A follow-up to: [very few worthwhile tasks? (weblink)](/weblinks/2023/06/29/very-few-worthwhile-tasks), with the subject tweet shared below.

As I thought more about the [\@benedictevans tweet](https://twitter.com/benedictevans/status/1674413748272939009) (at bottom) from yesterday, I hit on another, and perhaps clearer, connection to a core discussion in [my dissertation](/diss). One struggle people have with finding effective uses (or developing fruitful practices) for ChatGPT, etc., is _not_ that "very few worthwhile tasks can be described in 2-3 sentences typed in or spoken in one go". The struggle is that very few 2-3 sentence prompts/queries, in particular contexts, can effectively describe the imagined referent worthwhile tasks. The problem here is not that worthwhile tasks cannot be described in 2-3 sentences. The problem is finding some of those, perhaps very few, 2-3 sentences, for particular task contexts, that connect the searcher with their next steps in relation to their worthwhile task. The data engineers I interviewed were particularly successful, in part, because their work was organized in such a way that they could grow their queries out of [seeds in the code or text or conversations](/diss/extending_searching#search-seeds) they were working with. So, perhaps "in one go" is apt and we need to think about how to see and seed how to construct effective queries/prompts.

Here is the tweet in [very few worthwhile tasks?](/2023/06/29/very-few-worthwhile-tasks) (weblink post):

[highlighting added]

\@benedictevans

The more I look at chatGPT, the more I think that the fact NLP didn’t work very well until recently blinded us to the fact that very few worthwhile tasks can be described in 2-3 sentences typed in or spoken in one go. It's the same class of error as pen computing.

On Twitter Jun 29, 2023