This week, we’ll dig into a wide variety of topics, and feel free to spend more time on the topics you find the most interesting!
AI safety is the concern that AI could cause human extinction and should be a global priority alongside other societal risks such as pandemics and nuclear war. Given rapid AI progress that is unlikely to slow down, AI safety emphasizes that we need to invest research into aligning AI incentives with human values to avoid catastrophic scenarios. For example, one common thought experiment is the paper clip maximization problem:
An AI system instructed to maximize paperclip production will seek power/resources to fulfill its objective. As its only goal is to produce more paperclips, it would not care about human values and could take extreme actions (such as converting hospitals into paperclip manufacturing plants or manipulating humans into doing its bidding) that satisfy its programmed objective.
Recommend time: 11 minutes In this article, the Machine Intelligence Research Institute (MIRI) provides a basic outline of why we should worry about superintelligent AI and the danger it may bring.
Recommend time: 7 minutes. Article PDF without paywall In this article, Richard Ngo provides context to why AI has progressed so quickly and why AI safety researchers are concerned about the rapid progress in the field.
Recommended time: 4 minutes In this reading, Anthropic, an AI safety research company, describes an approach to building value-aligned systems by providing a constitution for AI agents to assess their responses.
In the US, educational performance for teens in reading and math has been stagnant for 15+ years. There is also a significant academic achievement gap between students from high versus low socioeconomic backgrounds. Because education attainment is a big driver of social mobility, this achievement gap contributes to growing wealth inequality in the US.
Around the world, there are 260 million out-of-school children. Research has indicated that each year of primary schooling increases earnings by 8 percent.
Recommend time: 6 minutes Massive open online courses (MOOCS) are often touted as the solution for universal education. Unfortunately, due to low completion rates and inequity in access, they have not brought about the education revolution that many thought might happen.
Recommend time: 4 minutes In this article, OpenAI highlights some of the applications of its newest large language model, GPT-4, in their partnership with Khan Academy to produce a personalized learning assistant.
Recommend time: 10 minutes. Article PDF without paywall In this article, journalist Kevin Roose poses an argument for why banning ChatGPT won’t achieve the intended effect, and that the best solution might actually be to incorporate it.
Recommended time: 6 minutes In the Context section, the World Bank highlights key questions and opportunities in the edtech world in low- and middle-income countries.