This is a nice thread. I similarly have been admiring recent AI demos but struggle to articulate why I'm not in an excited frenzy. I think it's bc the results are precisely what I'd expect from throwing more data at more compute. I don't think we're any closer to "productizing"
AI, or turning these innovations into tools for creators / writers / etc. Are we (tool builders) measuring metrics tool users care about, e.g., iteration speed? Last time I tried to use an image generation model it took me many minutes, which is unacceptably long
I think Github Copilot was one of the biggest leaps we've had in AI productization. Lots of training data (raw code) available, and the inference speed is slow but faster than the "think time" (Shneiderman et al.) of a programmer so it enhances, not derails, productivity, but
every time I type // TODO or # TODO, some rando's name is suggested, which gives me a really "yikes" feeling. I don't hear AI product teams talking about such feelings, what they will do to mitigate them/optimize for a better user experience rather than raw model loss
Re: user experience -- I commend the openAI teams for having decent toxicity filters baked into their playgrounds, but as everyone & their parent & sibling & child is starting a LLM or generative model or AI company, I want to remind them that building product is more than
wrapping a group of researchers in a corporation, raising big $$, and continuing to optimize for ML conference metrics. These are new kinds of tools & thus there is no playbook. I'm really looking forward to future Twitter discourse on how to solve creator/writer/developer needs
Loading suggestions...