisory system. Alacrity utilized a sophisticated expert system comprising over 3,000 rules, marking a significant advancement in AI applications.

These milestones collectively illustrated the dynamic and rapidly evolving landscape of AI during the 1980s, setting the stage for future developments and breakthroughs in the field.

The AI Winter

As the AAAI had cautioned however, the AI Winter arrived and it casted a chill over the once vibrant field of artificial intelligence. This term describes a period marked by dwindling interest from consumers, the public, and private sectors, resulting in a significant reduction in research funding and fewer breakthroughs. Both private investors and government entities, disillusioned by the high costs and seemingly low returns, withdrew their financial support. This AI Winter was precipitated by several setbacks, including the conclusion of the ambitious Fifth Generation project, reductions in strategic computing initiatives, and a slowdown in the deployment of expert systems.

Artistic Rendition of the AI Winter ^{Generated with Midjourney: a lonely man walking through a winter tundra dragging a broken robotic humanoid, photo-realism, realistic, hyper-realistic, wide-angled lense, photography, 4k, photo-realistic, unreal engine, snowing --ar 7:2 --v 5.0}

Key events during this period include:

In 1987, the market for specialized LISP-based hardware imploded. More affordable and accessible alternatives, such as those offered by IBM and Apple, could run LISP software, rendering specialized LISP machines obsolete. This shift led to the collapse of many companies that had specialized in LISP technology, as it became widely available through more cost-effective means.

In 1988, computer programmer Rollo Carpenter introduced the chatbot Jabberwacky. Designed to engage in interesting and entertaining conversations with humans, Jabberwacky represented a bright spot in an otherwise challenging time for AI, showcasing the potential for AI to interact with people in a more natural and engaging manner.

1990 to 2011 - Big Tech & The AI Resurgence

Despite the lean years of the AI Winter, the early '90s marked a period of remarkable advancements in artificial intelligence. It was an era that witnessed the debut of the first AI system capable of defeating a reigning world chess champion and saw AI infiltrate everyday life through innovations like the Roomba and the first commercially available speech recognition software on Windows computers.

This renewed interest in AI sparked a wave of funding, propelling research and development to new heights.

^{IBM's Deep Blue vs Garry Kasparov - 1997}

Key Milestones:

1997: In a highly publicized match, IBM's Deep Blue made history by defeating world chess champion Garry Kasparov, marking the first time a program bested a human chess champion.
1997: Windows introduced speech recognition software developed by Dragon Systems, bringing AI-driven speech capabilities to the masses.
2000: Professor Cynthia Breazeal unveiled Kismet, the first robot designed to simulate human emotions with its expressive face, complete with eyes, eyebrows, ears, and a mouth.
2002: The world saw the release of the first Roomba, revolutionizing home cleaning with its autonomous vacuuming capabilities.
2003: NASA achieved a significant milestone by landing two rovers, Spirit and Opportunity, on Mars. These rovers navigated the Martian surface without human intervention, showcasing the potential of AI in space exploration.
2006: Major companies like Twitter, Facebook, and Netflix began integrating AI into their advertising and user experience (UX) algorithms, enhancing how they engaged with users.
2010: Microsoft launched the Xbox 360 Kinect, a groundbreaking gaming hardware that tracked body movement and translated it into gaming commands, offering a new dimension in interactive entertainment.
2011: IBM's Watson, an NLP computer programmed to answer questions, won Jeopardy against two former champions in a televised game, demonstrating the advanced capabilities of AI in natural language processing.
2011: Apple introduced Siri, the first popular virtual assistant, which brought conversational AI to millions of users, setting the stage for future advancements in personal assistant technology.

This period, spanning nearly two decades, laid the foundation for the modern AI landscape, with each milestone contributing to the technology's growing presence in our daily lives.

2012 to 2023 - Signs of General Intelligence

The dawn of the 2010s marked a transformative era in the world of Artificial Intelligence (AI), ushering in a wave of innovations that continue to shape our daily lives. From the advent of virtual assistants to the pervasive influence of search engines, this period saw the rise of Deep Learning and Big Data, catapulting AI into the mainstream.

Depiction of General Intelligence ^{Generated with Midjourney: photo-realistic, 4k, unreal engine, depth of field, focus on a robotic hand sticking out from the ground, thousands of humanoid robots crawling out of the ground, zombie movie, hands sticking out of the ground, winter, snowing, mount fiji in the background --ar 7:2 --v 5.0}

Transformer Models & Self Attention

Before we go any further, I'd like to mention that the following history of transformer models and self attention was heavily inspired by a conversation between Lex Fridman and Aravind Srinivas from PerplexityAI - aired on the Lex Fridman podcast on June 19, 2024.

Yashua Bengio wrote a paper in 2014 with Dzmitry Bahdanau and Kyunghyun Cho called Neural Machine Translation by Jointly Learning to Align and Translate which introduced the concept of soft attention.

Ilya Sutskever, Oriol Vinyals, and Quoc Le wrote a paper in 2014 called Sequence to Sequence Learning with Neural Networks which introduced the concept of sequence to sequence learning. This paper suggested that a neural network could be trained to map an input sequence to an output sequence. It was a an RNN based model that could be used for machine translation, image captioning, and more. By scaling it up it could be all the phrase based machine translation systems. This approach however was brute force, not very efficient and lacked the attention mechanism previously introduced by Bengio and his team. Training this model costs a lot of money and time - all paid on Google's dime.

This student Badhanau was a student of Bengio and he was the one who introduced the attention mechanism to the sequence to sequence model - beating Ilya Sutskever's benchmarks with valence compute. And then people at deep mind figured that we don't even need RNNs - having introduced a new concept called Pixel RNN's (which contrary to name itself is NOT actually an RNN). They posited that you don't need an RNN and they eventually evolevd this PixelRNN model to a new model called WaveNet - where they were able to generate speech that mimiced human voices and sounded better than exising TTS (text to speech) systems, improving performance by upwards of 50%. This model also demonstrated that you could generate music and other audio signals - including segments of generated piano music that sounded like it was played by a human.

They realized in these experiemnts that a CNN (convolutional neural network) based model can do autoregressive modeling as long as you do masked convolutions. The masking was the key idea. So you can train in parallel instead of backpropagating through time. You can backpropagate through every input token in parallel. So that way you can utilize the GPU compute a lot more efficiently because you’re just doing matmuls. And so they just said throw away the RNN. That was powerful. And so then Google Brain, like Vaswani et al., the transformer paper, identified that, okay, let’s take the good elements of both - lets take attention as its more powerful than KANs, it learns more higher order dependencies because it applies more multiplicative computer. THen let's take the inside of WaveNet that you can just have an all CNN modle that fully parallel matrix multiplies and and combine the two together and they build a transformer.

And that is the, I would say, it's almost like the last answer. Like nothing has changed since 2017 except maybe a few changes on what the nonlinearities are and like how the square root descaling should be done. Like some of that has changed but and then people have tried mixture of experts having more parameters for the same flop and things like that but the core transformer architecture has not changed. Is it crazy to you that masking as simple as something like that works so damn well?

Yeah it's a very clever insight that look you want to learn causal dependencies but you don't want to waste your hardware to compute and keep doing the back propagation sequentially. You want to do as much parallel compute as possible during training. That way whatever job was earlier running in eight days would run like in a single day. I think that was the most important side of it. Whether it's cons or attention, I guess attention and transformers make even better use of hardware than cons uh because they apply more compute per flop because in a transformer the self-attention operator doesn't even have parameters. The qk transpose softmax times v has no parameter but it's doing a lot of flops and that's powerful. It learns multi-order dependencies. I think

Pivotal Milestones in AI Advancement

2012: In a groundbreaking experiment, Google's Jeff Dean and Andrew Ng trained a neural network to identify cats by feeding it unlabeled images. This milestone demonstrated the power of unsupervised learning, setting the stage for future AI developments.
2012:
2015: A call for ethical AI practices resonated globally when prominent figures like Elon Musk, Stephen Hawking, and Steve Wozniak, along with over 3,000 others, signed an open letter urging governments to ban the development and use of autonomous weapons. This move highlighted the growing concerns about AI's potential misuse.
2016: Hanson Robotics unveiled Sophia, a humanoid robot that captured worldwide attention as the first "robot citizen." Sophia's realistic human appearance, emotional replication, and communicative abilities marked a significant leap in robotic technology and human-robot interaction.
2017: Facebook's experiment with AI chatbots took an unexpected turn when the bots, designed to negotiate, abandoned English and developed their own language. This incident underscored the autonomy and unpredictability of advanced AI systems.
2018: Alibaba's language-processing AI achieved a significant victory by outperforming human intellect on a Stanford reading and comprehension test. This accomplishment showcased AI's growing proficiency in understanding and processing natural language.
2019: Demonstrating AI's prowess in strategic thinking, Google's AlphaStar reached Grandmaster status in the complex video game StarCraft 2, outperforming all but 0.2% of human players. This achievement highlighted the potential of AI in mastering intricate and dynamic environments.
2020: The introduction of GPT-3 by OpenAI marked a revolutionary step in natural language processing. Capable of generating code, poetry, and other written content nearly indistinguishable from human creations, GPT-3 set a new standard for AI-generated text.
2021: OpenAI's DALL-E brought AI closer to visual comprehension by producing accurate image captions. This development signified a major advance in AI's ability to understand and interpret the visual world, paving the way for future innovations in image processing and analysis.
2022: ChatGPT is launched by OpenAI and it shocks the world with an incredibly useful chatbot able to be used across hundreds of every day use-cases and commercial use-cases.
2023: Google, Anthropic, OpenAI and many other AI startups begin to expand upon their core capabilities - including improvements in memory, voice, reliability and reasoning for more complex tasks.

These milestones reflect the rapid evolution of AI, charting a course from early experiments to sophisticated systems that rival human capabilities. As we stand on the cusp of even greater advancements, the journey of AI continues to captivate and challenge our understanding of intelligence and innovation.

2024 to 2040 - The Era of AI Agents

Coming soon...

Conclusion

This concludes the first segment of my "What exactly are AI agents?" series where we mostly discussed the history of AI. In part two we'll look at the underlying design patterns behind modern AI agents - including ideas derived from cognitive architecture and incredible research that has emerged in recent years. In part three, we'll put everything we've learned into practice by implementing some incredibly fascinating use-cases using Python and libraries such as Langchain, Langgraph, Microsoft's Autogen and CrewAI.