How should we conceptualize language?
Lately, there's been much talk about what LLMs do as being in some sense inauthentic. I'm not going to talk about consciousness here. It's not a concept that I think has the right substance to be useful for my purposes. But I take issue
Cross-posted on the C&I substack Everything you really want people to understand should be encoded as a video game. By this, I mean we make a game that's stimulating (and potentially even enjoyable), where understanding the core concept is on 'the critical path'
If someone is using an unfalsifiable theory to tell you something, they are really using a metaphor and you should treat it like one. Listen to them. Stop inventing logical fallacyesque copout why someone should be treated like they’re in a formal debate instead of trying to communicate
More than a year ago, Armen asked me "What's the next challenge for LLMs?" My answer was: 'Things that can't be specified in a few sentences.' This is still my answer, but it's a bad answer. Describing something negatively is
I'm infuriated by the use of AI today. Have you ever been reading some historical account and thought 'They have all the tools to concoct the solution we use today, why don't they just use it solve their Very Big Problem?' This is how
I spend most of my time thinking about how to conceptualize LLMs, understand their behavioral dynamics, and measure what they are capable of. I want to find clear ways of thinking about LLMs: do instruction-tuned LLMs just have low-diversity or are they really more consistent than base LLMs
Give an AI a story about a father teaching his daughter to ride a bicycle. It will extract themes: coming-of-age, trust, the transfer of knowledge through generations. Give it those themes, it will generate new stories: a master chef's first lesson with an apprentice, a grandmother