The Discussion:
Gib started us off talking about a new watch he has that monitors his pulse and O2 levels. This info can be sent wirelessly to a cell phone for alerts and to keep a log. There are a variety of devices at different prices with various features to keep people healthy. Carl reported that they now have a pacemaker smaller than a grain of rice that needs no battery and is completely biodegradable. The device requires no surgery and a replacement can be inserted after several years without needing to remove the old one. This pacemaker is still in the testing phase.
We than talked about a newly coined word, enshittifiation. Cory Doctorow used it to describe the pattern of decline in quality of online platforms, services, and products over time. And of course we used AI to learn more.
Rich mentioned that police shows on American tv tend to present heavy-handed characters, and that he prefers watching Canadian tv shows. Pat suggested some shows including Interrogation Raw, The Interrogation Room, Evil Talks, and See No Evil.
We talked a bit about handwriting and how well AI does (or doesn’t) recognize it correctly. Gib mentioned that MNIST (Modified NIST) is the benchmark for testing AI image recognition.
Then Richard fired up his llama3.1 running on ollama and we started tinkering. Carl suggested that we use this prompt: “Select 2 large three digit numbers, then print in this order, 1 their product, 2 the first prime, 3 the second prime.” We found that it is not uncommon for it to hallucinate a product. Sometimes changing the order to first prime, second prime, product let to accurate results. In this we could see that changing the order of the output could affect the results.
Carl talked a bit about the context window which can affect the accuracy and efficiency of results. In most cases it’s better to have some context given in the prompt or as an uploaded document. However if the content is too much for the context window, but results can ensue. And if the context window is too large, the system will slow down and possibly be prone to hallucination. There are no hard and fast rules for either context window size or temperature.
We had talked earlier about the actors who had played Greg Brady and how various models gave different answers. This led us to adjust the temperature parameter to various values, such as 0, 0.1, 0.3, and 2. The results varied significantly. And we ran out of time…