Brief Notes: Experimenting with an Agentic AI Browser (Fellou.ai)

Recently, Alan Liu posted the results of an experiment he did with a new Agentic AI Browser called Fellou.ai. This is not another GenAI chatbot, but rather a piece of software you download to your computer. 

For those who haven't encountered this term before, "Agentic AI" implies a generative AI that has the ability not just to find and summarize information or generate text based on your queries, but to actually perform real-world tasks you specify. A common use case might be to ask an Agent to research travel plans to another city along parameters you give it, and then actually book plane tickets and hotel bookings for you.  

When you install Fellou.ai, it does ask whether you want to give the browser access to your passwords and logins or import them from other Browsers. I declined to do this, since this is a new company (based in China, I believe) and I'm not sure whether I should trust their security. So I'm not making any travel bookings or doing financial transactions with them anytime soon. 

1. Modifying Alan Liu's prompt for my own research question. Instead of anything involving passwords and financial transactions, for my first experiment I decided to emulate what Alan Liu did, but with a research project of my own related to work I've been doing related to Canonicity and the Digital Humanities. The prompt I used was as follows: 

Search the web, including scholarly resources such as those found in Google Scholar, Web of Science, JSTOR, and so on to collect 50 essays arguing for expanding or transforming the Anglo-American Canon of Literature, As the output of your search, create a spreadsheet in which you collect quoted excerpts of arguments for expanding or transforming the canon. Include citations of sources, where citations include where possible author(s), titles of works, the publication or website venue, the date of publication when that is declared, and URL, DOI, ISBN, ISSN, or other identifiers.

The results are in the table embedded below. I would describe them as pretty mixed. Some of the sources are actually really good scholarly sources and on target. Others are more bloggy, and a few are a little off-target. That said, the experiment does help me along on my research a bit, and gives me some sources I want to explore a bit more. Here's a sample, and you can see the full table at the link below (I don't have a good way to embed the frame directly onto this page)

https://docs.google.com/spreadsheets/d/e/2PACX-1vRxqhMKXgW5SNGIOIGEDvjCUrnyYn6kIG1cBsCY1EvdYLh7m6HxnyPysT0pRfCdhgGrMnbqfrjgz-5e/pubhtml

The most interesting part of the process was watching the browser work -- it went into Google Scholar and tried various searches, including phrasing configurations I myself might not have thought of. Instead of just doing one general search, it divided the queries up into groupings (African American literature and canonicity, Feminist Criticism and Theory, Latino Literatures, Indigenous literatures, Postcolonial literature, Multiethnic American literature, LGBTQ+ literature, etc). This is actually exactly what I'd already been doing in my own research, though some of the search phrases were ones I might not have thought of using (and they turned out to be productive). 

A sampling of output from Fellou.ai:

Content Coverage: 

The spreadsheet includes diverse perspectives on canon expansion: 

Decolonizing Canon (Entries 1-5, 20): McDonald, Gugelberger, Bakshi, Bernard, Chin African American Literature (Entries 1, 7-8): Corse & Griffin, Fox, Chin 

Latinx/Hispanic Literature (Entries 9-13): Texas Monthly, Dalleo & Saez, Latino Studies 

Native American/Indigenous Literature (Entries 14, 23-32, 50): Krupat, Momaday, Silko, Erdrich, Harjo, Orange, Alexie, Jones 

Queer Theory & LGBTQ+ Canon (Entries 15-19, 33-45, 47, 49): Butler, Sedgwick, Pruitt, Ferguson, Muñoz, Snorton, Hall, Baldwin, Bechdel, Lorde, Kushner, Woolf, Winterson, Norton 

Feminist Theory (Entries 21-22, 46, 48): 

Feminist scholars, Wollstonecraft, Woolf, Lorde 

Comparative/Multicultural (Entry 6): Canon Reformation in Comparative Literature

In the end, I might have benefited more from watching the Browser work and "think" than from the actual output it produced. (One other thought: if I had allowed the browser access to my library logins to get fuller access to scholarly databases, the results might have been better.) 

Grade: B+

2. Trying to use the Fellou.ai to find and download some files from Project Gutenberg. 

For my second experiment, I thought I would try something relatively simple to save me some time. I have been meaning to make a small corpus of modernist authors to do some text analysis queries related to gendered language in BookNLP. I gave the browser a list of 35 or so authors, mostly in English (or likely to have public domain translations in English) whose works might or might not be available on Project Gutenberg. (For the moment, I did not include many African American writers since I already have all of those in my African American literature corpus.)

Prompt used: Using the Project Gutenberg website at https://gutenberg.org, locate and download the Plain Text-UTF8 (.txt) files for each of the following authors: Virginia Woolf, James Joyce, D.H. Lawrence, Ernest Hemingway, Ford Maddox Ford, Radclyffe Hall, William Faulkner, Dorothy Richardson, Marcel Proust, Joseph Conrad, Djuna Barnes, Mina Loy, E.E. Cummings, Bryher, William Carlos Williams, Jean Rhys, Mary Butts, E.M. Forster, Hope Mirrlees, John Dos Passos, Knut Hamsun, Katherine Anne Porter, Rebecca West, Nathanael West, Wyndham Lewis, Sherwood Anderson, Jean Toomer, T.S. Eliot, Ezra Pound, Gertrude Stein, Katherine Mansfield, H.D., Marianne Moore, Hart Crane, F. Scott Fitzgerald, and save them to the folder C:\Users\amard\Downloads.

Watching the browser work, it seemed like it was working -- it was finding files and trying to download them somewhere. But perhaps the list was too ambitious -- at a certain point, the Browser crapped out and told me I was out of Tokens ("Sparks"), and the process could not be completed as requested. (One thing I should have specified was that I only needed English-language versions.) Unfortunately, though it worked through 20 or so of the authors, the Browser did not in the end download anything at all to my computer -- and I used up all my "Sparks" without anything at all to show :-(  

Admittedly, this is not a hard project and I could just do it by hand in an hour or two. 

Was it a good use of my time and AI resources to try this? Hard to say, but I'm leaning towards 'no'. 

Grade: F