apwheele
Joined 5 years ago
Karma
1,909
Submissions
407
Member since
Nov 10
Recent submissions
Comment · 13 days ago
I am currently using Perplexity for https://veruscite.com/. Verify citations actually exist -- so get in a list of 100, half I can check with crossref, but the other half I need to do a web search and fetch to make sure they exist.
So I need higher rate limits to not make the tool come to a crawl. I probably will not consider DeepSeek API here (I have tried DeepSeek hosted at US providers, like Perplexity). The other search/fetch tools I try do not have great web coverage of Asian journals, but the fact DeepSeek API is not zero data retention I cannot get over.
Comment · 17 days ago
I know what you mean by "cannot write out of a paper bag". Agree AI won't help those folks (I honestly don't know how anyone has enough patience to help them).
My above prompt does indeed work well for me, but agree would likely not catch those example errors you list.
Comment · 17 days ago
I am maybe not using copy-editing the way you are (feel free to correct me what is more typical in the field). My prompt is not much different than "review the text for minor grammar and spelling errors, and list the corrections you would suggest". I accept the large majority of the suggestions.
For technical review of the content, I just send it out to other experts.
Comment · 21 days ago
I am confused by this as well, shouldn't it be all white listed sites at the start?
Comment · 26 days ago
Put in the todo list to check out. If you cut web search costs even further for flash/fast search (similar to Perplexity, which is $1 per 1000 for fast search) would make it more competitive (at least for my app!)
Comment · 26 days ago
So I wish I could use Google for https://veruscite.com/, but the number of Google searches are a hard cap on the account! So yes that is fine for agentic coding, but for an app that relies on web-search is not sufficient.
I am currently using Perplexity fast search and fetch, and I am happy with that. I would try our Ceramic.ai, but I need to be able to fetch the pages as well (I do not want summaries).
Comment · 29 days ago
If you check out https://veruscite.com/hallucinations there are much more banal examples than that article.
Part of the forensic work with this, I won't ever be able to tell the difference between "I as a human messed up" vs "LLM go brr and filled in everything". So I think editor discretion when to correct vs when to retract is fine. (I do think that PLoS One correction should have been retracted as well though.)
Comment · 1 month ago
While this is true, it is still possible to take the probabilities and make determinations of recall and false positives using conformal sets, https://crimede-coder.com/blogposts/2026/ConfClassification
(Or just use a model to re-calibrate the probabilities, I like the conformal approach though as those rates are what I often care about.)
You just need some labelled data to generate the "corrected" probabilities (or thresholds to meet the specified error rates).
Personal blog at https://andrewpwheeler.com/
Consulting at https://crimede-coder.com/
Large Language Models for Mortals: A Practical Guide for Analysts (book), https://crimede-coder.com/blogposts/2026/LLMsForMortals