Quote from: Seeking Clarification on December 25, 2025, 16:57:25Is there a killer llm or killer app that uses llm yet? Something that has massive impact and actual benefit to actual humans?

I've yet to see it, everything it seems to do is mediocre unless it's very simple specific task like searching through documents for names and numbers which is something that could of arguably be done before the rise of llms.

The most advanced and transformative "killer app" socially beneficial use for LLMs I've seen is in the rollout of agentic AI for software development. As a lifelong (multiple decades) software developer, this past year (really the past 6 months) has been the biggest change, in my personal experience, since the introduction of web search. Maybe it'll be even bigger than that by next year (eg, punch-cards). It not only let's me, as an expert, do things I couldn't do before and at a pace I couldn't before, but it's been interesting to see it lowering the barrier of custom software development (with caveats) to basically anyone that can string a sentence or two together. There's a near infinite amount of software that never gets written because it's too expensive, so I think this latter shift may be an even bigger deal than the former.

There are a few other things I think that are pretty transformative within their domains - language translation at a quality that was impossible before (even with deep learning approaches like Google Translate). Real-time conversational and vision AI (impossible w/o LM-backed "understanding") that while has frivolous or negative uses, is also huge in terms of accessibility (real-time captioning, image descriptions, voice interfaces), language learning, etc.

I also think that being dismissive of "just" improved quality is a mistake. Modern classifiers/rerankers/embedding that take advantage of modern architectures/data/training techniques that couldn't exist without the LLM push are so much better than what came before that I believe they are a quantum leap/different from what came before. Same with the level of object detection, segmentation, etc in modern vision models. While pre-LLM "traditional" OCR like Tesseract can be good, they're terrible at layout and image extraction that VLMs do great on. (For multilingual or real-world images btw, it's no contest, even a hybrid architecture like Tesseract can't compete at all w/ modern OCR VLMs).

There's a lot of emerging applications in STEM, but I don't think I'd count any of them as "killer apps" yet except maybe medical notetaking (which is still error-prone but such a time-saver/QoL improver that it may already be considered a net positive). I do think that the ambient AI note-taking/personal assistant thing is going to be mainstream soon. The AI models are there already, it's just a matter of product/scaffolding more than anything else at this point.

Quote from: Lasdivine on Today at 15:28:19American here, will never see the light of day here

Unfortunately, Americans don't know what a free market is