A pdf is not a good format for human knowledge retrieval. This is particularly the case when it comes to technical documents such as research papers which comprise features like tables, figures and references in addition to text. The format makes sense in print when you have limited space and need to pack in all the relevant information. In other words it makes sense for storing information in a spatially efficient format.
However these very qualities make it extraordinarily difficult to read particularly if you have problems paying attention. You are frequently obliged to jump between pages and whilst many pdfs hyperlink figures, tables and references they usually lack back navigation. The double column format, another feature designed for spatial efficiency, is highly cumbersome for reading, increasing the need to move back and forth, side to side and, when reading on smaller devices, zoom in and out 1.
I find it astonishing that whilst AI has moved by leaps and bounds and has become infamous for the number of research papers published and there are so many tools and applications to find and share information and data about AI, there has been no standardised tool for viewing papers in a convenient format. The last time I looked there was arXiv vanity, which renders papers as HTML and another tool whose name I forget but which I believe uses AI to help with the conversion. In addition there are features in file readers like liquid mode in Adobe.
Of these I have mainly tried arXiv vanity but there is a certain clunkiness involved. First you have to find the url and paste it. Then it sometimes fails to render. And even when it does the rendering of key features like equations leaves much to be desired. The Adobe liquid mode, which I have also occasionally used, has proven similarly disappointing for rendering anything other than regular text. Moreover neither of these tools seek to address the navigation issue 2.
I have been thinking about the problems of the pdf format for papers on and off for a while. Then when I got Covid a while ago I found myself compulsively reading feel-good e-books on Kindle. These were not normally the sorts of books I would favour and at first my preference seemed to be due to their soothing and undemanding subject matter and style. However I realised they were easy to read in other ways as well. I began to wonder if one of the reasons I could not consume research papers as addictively was because of the inconvenience of the format, which made it difficult to get into them to the stage where the interest of the subject matter took over.
What I would want is something equivalent to an e-book format. There are really only a few important features. First, the document should be easy to read on any device which means having the ability to adjust font sizes, a flexible layout and no issues with rendering equations and tables at different scales. It is a lot more comfortable to read on a lightweight device like a phone (or a Kindle) and that can facilitate entry into the flow state. Another essential requirement is ease of navigation between parts of the document. Ideally it should be also compatible with existing e-reader applications like Kindle.
It has been some months since I last investigated alternatives to the pdf format and it is quite likely that by now some improved solution exists. (If not perhaps I ought to build one myself). Let me note though that the use of ChatGPT to summarise or explain a paper would not constitute a solution. It simply sidesteps the issue and is more akin to making a blog post or a video about a paper rather than improving access to the original material. However even if it turns out that an effective solution now exists, the fact that it has taken so long to come about, in my view, undermines the ideals of openness and democracy that supposedly underpin present day AI research.
- These problems become much more noticeable in the electronic format. For instance it is easier to flip between printed pages without losing your place. ↩
- I think the LiquidText handles this issue but I’ve not tried it. However I am not sure it would be an ideal solution due to its platform dependence. ↩