From BFS to AGI: AI history has more than one lane
Search, learning, transformers, and the unfinished question of general intelligence—a short history without the inevitable-destiny soundtrack.
By augova-tech-team
The title needs a small plot twist
“From BFS to AGI” sounds like a tidy journey: one humble search algorithm grows up, gets a neural network, and eventually starts negotiating its own salary.
Actual AI history is messier. Search, logic, statistics, and learning have developed in overlapping lines. They still coexist. Breadth-first search is a specific algorithm; artificial general intelligence is a contested description of broad capability. These are very different kinds of thing.
So treat this as a tour through ideas, not a straight evolutionary ladder with a guaranteed destination. The interesting part is how different approaches help solve different problems.
1955–1956: a name and a large ambition
The Dartmouth proposal, dated August 31, 1955, proposed a summer research project for 1956 and used the term artificial intelligence. Its ambitions included language, abstraction, problem solving, and self-improvement.
It also discussed neural networks. Learning was already in the conversation; it did not arrive decades later as a surprise replacement for all earlier ideas.
The proposal makes wonderful historical reading because the ambition is enormous and the proposed gathering lasts two months. Anyone who has estimated a software project can appreciate the optimism. It helped establish a research agenda, while leaving plenty of difficult questions for subsequent generations.
Search: explore possible next steps
Breadth-first search explores a graph level by level. Starting at one state, it visits immediate neighbours before states two steps away. With suitable visited-state handling, it finds a shortest path measured in edges in an unweighted graph.
In 1959, Edward F. Moore described this kind of level-by-level exploration for finding the shortest path through a maze. That can solve a maze without learning anything from a dataset. The algorithm follows an explicit procedure over possible states. Its drawback is that the frontier can grow very large.
Other search methods use heuristics to prioritize promising possibilities. Search remains useful in planning and problem solving. Calling it old-fashioned is a little like retiring maps because somebody invented a better camera.
Learning: improve from examples
Learning takes a different route: fit a model using data or experience. A recognizer might learn from labelled images instead of receiving a handcrafted rule for every visual pattern.
The 2012 ImageNet paper by Krizhevsky, Sutskever, and Hinton demonstrated strong image classification results with a deep convolutional network trained using GPUs. It was an influential demonstration of what that combination could achieve, rather than the invention of neural networks themselves.
Search and learning can also work together. A learned estimate can help choose which possibilities a search explores. The toolbox gains another instrument; it does not need to throw the previous instruments into the sea.
1997 and 2016: games become public scoreboards
Games are handy for researchers: the rules are clear, and the scoreboard does not accept excuses.
In May 1997, IBM's Deep Blue beat world chess champion Garry Kasparov 3.5 to 2.5 under standard tournament time controls. Its strength came largely from searching enormous numbers of chess positions, guided by carefully engineered evaluation. No robot uprising followed; chess clubs stayed open.
In 2016, a Nature paper described AlphaGo, which combined deep neural networks with tree search and beat the European Go champion five games to zero. That is a neat plot point for our story: search and learning were not rivals fighting for the throne. They were teammates, and one of them had very good pattern recognition.
2017: attention gets a new architecture
Attention Is All You Need introduced the Transformer architecture in 2017, initially reporting machine translation experiments. Attention lets a model combine information across positions in an input sequence.
The paper is a concrete architectural milestone. It is not a declaration that language, reasoning, or intelligence had been solved. A useful historical habit is to separate what a paper demonstrated from everything people later hoped its descendants might do.
For builders, the lesson is practical: architecture, data, computation, and evaluation all matter. One elegant idea can enable substantial progress while still leaving demanding questions about reliability and real-world use.
AGI: define the finish line before announcing it
Artificial general intelligence usually points toward broad capability across tasks, but definitions and evaluation criteria differ. The research paper Levels of AGI proposes a framework involving both breadth and performance, illustrating why a single flashy demonstration is an incomplete measurement.
AGI remains an uncertain aspiration, with no agreed arrival date. Fluent conversation does not by itself establish reliable general competence, and predictions about future systems should remain predictions.
For a business deciding what to build today, ask narrower questions: can this system complete our task, under our constraints, on unfamiliar examples? What happens when it fails? Those questions are less cinematic than declaring a new era, but they produce much better engineering decisions.
Sources and further reading
McCarthy and colleagues: Dartmouth proposal, 1955
https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
Russell and Norvig: Search algorithms (PDF, chapter 3)
https://aima.cs.berkeley.edu/algorithms.pdf
Krizhevsky, Sutskever, Hinton: ImageNet classification, 2012
https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
Vaswani and colleagues: Attention Is All You Need
https://arxiv.org/abs/1706.03762
Morris and colleagues: Levels of AGI
https://arxiv.org/abs/2311.02462
IBM: Deep Blue
https://www.ibm.com/history/deep-blue
Silver et al. (2016), Mastering the game of Go with deep neural networks and tree search
https://www.nature.com/articles/nature16961