Crawford, Atlas of AI (Excerpt)
I enjoyed reading the article, “Ethics at arm’s length”: Kate Crawford book excerpt because it talks about the downsides of AI, like taking user data. First of all, there seem to be a lot of bias generated by AI as most systems seem to reflect the “characteristics of the dominant voice in their code and in the data they use to learn, and that is male and white.” This is the result of when “researchers slowly disconnected with the subjects they were investigating.” I learned most university-based AI research is done without any ethical review process which is surprising because I assumed that universities cared about ethics, especially regarding user data. I was surprised when I read the Calgang database case because it seems to have many inaccuracies such as adding infants, and adults without any charges to the database from which their names can’t be removed now. The database also has a disproportionate number of Black and Latinx people on the list which proves AI’s biases and inaccuracies. Separating ethics from scientific research has become very problematic because it can harm subjects, especially in the age of modern AI. Large amounts of data are being collected everywhere online and by every AI model which is harmful to humans. As stated at the end of the article, “Weizenbaum hoped that scientists and technologists would think more deeply about the consequences of their work — and of who might be at risk. But this would not become the standard of the AI field. Instead, data is more commonly seen as something to be taken at will, used without restriction, and interpreted without context.” I now realize how big of a problem this is and hope that more people become aware of it and policy makers take action to propose regulations on AI companies so that our data is protected.
IBM What are LLMs
The article about “What are LLMs” by Cole Stryker at IBM was very interesting because it went into depth about how Large Language Models (LLMs) work. It focused on the history, training, and even limitations of LLMs. Since I am mostly uneducated on the topic of Artificial Intelligence, I learned many new concepts such as the process of tokenization where large amounts of texts are broken down into machine readable units called tokens to make it easier for models to understand and handle words consistently. I also learned about embedding which allows models to learn semantic associations between words and fine-tuning which adapts a general-purpose model to specific tasks like summarizing and classification. Additionally, Reinforcement learning from human feedback (RLHF) seemed very interesting to me because the author states it can be “adjusted to respond in a way that’s more casual, humorous or brand-consistent” and I believe I have experienced that with different generative AIs. Some seem more friendly, while some don’t, and it is probably due to RLHF. I also found out that generative AI and LLMs are not the same as generative AI is a part of LLMs which was informative. I always thought that LLMs predicted the “final answer” before giving a response, but it turns out that they use everything they learned during training to predict one token at a time and take their best guess at the answer at every step. In the end, the author talks about the limitations of LLMs, which include inaccuracy, amplifying biases, and increased cost/environmental concerns.
Sep 15
