Mitchell Bosley
@mitchellbosley
Followers
301
Following
2K
Media
13
Statuses
239
Postdoc at CENIA | Previously at the Schwartz-Reisman Institute at UofT | PhD from University of Michigan Poli Sci
Santiago, Chile
Joined January 2009
RT @JohnHolbein1: Did you know that your siblings can influence whether you vote?. Well, now you do!. Social scientists have long known tha….
0
17
0
RT @YamilRVelez: Excited to share a new working paper with @SemraSevi and Don Green. We built a chatbot-based Voting Aid Application (VAA)….
0
20
0
RT @apsrjournal: From our new issue: "Improving Probabilistic Models In Text Classification Via Active Learning" by MITCHELL BOSLEY (@mitch….
0
14
0
RT @JohnHolbein1: Prejudice is common in modern societies. Reducing that prejudice is difficult to do, particularly in a cost-effective a….
0
20
0
I don’t really get the point of exercises like this if the end result (the “book” in this case) isn’t shared. The problem is not “how can we get a language model to write a book”, it’s “how can we make it write a book that’s not just slop?”.
Let's build a multi-agent book writer, powered by DeepMind's Gemma 3 (100% local):.
0
0
2
Upweighting the importance of in-person exams relative to e.g. term projects is a short-term fix. There has to be room for students to develop research and essay writing skills, and there has to be an incentive structure that rewards students who take the time and effort to
0
0
3
both of these are bad.
Left is Gemini, right is GPT-4.5. This model is WAY better at nuance in writing
0
0
1
this imo speaks to the duality of current research into LLM capabilities: on one hand, very valid warnings not to anthropomorphize models; on the other; the abstract was clearly written with the help of LLMs (“nuanced” in the last sentence being the clear tell).
an efficient market would see apple's stock down 10% after publishing this
0
0
3
I’ll be presenting new work showing how we can use AI models from ChatGPT, Claude, and Meta to do qualitative content analysis from political debates at scale using prompts that combine detailed code books with human-coded examples
1
0
7
If you’re at APSA at 8am Thursday morning and interested in LLMs and AI for measuring concepts from text data, come see @yaoyao_dai @hauke_licht @miserlis_ and I present, with comments from @arthur_spirling and Walter Mebane!.
1
5
32
RT @verohurtado_: Oficialmente, hoy empecé a trabajar como Profesora Asistente en @GobiernoUAI. Muy feliz de empezar esta nueva etapa en m….
0
6
0
Great to see this finally accepted. Thanks @yshiraito @TedEnamorado and Saki Kuzushima for all your hard work!.
Our paper on the use of active learning for text classification is forthcoming at @apsrjournal! The author version of the accepted manuscript and the R package to implement our proposed approach are available at
2
1
18
fwiw, my prediction re: AI-generated writing is that there is likely to be a separating equilibrium where “high” types intentionally write in ways that distinguish themselves from AI, rather than a pooling equilibrium where everyone just churns out pablum.
I’m assuming chatGPT will basically kill off the college admissions essay as a value proposition. Of course, there will also be downsides.
0
0
3
RT @verohurtado_: I'm very happy to share that I'll be heading back to New Haven this Fall as a postdoctoral fellow at @YaleMacMillan! Afte….
0
4
0
RT @verohurtado_: So excited for my first @APSAtweets conference! I'll be presenting my research on #misinformation during the 2021 Peruvia….
0
9
0
I studied the writing routine of 500 legendary writers. Turns out, great writing is simple—if you do these 6 things every day:.
0
0
3