Russ Poldrack @russpoldrack.org · 07/10/2026"Future research on disease‐modifying treatments for Alzheimer’s disease should focus on other mechanisms of action." I'm sure that will totally happen. www.cochranelibrary.com/cdsr/doi/10.... 163
Russ Poldrack @russpoldrack.org · 24/09/2026ChatGPT Images came up with a really interesting hallucination of a brain, sort of a chimera of axial and coronal slices 2243
Russ Poldrack @russpoldrack.org · 29/08/2026Sid and Nori would like to introduce themselves to all of you. And no, there was absolutely no mischief going on right before this photo was taken by @jenott.bsky.social 1554
Russ Poldrack @russpoldrack.org · 26/07/2026Next time you get excited about how your coding agent solved a really hard problem for you, splash a bit of this cold water on your face - the source (metr.org/blog/2026-05...) should be essential reading for anyone using coding agents. 7282
Russ Poldrack @russpoldrack.org · 21/07/2026Performance optimization (and why you shouldn't optimize prematurely) russpoldrack.substack.com/p/optimizing... - the latest in my Better Code, Better Science series, featuring this great quote from my Stanford colleague Donald Knuth 150
Russ Poldrack @russpoldrack.org · 05/07/2026This is an incredibly insightful article by my Stanford colleague John Ousterhout about all the ways that performance measurement can go wrong in software development. I think it could be particularly useful reading for people chasing benchmarks in AI right now. cacm.acm.org/research/alw... 270
Russ Poldrack @russpoldrack.org · 14/06/2026Congrats to our own @markiewicz.bsky.social for winning the OHBM Open Science Award! 26911
Russ Poldrack @russpoldrack.org · 10/06/2026wow, Fable is basically useless for anything related to neuroscience. e.g., in response to the prompt "What is the neural basis of cognitive control?" it falls back to Opus with the attached message. I appreciate guardrails but this makes the model useless for science. 2223
Russ Poldrack @russpoldrack.org · 01/05/2026People in the AI world seem to adore the System 1/2 dichotomy. I love this quote from Gerd Gigerenzer about the concept: 26111
Russ Poldrack @russpoldrack.org · 27/04/2026Claude sounds a bit like a human with a prefrontal cortex lesion here www.tomshardware.com/tech-industr... 081
Russ Poldrack @russpoldrack.org · 08/03/2026In case you haven't seen the latest METR performance assessment, Claude Opus 4.6 is literally off the charts. Doubling time over the last two years is just over 100 days. metr.org 3120
Russ Poldrack @russpoldrack.org · 04/03/2026Yesterday we said goodbye to our beloved kitty Coco. We miss her terribly but I am glad I was able to immortalize her in Figures 1 and 2 of my 2021 Synthese paper. Hopefully she will meet Fodor's Greycat in kitty heaven. 1502
Russ Poldrack @russpoldrack.org · 24/02/2026To what degree do you think that transition is a feature of DL versus more generally of overparameterized models? e.g. even overparameterized regression models show this behavior - e.g. below from @mdubova.bsky.social & @sabinasloman.bsky.social - seems a bit like magic! 130
Russ Poldrack @russpoldrack.org · 12/02/2026probably the most important lesson (similar to what www.youtube.com/@NateBJones has been saying): 5111
Russ Poldrack @russpoldrack.org · 06/01/2026Wow, I've never seen anything like this from an LLM before (Claude Opus 4.1): 1190
Russ Poldrack @russpoldrack.org · 04/11/2025It's really nice when people actually use software that you create. This is an updated usage figure for the MRIQC.org Web API showing continued usage for the last 7 years, with more than half a million unique BOLD images in BIDS datasets! If you aren't already using MRIQC, check it out! 1261
Russ Poldrack @russpoldrack.org · 25/10/2025Great piece by @natolambert.bsky.social on the current state of human exhaustion in the AI world. Makes this important point: 2268
Russ Poldrack @russpoldrack.org · 24/10/2025Apple's Mail.app AI summaries have officially become useless. 1200
Russ Poldrack @russpoldrack.org · 14/10/2025OpenNeuro @openneuro.bsky.social just hit a huge milestone: 1500 datasets! Congrats to the team on making this project so successful over the last 7 years. 214230
Russ Poldrack @russpoldrack.org · 20/09/2025I know that citation statistics are faulty indicator, but I have to say that there is something humbling in knowing that people have decided 100,000 times to cite a paper that I was involved in. 1024610
Russ Poldrack @russpoldrack.org · 16/09/2025It's interesting how some instructions are really hard for coding agents to follow. Recently I have been using a problem file, in which I include an instruction to only mark a problem as fixed once I have confirmed that the fix worked (see image below). Claude Code basically refuses to do this. 040
Russ Poldrack @russpoldrack.org · 02/09/2025*Sharing for our department’s trainees* 🧠 Looking for insight on applying to PhD programs in psychology? ✨ Apply by Sep 25th to Stanford Psychology's 9th annual Paths to a Psychology PhD info-session/workshop to have all of your questions answered! 📝 Application: tinyurl.com/pathstophd2025 0209
Russ Poldrack @russpoldrack.org · 16/06/2025I ran into a former student at the Stanford graduation today. My stats class clearly had a powerful effect! 0300
Russ Poldrack @russpoldrack.org · 02/06/2025I agree that NHST is not an optimal way to do science, but if you are going to do NHST (which many continue to do), then I do think that well-powered experiments are essential. cf. @talyarkoni.com quote from 15 years ago: 100
Russ Poldrack @russpoldrack.org · 02/06/2025The quality of evidence in psychology publications has improved substantially over the last two decades, across all subareas. journals.sagepub.com/doi/10.1177/... 07517
Russ Poldrack @russpoldrack.org · 31/05/2025Great piece on the need for urgent educational reform to deal with the impact of AI. @tunguz.bsky.social www.aei.org/domestic-pol... 0297
Russ Poldrack @russpoldrack.org · 21/05/2025Very important point raised by @petersalib.bsky.social and Simon Goldstein regarding AI risk and alignment: www.ai-frontiers.org/articles/tod... 052
Russ Poldrack @russpoldrack.org · 17/05/2025Reviewing scientific papers with AI: Part 1 of a series of posts on using AI tools for peer review. russpoldrack.substack.com/p/reviewing-... 3112
Russ Poldrack @russpoldrack.org · 01/03/2025I am very thankful that 1) I grew up in a time when vaccines were universally recognized as the medical miracle that they are, and 2) my mother saved my childhood vaccine card so I have a record 2610
Russ Poldrack @russpoldrack.org · 14/02/2025In the "words no longer mean what they say" category, this from Hulu today: 2210
Russ Poldrack @russpoldrack.org · 14/02/2025A great quote from @ylepidemiologist.bsky.social (which you should definitely be reading!) about what we need to do establish trust for science going forward open.substack.com/pub/yourloca... 0155
Russ Poldrack @russpoldrack.org · 30/01/2025Girl Talk, Austin City Limits Festival, 10/4/2009 Our first ACL after moving Austin, our friend introduced us to Girl Talk, an amazing mashup artist. Absolutely crazy show. This was also the year of the legendary ACL mud fest. 210
Russ Poldrack @russpoldrack.org · 30/01/2025Dead Kennedys, Circle A Ranch (Dallas), December 1985 One of the scariest shows I have ever been to - skinheads were beating people up randomly outside. I would love to have seen them earlier but this was still definitely a bucket list item. But Jello Biafra was in perfect form. 170
Russ Poldrack @russpoldrack.org · 28/01/2025Current guidance from Stanford leadership regarding OMB memo: 37818
Russ Poldrack @russpoldrack.org · 23/01/2025Have had a great speaking trip across Asia, giving talks in Singapore, Bangkok, and now wrapping up in Hong Kong. 0292
Russ Poldrack @russpoldrack.org · 06/01/2025largely agree, though in my talk here I have a good example of the current challenges of fully automated test development without domain knowledge: poldrack.github.io/talks-AIAssi... 110
Russ Poldrack @russpoldrack.org · 01/01/2025Here's a replotting of the decoding results from Table 1 - the "linear" baseline is equivalent to training only on specific data, though using a different model (i.e. a simple linear decoder rather than a transformer + decoder head). These seem compelling to me, though the baseline may not be fair 220
Russ Poldrack @russpoldrack.org · 01/01/2025finally was able to get it pretty much correct with a bit more prompting. 140
Russ Poldrack @russpoldrack.org · 01/01/2025and interestingly, if you ask for the face back, the spelling degrades badly 160
Russ Poldrack @russpoldrack.org · 01/01/2025GPT-4o can fix spelling errors in generated images, but apparently it costs a face to do so. 2172
Russ Poldrack @russpoldrack.org · 19/12/2024I asked Meta AI to "generate an image that depicts the following equation: AI = neural networks + big data + GPU computing. make the image as realistic as possible and avoid any fantastic or sci-fi like imagery". I find the result really humorous for some reason. 2280
Russ Poldrack @russpoldrack.org · 30/11/2024Jen @jenott.bsky.social and I took two rides in Waymo robotaxis last night in San Francisco. The experience was seamless and the driving was impeccable, probably better than what I would expect from the average human driver. It even gunned it to go through yellow light! 6435
Russ Poldrack @russpoldrack.org · 29/11/2024You might try LLM-based code review as well. In Github you can ask the copilot assistant to do a code review. I just tried this: 271