ektasis by pando

ektasis by pando

  1. Keeping our eye on the ball

    There's a popular tip in most activities that involve motion: focus your sight on your goal. If you're driving a car, stare down the turn; if you're catching a ball, focus on the ball. This tip is an a key to success in most activities. I believe this works for engineering and research as well: a well-articulated vision for the world we want to end up in is critical when the stakes are high. And the stakes are high for AI. I don't think we, as the human race, are doing a great job agreeing on an ideal future for AI development. I'm talking about writing down a story for the average day a human might live 20 years from now.

  2. Culpability of AI is a Major Issue

    When the first AI system commits a serious crime, who is to blame? I don't think we really have a grapple on that as a society right now. The crime may not be requested: a student who tells his AI that he just has to have the set of answers for tomorrow's test might induce a strong misaligned AI to simply hack the computer systems of the teacher to get them. This is a crime, and it is something that could plausibly happen right now. Who is to blame here though? Is it OpenAI/Anthropic? Or is it the kid?

  3. Lab Alignment

    I think the content of this post is self-evident. I write it because the core concept underlying it is what bothers me most about the point in time that we are in. It is helpful to me to write things out sometimes, and I suspect it’ll also be helpful to have something to point to in the future. I don’t believe any of the big research labs can be aligned with the best interests of humanity. Sometimes they appear to be aligned, but this is through the force of will of their leadership. It is not the natural propensity of these organizations. Here is my argument: Each of the big labs (Deepmind and MSL included!)

  4. The Epoch

    In March of 2026 I experienced an acute awareness of something that I would refer to as “artificial general intelligence”. In various forms it had existed before this time: I think AGI development is a gradient and the definition of the term is vague enough that we will hit definitions of it for individual humans throughout time - but to me it happened around March this year. Humans built what I would consider a truly general purpose artificial intelligence. Development has not slowed down. These things were incredible but they are getting more incredible. A couple of years ago a group of people got together and built something called “humanities last exam”. It was a list of questions assembled by human experts in various domains of math and science that required true domain expertise to answer.