@phrees @ErikJonker @f4grx @jeffjarvis I'm surprised you didn't sprain your ankle shifting your argument that hard. Plus, there are ethical and profitable applications for the tech, especially in medicine. But enjoy your crusade.

@LouisIngenthron @ErikJonker @f4grx @jeffjarvis

Even in medicine, with tightly constrained training data, AI is totally capable of producing bullshit.

> In our dataset, images with rulers were more likely to be malignant; thus the algorithm inadvertently “learned” that rulers are malignant.

https://www.sciencedirect.com/science/article/pii/S0022202X18322930

@phrees @ErikJonker @f4grx @jeffjarvis That's a *bug*, not "bullshit". All software has bugs. They can be fixed.

@LouisIngenthron @ErikJonker @f4grx @jeffjarvis How do you debug an LLM without telling it that its output is bullshit?

@phrees Every algorithm ever written by any computer scientist ever has the exact same problem:

Put garbage in, you'll get garbage back out.

So, you debug an LLM by getting rid of the garbage in the training data. In your example above, you'd either scrub the rulers from the images entirely, or more likely you'd make sure there was an even distribution of rulers among the various diagnosis category input data sets.

@LouisIngenthron @phrees @ErikJonker @f4grx @jeffjarvis Are you talking about LLMs or neural networks? I don't think I would accept an LLM's diagnosis, but neural networks seem to be producing some uncanny results.

@jhavok Yeah, I was referring to the larger ML field.