Can artificial intelligence do science?
Science is still largely performed by humans and has avoided computerisation. We need to get science moving at computer speed with A.I. and lab automation to speed up discovery, cures, and inventions. We still must submit a peer reviewed paper and so the A.I. should perform the scientific method without any human intervention. For example, switch on the A.I. and on it goes doing science, come back later and get a half acceptable scientific paper fit for any journal.
Systematic problem-solving identifies steps that are followed to arrive at a conclusion and result. One of the major systematic problem-solving methodologies is the scientific method. Systematic problem-solving methods exist in many fields from science, engineering and business. Here we focus on them all with 3 ways to implement these system into A.I.
- Prompt Engineering
- Fine Tuning existing model
- Specialized new model
These models can use different A.I strategies.
Add which systematic problem-solving technique you are working on. Add either your prompt engineering, notes on fine-tuning an existing model, or notes on building a specialized model. Along with any issues that stop your progress.
Look at the scientific method. Each step in the systematic problem solving technique is termed a model. The scientific method comprises of 6 steps, the last step being, communicate results. So the A.I. really needs to perform 5 of the 6 competencies. The base folder is named after the systematic technique in question, here "scientific_method". Each subdirectory is named after the step in the method. So the 5 models for the scientific method are...
The scientific method starts when you ask a question about something that you observe: How, What, When, Who, Which, Why, or Where? Rather than starting from scratch in putting together a plan for answering the question, a savvy A.I. scientist does the research to produce important questions that do not repeat mistakes from the past.
A hypothesis is an educated guess and a prediction about how things work. It is an attempt to answer your question with an explanation that can be tested. A good hypothesis allows you to then make a prediction: "If _____ I do this _____, then _____ this _____ will happen." State both the hypothesis and the resulting prediction to be tested. Predictions must be easy to measure.
Design an experiment that can be performed that tests whether the prediction is accurate and thus the hypothesis is supported or not.
The experiment must be performed by the A.I. and be of high standards and fair. Repeat the experiments several times to make sure that the first results were not just an accident.
The A.I. generates a report, journal paper. Analyse results to see if they support the hypothesis or not.
Scientists often find that their predictions were not accurate, and their hypothesis was not supported, and in such cases they will communicate the results of their experiment and then go back and construct a new hypothesis and prediction based on the information they learned during their experiment. This starts much of the process of the scientific method all over again. Even if they find that their hypothesis was supported, they may want to test it again in a new way.
Iterate and return to Step 1 or conclude...
Auto-Science Journal for A.I. generated papers - https://auto.imt.cx
Communicate the results to others in a final report. Professional scientists do almost exactly the same thing by publishing their final report in a scientific journal or by presenting their results on a poster or during a talk at a scientific meeting. Other A.I. to counter argue and perform their own experiments. We cannot risk this becoming computer generated spam so there is a strict editorial process.
Evaluation: the hypothesis is evaluated against the findings in known experimental outcomes to determine hypothesis model competency.
These competencies can be met in 3 different ways currently, by prompt engineering, fine-tuning and building new models from scratch.