Artificial Intelligence has a knowledge problem. The latest wave of big AI tools, Large Language Models (LLMs) like ChatGPT/GPT4 and Llama, don’t encode actual knowledge about the world. They are only models that predict what words should come next, based on many billions of words of training data. LLMs usually sound like they are providing “correct” answers to our prompts, but this is merely a side-effect of the fact that LLMs are trained on data that contains enough correct information that it usually ends up sounding correct.
We are already seeing people getting frustrated with LLMs for “hallucinating” or “making things up” when the LLM generates text that is wrong or nonsensical. In Europe, government regulators are hard at work trying to mitigate potentially harmful uses and outputs of LLMs. OpenAI, the company behind ChatGPT, is working hard to make sure their product is safe, and at the same time, they are contributing to the problem by selling ChatGPT as a problem-solving tool.
A Very Brief History of AI
In the 1940s and 1950s, mathematicians and computer pioneers (most of whom were mathematicians) got excited about the potential of machines to think like humans. In the first wave of AI, this looked primarily like programming computers to play chess and checkers. These programs are search algorithms – they build a tree of possible moves and then search it for winning moves based on rules that assign values to different board positions. It soon became apparent that because of the number of possible positions in chess (about 10^44), we wouldn’t have enough computing power to search deeply enough to beat humans at chess for a very long time, and researchers moved on to other problems.
The second wave of AI in the 1960s-1980s consisted largely of heuristics engines and expert systems. Heuristics engines use a set of shortcuts or rules of thumb to quickly find a “good enough” solution to a complex problem that might not be solvable with traditional methods. Most anti-virus software uses a heuristics engine to identify suspicious behaviors. Expert systems encode knowledge about a specific domain of expertise in a rigorous way and use formal logic to solve problems within that domain. Examples include programs that can generate and verify mathematical proofs and programs for medical diagnosis. These systems were very highly developed by the early 1980s but fell out of favor for two main reasons: (1) they were extremely expensive to develop and maintain, and (2) they behaved unpredictably when faced with inputs that didn’t map directly to the underlying knowledge.
In the 1990s, computing power caught up to the chess problem enough to allow Deep Blue to beat Gary Kasparov, the top human player of the era. It was a landmark moment in AI, but looking back, probably said more about the weakness of human chess players than the strength of computers. We still can’t compute the full game tree of chess, or even exactly how many legal, reachable board positions there are.
By the early 2000s, large pattern recognition problems started to gain favor, as the combination of increasing computing power and improved neural network methods captured the imagination of researchers and programmers. Significant advances were made in techniques that would eventually prove fruitful in areas like voice recognition, handwriting recognition, facial and object recognition, and language translation. In 2012, a major neural network advance made image recognition practical, and neural networks have dominated the development of AI since then.
In the 2010s- present, the advent of compute-as-a-service enabled by cloud computing and the use of parallel processing GPUs to implement neural networks have lowered the barriers to AI development. LLMs were initially made possible by this new compute environment, and are gradually being moved onto smaller platforms. What we haven’t seen yet is a return to the knowledge-based approaches of the 70s and 80s.
A Call to Action
As we are poised to become dependent on LLMs to enhance our productivity, it will be increasingly urgent that we build a knowledge base under these tools. If we fail to do so, we risk granting authority to LLMs that are fundamentally untethered from actual knowledge of the world.
In future posts, we will examine the human knowledge problem, ponder what a return to knowledge-based AI might look like, and consider how incorporating a bottom-up knowledge representation might help us solve some of the biggest problems in scientific research today.