Any GPT-series model would qualify. Any app using an LM-pretrained transformers for token generation would qualify, if most app interactions include queries to the model.
Currently I use Github copilot, but most interactions with my code editor do not involve copilot, so this does not qualify yet.
For context, I'd estimate I currently spend roughly 30-60 minutes per work week interacting with LM. I work on language model alignment as a PhD student. If I work on a project which requires intensive direct work with an LM on a temporary basis, this does not qualify. I will resolve this question taking an average over a one month period.
Related questions
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ15 | |
| 2 | Ṁ9 | |
| 3 | Ṁ2 | |
| 4 | Ṁ2 | |
| 5 | Ṁ2 |
Resolving this as I frequently at least read copilot's recommendations these days when coding (though it remains highly unreliable). Also query GPT-4 more frequently as I've gotten a better sense of where it can speed me up vs just misguide me when it comes to coding related questions.
@BionicD0LPH1N Yes chat but not API access though that'll likely change soon.
There's a decent chance this will resolve soon. Within the next few weeks I intend to commit to a randomly selected subset of days over a month period where I'll monitor my AI usage time.
@JacobPfau I also find it more worthwhile than before to spend time probing capabilities and biases of LMs.
Does it count if increasingly large language models end up built into apps you use every day, providing higher-quality autocomplete, auto-fill, recommendations, content fill-in or similar? Or does it have to be something like hand-tuning prompts and feeding them to a model to complete/infill (like GPT-3 sandbox) or doing interactive completions (like ChatGPT)?
@ML Yes, if I start using auto-complete for email 15 minutes a day plus 15 minutes of QA interaction this would count. I use auto-complete very infrequently though.
Something I do use more is general spell/grammar check, I will not count this towards the 30 minutes even if a language model is used for this on the backend. I'm stipulating this a bit arbitrarily, but it's mainly because an LM would probably provide only a minimal improvement on that domain.