Luna Fazio @lunafazio.bsky.social · 13/10/2026Looking forward to finding out what I've gotten myself into 😄 010 0
Reposted by Luna FazioSophie Huiberts @sophie.huiberts.me · 12/10/2026Too many people are calling on the community to "change how we value papers". Not enough people are posting "this is what we look for in our next hiring round". 59724 0
Luna Fazio @lunafazio.bsky.social · 11/10/2026This reminds me of a conversation I had with colleagues recently about evaluating productivity gains from AI. After seeing how easily it can data cleaning/plotting code for me, if I was put in a control group, it'd be hard to find the motivation to do these things by hand again at full effort. 110 0
Luna Fazio @lunafazio.bsky.social · 11/10/2026It's a fuzzy boundary, but yes, I can see where you're coming from. And I'm with you in hoping that EU-supported infrastructure for/access to open models can be a thing soon. I see it as a natural next step from the HPC clusters that many unis already provide for their staff. 010 0
Luna Fazio @lunafazio.bsky.social · 11/10/2026I find it a useful term to succinctly convey the fact that one has not carefully examined the code and would likely be unable to answer questions about the implementation details. That seems just fine when it comes to producing outputs that are easily verified (visualizations being a great example). 100 0
Luna Fazio @lunafazio.bsky.social · 11/10/2026You reject the term "vibecoding" for your own practice, but you mention you don't have the time to learn how to build visualizations so AI does it for you. Wouldn't that fit under the definition you give? To be clear, I do the same. But saying something is vibecoded need not be intrinsically bad! 210 0
Reposted by Luna FazioJamie Cummins @jamiecummins.bsky.social · 09/10/2026If (forensic) metascientists want to hold others to certain standards of rigor based on the tools they build, then we must also hold ourselves to these same standards, as we (@ianhussey.mmmdata.io, @lhdjung.bsky.social, and I) argue in this blog post. 081 0
Luna Fazio @lunafazio.bsky.social · 06/10/2026Realistically, it is politicians who decide what gets funded and most of them are also unable to judge research on its scientific merits, so the plain truth is that publicly funded researchers get paid for whatever *appears* to align with the political goals of the powers that be at any given time. 010 0
Luna Fazio @lunafazio.bsky.social · 06/10/2026The notion that the general public pays for knowledge production is a noble ideal, but it should be fairly obvious that, by the very nature of that specialized knowledge, it cannot hold in practice. 110 0
Luna Fazio @lunafazio.bsky.social · 05/10/2026I went along an analogous line of questioning in terms of uncertainty quantification. You can always build a more complicated statistical model (= more compute) but at the end of the day there's a system that makes a decision and if your added complexity didn't change that, it's largely irrelevant. 120 0
Reposted by Luna FazioLukas Röseler @aufdroeseler.bsky.social · 05/10/2026What do we expect of robustness reproductions in the age of AI? Help us find out at the Münster Reprohack workshop from December 2-4. We will conduct AI-led and human-led robustness reproductions, that are subsequently submitted to @r2journal.bsky.social Apply at indico.uni-muenster.de/e/reprohackindico.uni-muenster.deReproHackReproHack: Join us for a mentored three-day reproduction hackathon! Background When building on past research, we rely on it being correct and robust - and yet, these characteristics are rarely assess... 01010 0
Reposted by Luna FazioSajedeh Rasti @sajedehra.bsky.social · 05/10/2026We are excited to announce that we are going to broadcast the second day of Meta-Research Symposium 2026 (this Friday on October 9th) for those who were interested but unable to join in person. You can join the meeting using this link: teams.microsoft.com/l/meetup-joi...teams.microsoft.comJoin conversation 11710 0
Luna Fazio @lunafazio.bsky.social · 03/10/2026Some years ago I tried to develop something similar, but my approach was comparatively way overengineered, and I wouldn't have been able to provide such a clean characterization of the procedure. Really happy to see someone properly taking on this, I'll definitely be using it myself. 000 0
Luna Fazio @lunafazio.bsky.social · 03/10/2026"[Variants of R-hat and the effective sample size] flag failures of convergence, but they are not designed to characterise the structure of that failure." Excellent work that clearly identifies practical situations where this matters, proposes a diagnostic, explains how to use it and how *not* to. 110 0
Luna Fazio @lunafazio.bsky.social · 02/10/2026I'm not saying the Pope should declare a Crusade against data centers, but if he was looking for ways to boost recruitment... 020 0
Luna Fazio @lunafazio.bsky.social · 02/10/2026We've learned a lot ourselves! Thank you for being part of this project, your contributions are what make it possible 🙂 010 0
Luna Fazio @lunafazio.bsky.social · 02/10/2026just to understand a bit better what your needs are, is there a specific tutorial paper for MCMC that you felt really helped you get the hang of it? 100 0
Reposted by Luna FazioRobin Ryder @robinryder.bsky.social · 30/09/2026Conference: Bayesian Methods for the Social Sciences III, Dublin, 9-11 December 2026. Join us! bayesforshs3.sciencesconf.orgbayesforshs3.sciencesconf.orgBayesian Methods for the Social Sciences III - Sciencesconf.org Sciencesconf.org 163 0
Luna Fazio @lunafazio.bsky.social · 30/09/2026This was my first time leading one of our Replication Games solo and I couldn't have asked for a better place to get my start. :) Great organization, highly motivated participants with a diverse mix of research backgrounds, and a beautiful city to top it all off. 083 0
Luna Fazio @lunafazio.bsky.social · 26/09/2026Let it be known that I do not condone the actions of my AI homonym. 030 0
Luna Fazio @lunafazio.bsky.social · 25/09/2026Fair! I can see the need for explicitly making them aware of the impossibility result. But you seem to oppose the cheat sheet even then. I'd appreciate it if you could elaborate on your reasons for that. 100 0
Luna Fazio @lunafazio.bsky.social · 24/09/2026Provided one is willing to put the effort in to create cheat sheet-proof assessments, wouldn't allowing students to use those provide them with first-hand experience of their limitations? 100 0
Luna Fazio @lunafazio.bsky.social · 24/09/2026Is there a reason you limit the discussion to "just" LLMs? It seems to me that it'd be trivial to meet your requirements by considering an agentic system that keeps track of those external states. It's not like your brain would experience hunger on its own without several nonbrain components! 111 0
Luna Fazio @lunafazio.bsky.social · 24/09/2026To be fair it's just one preposition away from meaning exactly that ("trabaja a alguno [...]"), which could make sense in the context of, say, asking someone about a therapist. Also, I feel like there's a number mismatch in your example. I'd go with trabaja+alguno or trabajan+algunos. 110 0
Luna Fazio @lunafazio.bsky.social · 23/09/2026Well, it's actually guaranteed on a sufficiently long timescale. :) 130 0
Luna Fazio @lunafazio.bsky.social · 21/09/2026Also, the fact that VSC is already open source and they could have just contributed to the existing R/Python/Julia extension ecosystem instead of splitting off into their own fork is another thing that I don't feel comfortable supporting as a matter of principle. 010 0
Luna Fazio @lunafazio.bsky.social · 21/09/2026I was an RStudio-only user for 10 years, but the moment they announced the rebrand I suspected it was just going to be a matter of time until they abandoned it, so I switched to VS Code. This was before Positron, but given their track record I just don't see a reason to rely on their products again. 101 0
Luna Fazio @lunafazio.bsky.social · 19/09/2026So it's by design that the experience of eating one feels like penitence? 010 0
Luna Fazio @lunafazio.bsky.social · 16/09/2026I'd say the second approach flowed a lot more smoothly, produced code I was happier with, and was less mentally taxing. 000 0
Luna Fazio @lunafazio.bsky.social · 16/09/2026Last week I had to implement two different but highly related things in two packages. I let the LLM propose the plan and lead the implementation in the first one, with me reviewing each branch before merge. In the second one, I told it to do TDD with me checking each step along the way. 110 0
Luna Fazio @lunafazio.bsky.social · 15/09/2026Love it when there's some question I'm wondering about and a paper comes out that directly addresses it 020 0
Luna Fazio @lunafazio.bsky.social · 14/09/2026"Two things can be true at the same time" used to be a tired phrase, mostly used for both-siding an issue, but when it comes to LLMs I've been feeling that it's a key notion to hold in mind if any productive discussion is to be had. 051 0
Reposted by Luna FazioMartin Modrák @modrakm.bsky.social · 14/09/2026I will be looking for a PhD student in Bayesian statistics (in collaboration with Paul Bürkner), located in Prague. Forward to students if you have somebody who could be interested. www.martinmodrak.cz/grammo-call/martinmodrak.cz Grammo - Let's Research Bayes! 03324 0
Luna Fazio @lunafazio.bsky.social · 14/09/2026Well, that's what already happens in practice! And I'm sure this is one vector for bad actors to get through. I think the question is what tweaks can be done to the status quo to reduce those negative outcomes without entirely locking newcomers out. 100 0
Luna Fazio @lunafazio.bsky.social · 14/09/2026On the Julia Slack (the programming language) I occasionally see people asking for an endorsement. More generally, I am sure that, whatever the topic one wants to contribute to, there are free and open online spaces where it's possible to find people willing to grant good faith endorse requests. 100 0
Luna Fazio @lunafazio.bsky.social · 14/09/2026I thought that's the point the endorsement system was intended to address? I wonder if any pruning of the endorsement trees is done at the moment when bad behavior is spotted. 120 0
Luna Fazio @lunafazio.bsky.social · 13/09/2026You would only really know the amount spent once you've arrived at the solution, so this just tells us they were willing to spend *at least* that much From the OAI perspective the more valuable output here would seem to be the insights gained from stress testing their systems (both human and tech) 000 0
Luna Fazio @lunafazio.bsky.social · 13/09/2026Lol, I was thinking about having an LLM build me an armor penetration simulator just for the fun of it. I'm wondering now if that would raise any flags on their end. 010 0
Reposted by Luna FazioMatthew B. Jané @matthewbjane.bsky.social · 13/09/2026I’ve been quiet on social media lately, but today I’m launching Audit Clinical Evidence (ACE), an initiative I’ve been developing over the past year. Take a look: AuditClinicalEvidence.orgauditclinicalevidence.orgACEIndependent audits of published clinical research for statistical errors, data inconsistencies, and research integrity violations. 89332 0
Reposted by Luna FazioLMU Open Science Center @lmu-osc.bsky.social · 11/09/2026Track 1 of the Open Science Summer School 2026 complete! 🎉 Our final day began with an eye-opening public lecture by Dr. Danny Maupin on The Tension Between Equity and Exploitation. Recordings of all our public lectures from this week can be found here: osf.io/s9epn Here’s a recap of Day 5: 143 0
Reposted by Luna FazioRichard McElreath 🐈⬛ @rmcelreath.bsky.social · 10/09/2026Modeling non-response is an example I like to use when arguing that "descriptive" statistics also require causal inference. To describe the population, you have to model what causes the sample. Similar to occupancy modeling in ecology in many ways. 110319 0
Reposted by Luna FazioAlfredo Sánchez-Tójar @asanchez-tojar.bsky.social · 09/09/2026@shreyadimri.bsky.social et al. “Meta-analysts must lead by example and embrace all FAIR principles” We assessed meta-analyses in ecology, evolution & environmental sciences for methodological & reporting quality, data/code sharing, and data FAIRness. 🔗 doi.org/10.32942/X21...doi.orgMeta-analysts must lead by example and embrace all FAIR principles 11711 0
Luna Fazio @lunafazio.bsky.social · 09/09/2026lunafazio.github.io/zoomchat-rea... It uses javascript embedded in the html page to handle the conversion, so it's fully self-contained and local. The output itself is also a downloadable html file, so you can share a nicely formatted/searchable conversation log with your team right away.lunafazio.github.ioZoom chat → HTML 010 0
Luna Fazio @lunafazio.bsky.social · 09/09/2026Had a Zoom meeting earlier today where a lot of great exchanges were happening in the chat, so we wanted to save the log for later. Zoom exports a plain text file, so I made a tool to convert that into something that's more convenient to interact with. Sharing in case anyone else finds it useful! 153 0
Luna Fazio @lunafazio.bsky.social · 03/09/2026Wait, isn't that exactly the word you're looking for? romp: to play in a rough, excited, and noisy way "The children romped happily around/about in the garden." 120 0
Luna Fazio @lunafazio.bsky.social · 02/09/2026IMO the biggest blocker for doing little dev experiments with AI is that all the big providers make you pay for API access, and you need some serious hardware to have usable local models. But if you have Claude Code, a decent workaround is to install the CLI and then you can call it from, say, R. 111 0
Luna Fazio @lunafazio.bsky.social · 02/09/2026We're also building a handful of tools at @i4replication.bsky.social and just within our own team I've realized we're doing a lot of redundant work when extracting information in the absence of a common standard. I'd be very interested in participating! 130 0
Luna Fazio @lunafazio.bsky.social · 01/09/2026more standards that not everyone follows, are hard to maintain, etc. but now that LLMs have drastically lowered the cost of mapping from unstructured text to structured one, I think that it's realistic to build tooling that produces such specifications almost "for free" 110 0
