Vasco Brazão 🏳️‍🌈
LightNews LightNews
banner
vascobrazao.bsky.social
Vasco Brazão 🏳️‍🌈
@vascobrazao.bsky.social
330 followers 950 following 290 posts
Recovering behavioral scientist, posing as a statistical consultant. Applied stats #RStats, neurodiversity, learning Ukrainian, writing things
Posts Replies Media Videos
vascobrazao.bsky.social
calling all #AI haters (complimentary): I participated in a project evaluating LLM-supported math tutoring and would love some criticism/references/whatever you've got that might influence follow-up projects.

technical report here (abstract in picture): storage.googleapis.com/deepmind-med...
abstract of the technical report. reads: One-to-one tutoring is widely considered the gold standard for personalized education, yet it remains prohibitively expensive to scale. To evaluate whether generative AI might help expand access to this resource, we conducted an exploratory randomized controlled trial (RCT) with 𝑵 = 165 students across five UK secondary schools. We integrated LearnLM—a generative AI model fine-tuned for pedagogy—into chat-based tutoring sessions on the Eedi mathematics platform. In the RCT, expert tutors directly supervised LearnLM, with the remit to revise each message it drafted until they would be satisfied sending it themselves. LearnLM proved to be a reliable source of pedagogical instruction, with supervising tutors approving 76.4% of its drafted messages making zero or minimal edits (i.e., changing only one or two characters). This translated into effective tutoring support: students guided by LearnLM performed at least as well as students chatting with human tutors on each learning outcome we measured. In fact, students who received support from LearnLM were 5.5 percentage points more likely to solve novel problems on subsequent topics (with a success rate of 66.2%) than those who received tutoring from human tutors alone (rate of 60.7%). In interviews, tutors highlighted LearnLM’s strength at drafting Socratic questions that encouraged deeper reflection from students, with multiple tutors even reporting that they learned new pedagogical practices from the model. Overall, our results suggest that pedagogically fine-tuned AI tutoring systems may play a promising role in delivering effective, individualized learning support at scale.
November 13, 2025 at 5:53 PM
vascobrazao.bsky.social
Finally!

#AI #GenAI
screenshot of empty day on Google Calendar with a fake gen-AI-like banner saying "generate a life". Text on top says "New Google Cal Feature Just Dropped"
October 21, 2025 at 11:28 AM
vascobrazao.bsky.social
amazing. it appears google translate has hurt itself in its confusion, see:
screenshot from a greek wikipedia page. reads: "Σε πολλές ξένες γλώσσες, η σημείωση της άνω τελείας γίνεται με το ελληνικό ερωτηματικό (;)." screenshot of the same text automatically translated into English. reads: "In many foreign languages, the colon is marked with the Greek semicolon (?)."
October 9, 2025 at 4:58 PM
vascobrazao.bsky.social
*shudders*
screenshot of the offending semicolon
October 9, 2025 at 12:58 PM
vascobrazao.bsky.social
just started reading, found a cute typo:
May 20, 2025 at 2:47 PM
vascobrazao.bsky.social
hey google docs, this is #RStats erasure 😠
screenshot from google docs, from the "insert" menu, "building blocks" submenu, "Code block" submenu. we can see several programming languages in alphabetical order. after "Python" comes "Rust" and then "Shell", nary an R to be seen.
May 19, 2025 at 5:24 PM
vascobrazao.bsky.social
se considerarmos direita vs esquerda parece que tem correlação: quanto mais à direita, menos confiança em cientistas
screengrab of the results for correlates of trust in scientists for Brazil, from https://www.tisp-manylabs.com/explore-tisp-data
January 20, 2025 at 11:54 AM
vascobrazao.bsky.social
#RStats puzzle. What happens after the third call, and why?

#DataScience #StatsSky
Three R commands, the first two are followed by their output. Literally, below: > c(!is.na(NA), !is.na("NA")) [1] FALSE TRUE > FALSE + TRUE [1] 1 > !is.na(NA) + !is.na("NA")
November 27, 2024 at 12:11 PM
vascobrazao.bsky.social
citation desperately needed... And "the soon-to-be for-profit enterprise whose profitability depends on saying so said so" is less than evidentiary 🙄
first two paragraphs from the piece, highlighted: "These combined efforts will likely lead to Artificial General Intelligence (AGI), where machines have human-like cognition, perhaps within just a few years."
November 15, 2024 at 2:46 PM
vascobrazao.bsky.social
things i will never understand
screenshot of warning on the PsycNet website about session timeout. Warning reads "Your APA PsycNet session will timeout soon due to inactivity."
October 22, 2024 at 10:23 AM
vascobrazao.bsky.social
ironically, the feed seems to be down at the moment (ping: @steveharoz.com )
October 18, 2024 at 8:49 AM
vascobrazao.bsky.social
!!!
Detected: 130
October 18, 2024 at 8:22 AM
vascobrazao.bsky.social
Bullshit, hard and soft.
Bullshit (general) Any utterance produced where a speaker has indifference towards the truth of the utterance. Hard bullshit Bullshit produced with the intention to mislead the audience about the utterer’s agenda. Soft bullshit Bullshit produced without the intention to mislead the hearer regarding the utterer’s agenda.
October 15, 2024 at 3:28 PM
vascobrazao.bsky.social
Finally reading "Is ChatGPT Bullshit?", which argues the falsehoods in the outputs of language models are not "hallucinations" but "bullshit", and explains why the distinction is useful.

link.springer.com/article/10.1...

#ScienceReading #ArtificialIntelligence #PhilSky #PhilTech #Psychology
Abstract of the paper, reads: "Recently, there has been considerable interest in large language models: machine learning systems which produce human-like text and dialogue. Applications of these systems have been plagued by persistent inaccuracies in their output; these are often called “AI hallucinations”. We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs. We distinguish two ways in which the models can be said to be bullshitters, and argue that they clearly meet at least one of these definitions. We further argue that describing AI misrepresentations as bullshit is both a more useful and more accurate way of predicting and discussing the behaviour of these systems."
October 15, 2024 at 2:58 PM
vascobrazao.bsky.social
don't know if to laugh or cry
asked Notion AI how to disable it, it responded: I apologize, but I couldn't find specific information on how to disable Notion AI in the search results provided. However, as Notion AI is an integrated feature of Notion, it's not typically something that can be fully disabled. If you're concerned about Notion AI's usage or visibility in your workspace, you might consider: - Not using the AI features when they appear as options - Adjusting your workspace settings to limit AI usage, if such an option is available - Contacting Notion support for more specific guidance on managing AI features in your workspace If you need further assistance or have specific concerns about Notion AI, I recommend contacting Notion support directly. You can do this by pressing the "Get Notion support" button below in the chat. They will be able to provide you with the most up-to-date and accurate information regarding AI features and any potential options for limiting or disabling them.
October 5, 2024 at 10:08 AM
vascobrazao.bsky.social
it is always disheartnening to see leftists who don't recognize russia for the violent colonial empire that it is. a new podcast can help educate those who are open to it -- listen to Matryoshka of Lies wherever you get your podcasts! spotify link below

#UkrainianView #Russia
Astronaut "always has been meme" with the focal question "wait russia is a violent colonial empire?"
June 13, 2024 at 3:00 PM
vascobrazao.bsky.social
the questions stare back at you
Screenshot from chapter 3 with the following text highlighted: However, let’s ask ourselves the following important questions: How easy would it be for someone else to rerun the analysis? How easy would it be to update the analysis once new data gets published? How easy would it be to reuse this code for other projects? What guarantee do we have that if the scripts get run in 5 years, with the same input data, we get the same output?
June 13, 2024 at 11:01 AM
vascobrazao.bsky.social
Finally taking the time to learn how to create fully reproducible analyses -- many thanks to
@brodriguesco.bsky.social
for this phenomenal resource raps-with-r.dev
#rstats #stats #AcademicSky #PsychSciSky #sociology
Screenshot from the linked website, a blogdown book called "Building reproducible analytical pipelines with R"
June 13, 2024 at 10:43 AM
vascobrazao.bsky.social
In the next section, the authors suggest psychologists start asking more questions amenable to predictive analysis. But I don't quite get why. E.g, in the example in the screenshot, what have we actually learned about human personality?
Screenshot from page 18, highlighted text: "In addition to unambiguously demonstrating that personality can be reliably inferred from online footprints, a major advantage of the Kosinski et al. study is its ability to provide good comparative estimates of how well different traits can be predicted using multivariate predictive models"
May 3, 2024 at 4:06 PM
vascobrazao.bsky.social
In the section "Increasing interpretability", they cover 5 benefits of machine learning approaches, the last one being thath "a prediction-focused perspective can help gain a deeper understanding of the general structure of one’s data" (see also screenshot) but I'm suspicious. Now we're using >
first paragraph of page 18, highlighted text: "And if very complex deep learning architectures perform much better than virtually all other models—as is increasingly the case in many psychologyrelated domains such as computer vision and natural language processing (LeCun et al., 2015)—one may want to consider the possibility that simpler, more interpretable models are simply not adequate for explaining the phenomenon of interest."
May 3, 2024 at 3:58 PM
vascobrazao.bsky.social
They give examples, like multilevel modelling, where introducing bias can be beneficial because the reduction in variance helps to reduce the total prediction error, and suggest that it's helpful to view p-hacking debates through this lens.
Last paragraph of page 7 previous paragraph continued, now on page 8. Highlighted text: That is, one can construe a research strategy that favors liberal, flexible data analysis as a relatively low-bias but high-variance approach (i.e., almost any pattern in data can potentially be detected, at the cost of a high rate of spurious identifications) and an approach that favors strict adherence to a fixed set of procedures as a highbias, low-variance approach (only a limited range of patterns can be identified, but the risk of pattern hallucination is low)
May 3, 2024 at 1:45 PM
vascobrazao.bsky.social
Next they go into the difference between bias and variance (how much the mean of your estimates deviates from the true mean vs how spread out around their mean your estimates are), suggesting that decomposing error into both components can be more informative than just looking at mean square error.
Figure 2 from the paper, illustrating the difference between bias and variance by displaying different patterns of dart throws onto four bullseyes. Horizontally, the quadrants are split into low bias on the left and high bias on the right; vertically, we have high variance at the bottom and low variance at the top. Figure 3 from the paper, showing two bullseyes with the same throws depicted on each. The left side shows arrows from each throw to the center of the bullseye and is labelled "sum of squared errors". The right side shows lines going from each throw to a point that represents the mean coordinates of the throw, and an arrow from that point to the center of the bullseye with the descriptor "bias" - this right side figure is labelled "bias-variance decomposition".
May 3, 2024 at 1:31 PM
vascobrazao.bsky.social
@devezer.bsky.social might appreciate this example of how replication can fail to add to our knowledge precisely because it sticks to rather than improves on the methodology of an earlier study
screenshot from second page of the paper. Section starting in "It has been reported that since there are 26..." and ending in "... between cervical cancer and smoking."
April 30, 2024 at 10:00 AM
vascobrazao.bsky.social
The next section opens with an acknowledgement of increased calls to contrain analytic flexibility. But again, I've become convinced that the "exploratory-confirmatory" distinction is overblown and not really relevant for theory building. E.g., read "Paths in Strange Spaces" osf.io/preprints/ps...
April 18, 2024 at 5:56 PM
vascobrazao.bsky.social
Here again, I've been more convinced by attempts to explicate the role of data and statistical evidence in the process of theory buildingm, but I'm interested in learning how the competing motivations Y&W describe (flexibility and robustness) can be balanced.
April 18, 2024 at 5:42 PM