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
Reposted by Vasco Brazão 🏳️‍🌈
rawx.bsky.social
What an amazing shot. www.instagram.com/p/DRBAaqLkvye
November 16, 2025 at 9:34 AM
Reposted by Vasco Brazão 🏳️‍🌈
sailorrooscout.bsky.social
It shows how vaccines that educate the immune system against one enemy may inadvertently help it fight others. For humanity, that implies that what is learned during a worldwide health crisis may go on to save lives in ways unforeseen.
November 14, 2025 at 5:12 PM
Reposted by Vasco Brazão 🏳️‍🌈
sailorrooscout.bsky.social
regardless of their tumor type. If it holds up in clinical trials, this discovery could fundamentally change the way cancer is treated everywhere. By linking mRNA vaccines to existing immunotherapies, doctors might be able to strengthen their patients' immune defenses without having to use
November 14, 2025 at 5:11 PM
Reposted by Vasco Brazão 🏳️‍🌈
sailorrooscout.bsky.social
GOOD NEWS! A groundbreaking study in MORE THAN 1,000 people reveals that cancer patients who received a COVID-19 mRNA vaccine within 100 days of starting immunotherapy lived dramatically LONGER than those who didn’t. In some cases, their survival rate was DOUBLED.
November 14, 2025 at 5:00 PM
Reposted by Vasco Brazão 🏳️‍🌈
vascobrazao.bsky.social
while I am very skeptical of "AI"-hype and don't expect tech giants to have anyone's best interest in mind, I do think we did some things well in this project, so keep these three things in mind >
November 13, 2025 at 5:53 PM
Reposted by Vasco Brazão 🏳️‍🌈
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
Reposted by Vasco Brazão 🏳️‍🌈
olivia.science
Love to see community action against this AI nonsense! neighborhoodview.org/2025/11/13/d...
As one Massachusetts school administrator recently said; this moment with AI is remarkably like the moment when we were introduced to asbestos. Yes, it had some remarkably promising characteristics – fireproofing! – and had some real utility in science, research, and industrial applications. But a profit-driven industry bullied us into inserting it everywhere; into our homes and schools and public spaces, before we really understood the risks. This resulted in decades, if not centuries, of illness, injuries, deaths, and the astronomical financial burden of trying to remove the stuff. As you, the leaders and policymakers in our schools, craft an AI policy for our district, we the undersigned call on you to: 1. Ban AI tools into the classroom, protect our students and teachers from de-skilling and allow them the space and time to engage in assignments themselves. 2. Resist any direct financial relationship or contracts with AI providers, as well as the “training” they might offer. 3. Provide a digital literacy curriculum to help students navigate the current digital landscape, and promote critical engagement with technology. 4. Guarantee that anywhere generative AI has already entered our classrooms or curriculum, an opt-out will allow students and teachers to refuse the use of these products at no risk to their grades, progress or employment.
November 14, 2025 at 11:42 AM
Reposted by Vasco Brazão 🏳️‍🌈
jeymccreight.bsky.social
Watson was deeply influential to me as a genomicist, but not how you may assume

From stealing Rosalind Franklin's lab notes to his frequent sexist & racist bioessentialist comments later in life, he showed how the scientific community would still worship a terrible person

Don't idolize bad people
November 7, 2025 at 9:04 PM
Reposted by Vasco Brazão 🏳️‍🌈
homebrewandhacking.bsky.social
Things that make generative AI dangerous:

1. Job market impact. As a cutting measure it's being used to generate "boilerplate". This is putting writers out of work.

2. Stock market bubble. When this inevitably crashes it will take out pension funds and savings.

1/
April 13, 2025 at 12:08 PM
vascobrazao.bsky.social
Will try to gather in this thread some steps I'm taking to put me in a better position to be an effective "AI" critic at my EdTech company. See quoted thread for a bit of context
vascobrazao.bsky.social Vasco Brazão 🏳️‍🌈 @vascobrazao.bsky.social · 3d
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 14, 2025 at 11:17 AM
Reposted by Vasco Brazão 🏳️‍🌈
ozkaterji.bsky.social
Last night was the worst bombing I have experienced in Kyiv since the war began. 7 hours of non stop aerial attacks across the city.
November 14, 2025 at 9:45 AM
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
Reposted by Vasco Brazão 🏳️‍🌈
abeba.bsky.social
flat out lies. these tech ceos take regulators and the masses for fools
“If you want to know how a model will behave, the only way of doing it is to run it,” Mensch also suggested. “You do need to have some empirical testing, what’s happening. Knowing the input data that has been used for training is not going to tell you whether your model is going to behave well in [a healthcare use-case], for instance. You don’t really care about what’s in the training data. You do care about the empirical behaviour of the model. So you don’t need knowledge of the training data. And if you had knowledge of the training data, it wouldn’t even teach you whether the model is going to behave well or not. So this is why I’m saying it’s neither necessary nor sufficient.”
November 12, 2025 at 10:38 AM
Reposted by Vasco Brazão 🏳️‍🌈
aboothagram.bsky.social
Knowledge of the training data has literally been used to understand the cause of errors in AI driven healthcare systems. For example the melanoma detection system that was found to be using the presence of rulers, stickers, and pen marks made by DRs to infer melanoma status, rather than mole itself
November 12, 2025 at 11:52 AM
Reposted by Vasco Brazão 🏳️‍🌈
benjaminjriley.bsky.social
Who's up for a short thread demonstrating just how wrong this claim is by Mistral's CEO, using research by @rtommccoy.bsky.social (an co-authors) regarding crossword puzzles? Just me? Well whatever, let's dive into a mini-exploration of the Embers of Autoregression and why they singe...
abeba.bsky.social Dr Abeba Birhane @abeba.bsky.social · 5d
flat out lies. these tech ceos take regulators and the masses for fools
“If you want to know how a model will behave, the only way of doing it is to run it,” Mensch also suggested. “You do need to have some empirical testing, what’s happening. Knowing the input data that has been used for training is not going to tell you whether your model is going to behave well in [a healthcare use-case], for instance. You don’t really care about what’s in the training data. You do care about the empirical behaviour of the model. So you don’t need knowledge of the training data. And if you had knowledge of the training data, it wouldn’t even teach you whether the model is going to behave well or not. So this is why I’m saying it’s neither necessary nor sufficient.”
November 12, 2025 at 1:16 PM
vascobrazao.bsky.social
therapist said my sadness is not pathological can i get a hell yeah
November 12, 2025 at 11:13 AM
Reposted by Vasco Brazão 🏳️‍🌈
mjcrockett.bsky.social
This NeurIPS workshop claims that LLMs "provide an important foundation for exploring human cognition, emotion, and social interaction"

This is flawed logic, as @lmesseri.bsky.social and I argue here:
www.sciencedirect.com/science/arti...
November 9, 2025 at 3:23 PM
Reposted by Vasco Brazão 🏳️‍🌈
mmitchell.bsky.social
Love the phrases "Cognitive Cost" and "Executive Function Theft" to pinpoint how exhausting it is to be constantly bombarded with apps telling you to use AI. I've been calling it "Corporate pressure" and "Force feeding". It's quite a lot right now. Seems a bit desperate tbh.
hedgielib.bsky.social Abigail Goben @hedgielib.bsky.social · 6d
The #ExecutiveFunctionTheft of having to opt out.
emilymbender.bsky.social Emily M. Bender @emilymbender.bsky.social · 6d
Feeling annoyed at the cognitive cost of having to dismiss all the offers of "AI" assitance every time I use Acrobat to provide feedback on documents my students wrote. (NO I do NOT want an "AI" summary of this. In what world???)

Decided to check settings:
November 10, 2025 at 6:53 PM
Reposted by Vasco Brazão 🏳️‍🌈
andriikatriuk.bsky.social
So last night ordinary good russians completely destroyed the generation of electricity of one of the electricity producing companies. Right before winter. Tell me more how much sanctions and consequences they got for years of terror and genocide.
#UkrainianView
November 8, 2025 at 3:29 PM
Reposted by Vasco Brazão 🏳️‍🌈
escapedexplorer.bsky.social
"World's first potential trillionaire has killed 600,000 people, including 400,000 children" is the only headline that anyone should ever read about Elon Musk
newyorker.com The New Yorker @newyorker.com · 10d
One analytical model shows that, as of November 5th, the dismantling of U.S.A.I.D. has already caused the deaths of 600,000 people, two-thirds of them children. https://newyorkermag.visitlink.me/jUzNSc
The Shutdown of U.S.A.I.D. Has Already Killed Hundreds of Thousands
The short documentary “Rovina’s Choice” tells the story of what goes when aid goes.
newyorkermag.visitlink.me
November 7, 2025 at 1:33 AM
Reposted by Vasco Brazão 🏳️‍🌈
evilbatwitch.cryptoanarchy.network
If you aren't seeing skeets from some of your fave people, they might have been labeled "Rude" by bsky mods. Idk what they are thinking, I don't need a mommy to tell me who to be buddies with. Anyway, go into Settings > moderation > bluesky moderation (advanced) and turn Rude off.
Rude Rude or impolite, including crude language and disrespectful comments, without constructive purpose. Highlighted Off, darkened Warn, and darkened Hide buttons.
November 5, 2025 at 7:10 AM
Reposted by Vasco Brazão 🏳️‍🌈
transmetropolitanr.bsky.social
lol
November 6, 2025 at 4:47 AM
Reposted by Vasco Brazão 🏳️‍🌈
techpolicypress.bsky.social
Elon Musk’s Grokipedia is one of many experiments designed to replace human-generated writing with LLMs. But Ryan McGrady argues that its defining feature is its use of technological power to re-exert top-down authority over information and knowledge.
buff.ly/odvgs4q
November 4, 2025 at 1:30 PM
vascobrazao.bsky.social
oh to be a salmon, lying in an fMRI, mentalizing...
November 4, 2025 at 1:30 PM
Reposted by Vasco Brazão 🏳️‍🌈
solomonkurz.bsky.social
In Ch 19 (nyu-cdsc.github.io/learningr/as...) of his 2nd edition, Kruschke used *residual* SD as a standardizer for group differences from a multilevel ANCOVA. Is there any precedent for using a *residual* SD as a standardizer for a standardized mean difference effect size? #RStats
nyu-cdsc.github.io
October 6, 2025 at 4:06 PM