I'm hopeful that new efficiencies in training (Deepseek et al.), the impressive performance of smaller models enhanced through distillation, and a glut of past-their-prime-but-functioning GPUs all converge make good-enough open/libre models cheap, ubiquitous, and less resource-intensive to train and run.
The premise for this research question is related in part to work from Dr. Eleanor Maguire (who I just learned died last year at only 55) and her team on hippocampal anatomy and careers involving spatial navigation, including taxi drivers. A connection that may be of interest to HN readers: Demis Hassabis was one of Maguire's doctoral students, although he does not appear to have worked on the project most relevant to this study.
If you're not interested in the new sets, the core product is readily available. Moreover, enjoy the fact that you can get that bucket of bricks for cheap partly because the expensive shiny high-margin SKUs provide a subsidy.
Tangential, but Darknet Diaries has been running a "hacker history" series to kick off 2026, starting with this one: https://darknetdiaries.com/episode/168/. The "Hackers" movie gets a few call-outs.
Sort of. The rebuttal (by Flandin and Friston) suggests that properly-applied parametric statistics of the kind they favor are valid. Eklund et al. wouldn't disagree with that because their own findings support it, but they would point out that not all researchers necessarily adhered to the conservative statistical approach that F&F discuss. More specifically, both sets of authors describe the importance of using a conservative "cluster defining threshold" to identify spatially contiguous 3D blobs of brain activation. Eklund et al. use their findings to raise the question of whether the bulk of fMRI reports were conservative in this regard.
I completely agree that this can be frustrating. For what it's worth, the data from a research study aren't necessarily going to be clinically informative to a GP or even to a radiologist. Structural MRI data could potentially be interpreted, but the sequences collected for research are still different than what you'd get if you went in for a clinical scan.
And if you're interested in having the data for your own purposes, there may still be some liability concerns for the researchers. There's always an agreement to keep MRI data under strict supervision to avoid leaks of personal health information, so providing those data to participants (safeguards or no) could potentially count as a breach of protocol.
Flaundin & Friston's response is interesting because it essentially endorses the findings of Eklund et al. (except for one element of the Eklund analysis that they suggest is a modest error). F&F believe that by setting one parameter correctly (i.e., using a conservative cluster forming threshold) the validity of their preferred parametric statistical approach is upheld. Eklund et al. might quibble because their take-home message is that non-parametric methods should be used instead, but their findings are not misrepresented by F&F.
Regardless, an open and important question is how often other authors used a sufficiently conservative cluster forming threshold for their fMRI analyses. If nothing else, Eklund et al. will cause future reports to be more cautious in this regard.
The dead salmon article is a bit of a red herring here. It's a clear demonstration that a shoddy statistical approach can undermine fMRI findings. Critically, the implications of the current paper extend to even research that has been rigorously analyzed using field-standard software. Statistical issues are at the heart of both papers, but the newer paper identifies problems that are subtle and ubiquitous.
This is right and wrong: AFNI's 3dClustStim did contain a specific bug (since corrected) and is used for fMRI analysis, but the problem that the authors identified is more general than that and encompasses all of the tested packages.
You're correct that the main finding reflects flawed statistical assumptions made by the major neuroimaging packages (or so these authors contend), but they did uncover a specific bug in one of the three packages (AFNI):
"... a 15-year-old bug was found in 3dClustSim while testing the three software packages (the bug was fixed by the AFNI group as of May 2015, during preparation of this manuscript). The bug essentially reduced the size of the image searched for clusters, underestimating the severity of the multiplicity correction and overestimating significance ..."