In artificial intelligence, there’s a major difference between the concepts of following instructions and executing tasks by a machine. High level tasks are only known to the programmer who writes instructions for the machine that, when followed, would ensure that these tasks are executed accurately.
To illustrate this, here’s an example of a simple algorithm for a chess game written by a programmer and expressed in a high level language, human language, defining practical steps which should be followed by a machine to achieve a higher level task, which is to win the game in this case.
The algorithm steps for the machine are as follows:
In order to achieve the higher level task, which is to win the game, the algorithm steps must be accurate and clear. In this example, if the target mentioned in the last step of the algorithm doesn’t include the phrase “before he takes yours”, the machine could continue playing while already having lost the game. This dependency on human instructions in this example shows that the machine cannot take its own decisions when it comes to achieving high level tasks defined by human.
Mohamed el Nawawy ,
The root cause of the problem is not that "the algorithm steps must be accurate and clear" it is that the AI algorithms are not designed in a way where they can really understand in the way that you are referring to.
To first give you why it does not depend on clear instructions, an AI right now can be designed on simple image processing to follow the trajectories of each piece from a video camera placed above the board. Given several hundred to several thousand videos of successful and failed games, the system can learn how to play based only on trajectories (I have oversimplified the complexity of the setup, but this setup illustrates how to setup an AI training where rules of the are not explicit, they are learned).
The second objection to this line of thought is that current deep learning does indeed learn context. To this, the counterargument is to define what type of context is being referred to. Deep learning does learn context in terms of previous information on a sequence of data given to it. But, is this the right context? This is where the response 'it depends' is usually brought up. The response in defense of deep learning is that the sequence can be as specific as required. But I have not seen any real mathematical proof that this can be done for all cases (universal proofs that I have seen are existential and not really universal in nature).
To finally answer your question whether they understand context, the answer would be that at this stage is a no because of inherent limitations in design and the definition of context. Does that mean that the goal is not achievable, the answer is also from my perspective a no.
Your example underscores the fundamental difference between a machine's ability to follow instructions and its inability to independently define or achieve high-level tasks. While machines can execute tasks with remarkable precision, their success depends entirely on the clarity and completeness of the instructions provided by humans. This highlights the critical role of programmers in translating high-level goals into actionable steps for machines to follow.
The proper comprehension of context remains a challenge for machines. The higher levels of intelligence will of course depend on the skills of programmers - In my career, in the energy sector, I co-developed a patent referred to as "data in context", that is intended to enable search algorithms and/or users of scientific software (geosciences in this case) to "learn" from the data already available, plus the actions of the user and the specificity of the task. If interested, please see USPTO publication: Chapter METHOD AND SYSTEM FOR DATA CONTEXT SERVICE
- Badges
- Science method