Probabilistic to Deterministic and back to Probabilistic (or Analog to Digital and back to Analog)
The events in this world can be probabilistic or deterministic. In many cases, the event is probabilistic, i.e., there is probability associated with if the event will happen or not. Will their be a rainfall? You can’t be sure, but you can assign a probability to it. On the other hand, there will be deterministic events, where we can be certain of the outcome — Sunrise for example. It will definitely happen at x am. I wonder if our life has gone from being primarily probabilistic (till the half century ago or so) to deterministic (during the last half century) thanks to the technology and computers. Now with the advent of AI and LLMs, it is becoming probabilistic again. Put in other words, technology has made our lives from analog to digital and now making it again analog.
During the human evolution, humans did not have control over the nature and hence, their actions were probabilistic. Take ‘starting a fire’ for example. You can attempt to start a fire, but there is always uncertainty associated with it. Many things we take for granted now such as getting food, traveling to other places, curing a disease often had probabilistic outcomes. Human endeavor was to improve the probability of success, but it would never reach certainty. With Industrial Revolution, the technology tamed nature, but the technology remained analog. You could start a train by burning coal, but the start was not with a button which always started the train. The motormen had to use their skills and luck to ensure that there was a high probability that the train started on time. The instruments continued to be based on the analog technology, which continued till the arrival of the computers. Even the initial computers were based on the mechanical valves and transistors which made them less reliable. There could be unexplained reasons because of which the equipment may or may not work.
Then came the digital revolution. With semiconductors, the transistors became reliable and the information could be coded in 0s and 1s. This provided tolerance to the systems and the information could be sent reliably. Those who have used analog telephones before, know the uncertainty and issues sometimes associated with them. There could be disturbance in line, dropped calls. Digital telephones made them more reliable. And this digitization touched most aspects of our life. Most electronic equipments and even cars became digital. In the manufacturing there was a focus on improving quality, so that there was a certainty associated with the product through initiatives such as six sigma. This ensured that even the things which continued to use analog technology behaved as digital as the probability of other outcome was driven drastically low through continuous improvements. Inputs were also digital which meant that you could input only limited set of options. In the case of analog inputs, as long as the input is within a certain tolerance limit (which kept on increasing), the output was certain.
But now with LLMs and other AI tools we are back to the uncertain probabilistic age. The output of LLMs is not deterministic. Efforts are on to make it deterministic and increase the fault tolerance, but that is still not the case. For the same input, the output may not be same for every occassion. There is some variation, which is difficult to predict beforehand. Also, for a slightly different input, the output can vary drastically. This has given rise to ‘prompt engineering’, a fancy term for a brute force approach of using different inputs to get the desired output. As we develop more and more tools using these technologies, developers need to ensure that they are deterministic and users can have confidence in their outputs.
This raises two questions.
- Why are systems probabilistic or deterministic?
- What impact do such systems have on the users? Have they changed the behavior of human beings over time?
Some systems are deterministic by design. The computer programming so far was deterministic. Programmers explicitely told the computer through codes about the action to be taken in case of a certain scenario. There used to be a proper flow diagram. The system returned error in case the computers received something which they were not told to do. Hence, in the last 50 years, all the systems people were used to were deterministic. But things changed with LLMs. In normal coding, the programmers provide input and logic and the computer provides the output. In the new paradigm of LLM, the programmers provide inputs and outputs and the computer ‘derives’ the logic in the training stage. In the inference stage, the computer uses this logic to provide response. So the programmer does not know what is the ‘logic’ which the computer has derived or how this ‘logic’ is applied to the input. Hence, we can only guess what could be the likely output. In simple cases, we can predict with high probability, but can never be sure as we do not know the logic. Even small changes in the input can create a large impact on the complicated logic derived and can dramatically change the output. Again, we do not know what that impact could be and hence, we can only say with a probability what the answer could be.
Thus, a system can become probabilistic if the inner workings are too complicated and/or if we do not understand what is going inside. This is what precisely happened in the earlier analog world. The analog signals flowing though the telephone cable could be impacted by multiple external factors, which one could not predict. In the digital world, though the signal passes the same cable, because the information is not encoded in the things which are impacted, the tolerance is higher. In the LLM world, the system could be impacted by some random text that the model learnt in the training phase and could result in a cascade and wrong results. In short, we are back in the uncertain probabilistic world.
But how does this impact the generation which has been living in the deterministic world? People have been accustomed to getting fixed results for their inputs (I remember old ad by TCS — Experience Certainty!). So probabilistic responses can throw them off guard. People used to light to turn on whenever they press the switch, can’t think of a situation where there is a possibility that the light might not turn on. Rather I have seen people using probabilistic tools as a deterministic. Google maps is a perfect example. The ETA on the map is a high probability time and there are possible scenarios where that time may not materialize. No map algorithm can predict uncertainties about traffic. All the route options within 1–2 minutes of ETA can potentially result in a similar ETA depending upon the live traffic conditions. But people do anchor on the ETA and determine their route solely on the software ignoring the real life traffic signals. With proliferation of LLM-like models, people may still anchor on the answer they received as the only truth further exacerbating the narrative battles in the post truth world.
One of the reason for this thought process is the belief that the computers or the technology always produces fixed results. Humans can sometimes fail, provide incorrect answers, but not technology. With LLM like technology are systems becoming human again? Are they prone to same errors and vagaries as humans? Perhaps people need to change their mental models about systems/ tech/ computers and think of them as similar to humans. We are entering a new world of AI and our perspective about machines needs to change to thrive in this new world!
PS: I had to check the spelling of analog before I started the blog and apparently analog is an American spelling while analogue is British. I am sticking with the American spelling for this post.
