ML for Science and Engineering - Lecture 29
Tools, responsibility, and how to act when you cannot know what is right
Joseph Bakarji
Engineers build tools. That is the job.
You learn to build small tools, then bigger tools.
When you build tools for a company, in a structural sense, you are a tool. The company uses you to build its tools.
The concept of using something and being used generalizes across humans and machines.
Tools are means to ends. Trace any tool far enough and you arrive at a small set of human ends.
You are building a screw in a very large machine.
AI systems are not only part of larger opaque systems, they are themselves unpredictable.
Traditional tool: You know the input-output relationship. The wrench turns the bolt.
AI tool: You release a neural network. People use it in ways you did not anticipate. The consequences branch unpredictably.
Unpredictability is not a defect. It is what makes intelligence interesting.
A perfectly predictable tool is not creative.
If you hire someone to be creative, you must tolerate unpredictability. You do not know what ideas they will produce.
The same holds for AI. If you want it to be useful in genuinely new situations, it will surprise you.
You already know this from this course.
Chaotic systems are deterministic. Given the same initial conditions, they produce the same trajectory. But tiny perturbations lead to vastly different outcomes.
No matter how many rules you impose on a high-dimensional system, the dimensionality defeats regulation.
One tool, many outcomes. Click nodes to explore branches.
Model consequences as a probability distribution over a value spectrum.
Every powerful technology comes with guidance. AI should be no different.
A practical approach to managing unpredictable tools.
The type of question you ask determines which level you are working at.
A concrete example of applied ethics with no clean resolution.
Each theory gives you a lens. Each lens distorts something.
| Theory | Core Principle | Limitation |
|---|---|---|
| Utilitarianism | Maximize collective happiness | You cannot sum happiness; ignores distribution |
| Deontology | Follow universal moral rules | Rules conflict; rigid in edge cases |
| Virtue Ethics | Cultivate virtuous character traits | Who defines virtue? Culture-dependent |
| Divine Command | Good = what the creator commands | Requires shared religious framework |
You have intuition that 1 + 1 = 2.
You have intuition that some things are good and some are bad.
Neither has objective grounding beyond the feeling.
Some positions:
Here is the deeper point. It is not just ethics that is uncertain.
Your entire understanding of the world is subjective. Every model, every differential equation, every scientific framework happens inside your head.
Perhaps it is collective subjectivity. But subjective nonetheless.
If scientific understanding is uncertain, ethics being uncertain is not a special crisis. It is the general condition.
Not a normative theory. Phenomenological principles for how to act when you cannot know.
This is not just a saying. It is a practice.
Every scientific model relies on assumptions: smoothness, observability, stationarity, linearity.
Every assumption eventually fails. This is what you learned in this course.
The awareness that your knowledge rests on breakable assumptions is uncomfortable. But it is the precondition for growth.
The counterweight to epistemic paralysis.
You will not discover whether something works by thinking about it. You have to try it, fail, and revise.
This is the train/test loop applied to life:
Humility and action are not in conflict. They are phases of the same loop.
Radical epistemic humility plus aggressive action.
These are not in conflict. They are the same cycle, viewed from two angles.
Knowing you don't know → fuels curiosity
Curiosity → drives action
Action → generates feedback
Feedback → refines knowledge
Knowledge → reveals deeper ignorance
The traditional model: a toolbox.
For each problem type, there is a skill.
Mechanical engineering problems → statics, dynamics, thermo
Data problems → regression, classification, clustering
Optimization problems → gradient descent, linear programming
In a world with language models, anyone can look up any skill.
You do not need an engineering degree to have access to a toolbox. You need a search engine.
"This is my problem. What tool do I need? What is the solution?" The language model answers that.
If skill lookup is commoditized, what is left?
Real problems are not given to you pre-defined. You have to define them.
There are infinitely many ways to define a problem.
You tend to define problems using the tools you already have. "I know ML, so this is an ML problem."
But maybe it is a social problem. Or a design problem. Or a policy problem.
Taste is not taught. It emerges from repeated cycles of failure and revision.
If you have never failed, you cannot tolerate the risk required for genuinely hard problems.
Startups fail ~90% of the time. Successful founders have usually failed before. They know what failure feels like, so they can tolerate the risk.
If your education has been frictionless, you have never developed tolerance for failure. You will avoid hard problems.
"Reality requires you to have taste, requires you to fail, and through the effort of building things and failing, you become good at your art."