Gary Ackerman @garyackerman@qoto.org

I used to think we were becoming a more educated population. The last 5 or so years proved me wrong.

Training data differs vastly: Traditional AI uses smaller, labeled datasets. Generative AI models are trained on massive datasets (millions of images, vast amounts of text for LLMs) requiring significant resources.

“The danger to society is not merely that I should believe wrong things, though that is great enough; but that I should become credulous, and lose the habit of testing things and inquiring into them; for that it must sink back into savagery.” -W. K. Clifford in 1877

Rationality does not cater to our whims.

How should we respond when those in education, but not teaching, feel they don’t need courses in teaching in their studies?

Anyone else get nervous when the incoming email traffic is light?

Sometimes the best response is to slowly back away.

When the argument is “it’s best for the students,” it isn’t.

We all study textbooks with linear models. Depending on the field, they are some variety of analyze, plan, deploy, evaluate. Does anyone follow these in real life? Maybe we should spend less time on models and value the diversity and variation that forces us to nonlinear design.

Anyone else noticing the educator who is reluctant to record, caption, and publish their sessions is the same one who insists all training sessions be recorded and feature step-by-step instructions in hardcopy?

So much IT frustration arises from ignoring recommended procedures.

One way to avoid “there is no way I can make all of my stuff accessible” is to not build your class with inaccessible materials in the first place.

Generative AI isn't a silver bullet: risks include incorrect or biased content (hallucinations) and potential misuse. Enterprises need a strategic approach with clear objectives and safeguards.

Traditional AI (Narrow AI) operates using classical data science for prediction or classification tasks. It's deterministic, acting within predefined boundaries and rules.

Need to see how well your sample data fits a predefined population distribution? The chi-square goodness of fit test is your go-to! It compares observed vs. expected frequencies. #Statistics

The "objective measurements" advocates sought for standardized tests are only possible when the environment's variation is removed. This is not feasible with human learning. #EducationResearch

Students live in a "rich and variable environment" that cannot be controlled like a laboratory experiment. This makes it difficult to attribute performance changes solely to instructional practices. #LearningEnvironments

Standardized testing was "doomed to fail from the very beginning". The rationale to measure "what works" in education, like a science experiment, was fundamentally flawed. #EdReform

"Nothing is more dangerous than a dogmatic worldview - nothing more constraining, more blinding to innovation, more destructive of openness to novelty." -- Stephen Jay Gould

A blog post, now with audio, focusing on the problems of testing as a data source in education. https://hackscience.education/2018/09/24/why-standardized-testing-flopped/