MYP Design Artificial Intelligence and Emerging Technologies Notes

Machine Learning Is Pattern Finding, Not Understanding

1. Ordinary programming hands a computer a set of rules, while machine learning hands it thousands of examples and lets it work the rules out.
2. Show a model 50,000 labelled photographs of chairs and it learns which arrangements of pixels tend to come with the label chair.
3. Training is the slow, costly part done once on huge computers, and inference is the fast part that runs when you actually use the tool.
4. A large language model is trained on enormous amounts of text and predicts which words are likely to come next, which is why it writes smoothly and can still be badly wrong.
5. None of these systems holds a model of what is true, so every fact one gives you needs checking against a real source.
   1. The same gap explains why an image generator will cheerfully draw a chair with five legs and no visible joints.

Where AI Tools Actually Help in Design Work

1. Image generation earns its place early, for mood boards and quick visualisations of a concept you have already chosen.
2. It is useless at anything dimensional, so a generated picture is not a drawing and cannot stand in for a CAD model or an orthographic view.
3. Code assistance speeds up the repetitive parts: reading an Arduino sensor, formatting a chart, finding the missing bracket you have stared past six times.
   1. Code you cannot explain is code you cannot fix at nine o'clock the night before your deadline.
4. Summarising helps you get oriented in an unfamiliar field, after which you go and read the actual sources yourself.
5. Speech to text for interview transcripts and background removal on product photographs are quiet, genuinely useful jobs that nobody argues about.

Bias Comes From the Training Data, Not the Code

1. A model absorbs whatever pattern sits in its examples, including the patterns nobody intended to teach it.
2. Train a hand tool recommendation on photographs of one group of users and the product it suggests fits that group and fails everyone else.
3. Image generators asked for an engineer have returned the same narrow set of faces, because those faces dominated the images they were trained on.
4. Sampling bias, data that does not represent the people who will actually use the product, is the version most likely to catch a student designer.
5. Voice recognition trained mainly on one accent misses words spoken in another, which turns a convenience feature into a barrier.
6. Reducing bias means changing the data and testing with real users from the groups you left out, not rewording the prompt.

Declaring AI Use Is Part of the Work, Not an Admission

1. Say where you used an AI tool, which tool it was and what you asked it for, at the point where it appears in your folder.
2. A generated image on a mood board gets the same caption treatment as a photograph you found online.
3. Code from an assistant gets a comment in the file naming where it came from and what you changed.
4. The work judged against the four criteria has to be your thinking, so a generated proposal you did not develop is not evidence of Developing ideas.
5. Your school sets its own policy on what is allowed, and that policy sits above any general advice you read anywhere.
6. Keeping a short log of what you asked and what came back turns the declaration into a five minute job instead of a last day panic.

Emerging Technologies Beyond AI

1. The internet of things puts small networked sensors into ordinary objects, so a bin reports when it is full and a radiator reports its own temperature.
   1. Every one of those devices is an embedded computer with a network connection, which makes each one another door into the network.
2. Augmented reality lays information over a live camera view, which lets a client stand a full size version of your product in their own room before you build it.
3. Virtual reality replaces the view completely, which suits walking through a space that does not exist yet and training where real mistakes are expensive.
4. Biotechnology in materials grows what used to be manufactured: packaging grown from mycelium, textiles made from bacterial cellulose, bioplastics from maize starch.
5. Grown materials are often compostable and low energy to produce, and they are also slower, more variable batch to batch and harder to scale than the plastics they replace.

Judge a New Technology by Who It Helps and Who It Harms

1. Start with who gains first, because early benefits nearly always land with whoever can afford the technology.
2. Then find who carries the cost: the workers whose task was automated, the neighbours of the data centre, the people whose data was used to train it.
3. Work out what happens when it fails, since a wrong film recommendation is an annoyance and a wrong reading in a medical or vehicle system is not.
4. Check who is able to opt out, because a technology nobody can refuse needs a far higher standard of evidence behind it.
5. Follow the whole life: materials mined, energy burned to train and run it, and what happens to the hardware when it is replaced in three years.
6. These questions belong in Evaluating whenever your solution uses a new technology, and they are the same questions the global context of scientific and technical innovation asks.

Testing the Answer on a Real User

1. Put the technology in front of somebody outside your class and watch what they do without explaining it first.
2. Ask them what they thought it was doing with their information, and compare that with what it really does.
3. Record the gap between those two answers, because that gap is the honest finding your evaluation needs.