Will Harrison @willjharrison.bsky.social · 6hI suspect this is what the companies were doing back in the 2010s, but the training datasets weren't sufficient to produce the appearance of "expertise" in all domains. 100
Will Harrison @willjharrison.bsky.social · 6hI can't see those results reported though? Am I missing something? 000
Will Harrison @willjharrison.bsky.social · 04/10/2026AI companies are waging and winning a war to absolve themselves of liability for their shitty products. The real danger has always been how AI tools will help Bad Humans do Bad Things, and that's exactly what is happening when an AI agent goes "rogue". 020
Will Harrison @willjharrison.bsky.social · 04/10/2026bro wants you to calmly chew your asparagus so strangers can tell you why you shouldn't exist He often gives seemingly good advice, but this is dog shit privilege. 100
Will Harrison @willjharrison.bsky.social · 04/10/2026This is critical and missed by many researchers. We don't necessarily understand something just because a DNN or LMM can reproduce him-like responses; that just means that the responses can be learned from a dataset/architecture. A psychologist/neuroscientist's target is the mind, not the model. 040
Will Harrison @willjharrison.bsky.social · 04/10/2026Can you comment on the team's knowledge contribution to agent development? It's easy to calculate direct costs, like subscription fees and time spent playing with a model, but do you also need to factor in your team's expertise? 110
Will Harrison @willjharrison.bsky.social · 04/10/2026Can you share more information about the "new" untrained tasks you tried out? (I've been arguing with my uni that online testing is not good enough quality for my research, so this is also a really critical study for uni policy) 010
Will Harrison @willjharrison.bsky.social · 04/10/2026This is really fascinating and potentially troubling. I'm curious if you tried any perceptual tasks? From reading the preprint and your thread, some important details are unclear about what information was available to the boys. I'm not skeptical of any of your claims in general. 1/n 200
Will Harrison @willjharrison.bsky.social · 02/10/2026This looks like a positive change, but the devil will be in the details: how many “highly competitive” allocations can be made, and what is the expected number of funded grants based on review vs lottery? 120
Reposted by Will HarrisonBruce Wilson @brucewilson.bsky.social · 30/09/20261,200 reported incidents of sexual assault, dating & domestic violence, and stalking lead to 7 expulsions or dismissals. Cornell has positioned itself as higher ed's industry leader in teaching rape culture to America's future male economic elites. www.usatoday.com/story/news/n...usatoday.comCornell rarely expels students in sex assault cases | ExclusiveA third of undergraduate women say they've been sexually assaulted. But on average, Cornell reported expelling just one attacker per year. 56525
Will Harrison @willjharrison.bsky.social · 01/10/2026I'm avoiding it and actively encouraging my friends and family to leave. 020
Reposted by Will HarrisonAlex Holcombe @alexh.bsky.social · 21/09/2026With the recent mainstreaming of AI concerns, I'm very disappointed that we haven't heard more about supporting proof of personhood tools and techniques, which some have worked on for more than a decade. 182
Will Harrison @willjharrison.bsky.social · 17/09/2026We just decided to trial Discord in our lab too 100
Reposted by Will HarrisonJ. Brendan Ritchie @jbrendanritchie.bsky.social · 14/09/2026If only there was a thing govs could do to make companies align with human values. Some kind of system of legally binding constraints for certain industries. Shame nothing like that has ever existed ever. 1145
Will Harrison @willjharrison.bsky.social · 13/09/2026isnt the joke on us, the people who use these tools for IRB-approved studies? 110
Will Harrison @willjharrison.bsky.social · 11/09/2026We let AI companies soak up all the IP in the world but dammit if we let them take our maths. If this is the straw that achieves change then so be it. 011
Reposted by Will HarrisonIcelandic Vision Lab @icevislab.bsky.social · 10/04/2026VIVAS correlates only moderately with the VVIQ, well below the ceiling set by their respective reliabilities, and shows striking individual-level dissociations: some individuals scoring in the bottom 10% on one measure fall in the top quartile on the other. 7/9 111
Will Harrison @willjharrison.bsky.social · 10/09/2026Why won't anyone think of where the codes are 000
Will Harrison @willjharrison.bsky.social · 10/09/2026I don't think I know what the binding problem is anymore. As an aside, I'd like to know if demo #1 still works when peripheral stimuli are scaled for cortical magnification. I suspect the conflicting motion signals are being pooled in larger RFs, and just adding noise. 120
Will Harrison @willjharrison.bsky.social · 10/09/2026Great letter and good on you for bothering to write, send and share. 000
Will Harrison @willjharrison.bsky.social · 09/09/2026Before everyone was talking about prediction errors, surprise was investigated in the context of attentional control settings, e.g. link.springer.com/article/10.3...link.springer.comOculomotor Capture by New and Unannounced Color Singletons during Visual Search - Attention, Perception, & PsychophysicsThe surprise capture hypothesis states that a stimulus will capture attention to the extent that it is preattentively available and deviates from task-expectancies. Interestingly, it has been noted by... 010
Will Harrison @willjharrison.bsky.social · 09/09/2026When I hear an AI model solved an old maths problem, I ask myself: 1) How many human mathematicians were involved? 2) How much money and freedom were the mathematicians paid to solve the problem? My null hypothesis is that the multi-billion dollar models are not the critical ingredient. 031
Reposted by Will HarrisonEric Schares @eschares.bsky.social · 31/08/2026🚨Update: Our analysis of global APC expenditure has now been expanded to cover more publishers (7 -> 14) and more years (2019-2025). We estimate $15B in APCs paid over seven years, $3.7B in 2025 alone. Elsevier crosses the $1B threshold by itself in 2025. arxiv.org/abs/2608.16322 #ScholComm 6179148
Will Harrison @willjharrison.bsky.social · 31/08/2026Yeh but the water temp is still only 20° and I'm such a baby I need it to be 22 before I go in 😭 010
Will Harrison @willjharrison.bsky.social · 30/08/2026This reads a bit like, "Now that we've built such huge complex models that aren't really interpretable, we need to start using the tools of science to understand them." 160
Will Harrison @willjharrison.bsky.social · 30/08/2026We really breezed right through winter this year didn't we. So uncomfortably warm for this time of year already. 110
Will Harrison @willjharrison.bsky.social · 30/08/2026Thanks for sharing - what a cool graphic. As someone who straddles behavioural psych and cog neuro/computer vision, I think the fracturing in neural methods will be lesser that traditional psych because of the better development and sharing of analytical toolboxes in neuro-areas. 010
Will Harrison @willjharrison.bsky.social · 28/08/2026@neuronomist.bsky.social will probably know more about this, or maybe can suggest some papers? 010
Will Harrison @willjharrison.bsky.social · 28/08/2026Sure, but your question involves attentional sets, which I think are more contextual than trial by trial. An observer doesn't know from one trial to the next what the orientations will be or the best saccade vector. They DO still know the relevance of colour, which is a more reliable cue. 110
Reposted by Will HarrisonIgor Utochkin @utochkin.bsky.social · 27/08/2026Our new preprint w/ @edvogel.bsky.social presents a neural marker of a competition/interaction taking place in working memory between memory and current goals. We label it r-CDA Check out preprint for details about r-CDA and other interesting neural dynamics osf.io/preprints/ps... 12110
Will Harrison @willjharrison.bsky.social · 28/08/2026My immediate thought is that orientation/shape is distorted across eye movements, but colours aren't. It's hard to tune to "horizontal" while making horizontal eye movements, because everything turns horizontal during the saccade, and static horizontal features get masked. 100
Will Harrison @willjharrison.bsky.social · 27/08/2026Good and useful summary, thanks. I skimmed through the more critical replies and thought they make assumptions about your motivation in preparing this, which I found confusing. 010
Will Harrison @willjharrison.bsky.social · 26/08/2026Would I have the power to stop the year-to-year arbitrary formatting changes for grant apps? #hahahgivememoremoney 020
Will Harrison @willjharrison.bsky.social · 25/08/2026Skeptical of some /AI/ math proofs banging around public sphere. How do we know what's in the training data? How much was the human doing? But honestly I don't care enough to debate. 100
Will Harrison @willjharrison.bsky.social · 24/08/2026By finding the local peaks and troughs of the filtered noise, it's easy to add specular highlights to give the appearance of gloss. 001
Will Harrison @willjharrison.bsky.social · 23/08/2026Sometimes I just want to try to code pretty shit low pass filter 3D noise (creates the undulation) -> project noise onto a colour wheel (with some other bells and whistles) 1124
Will Harrison @willjharrison.bsky.social · 20/08/2026It’s very difficult to assess what tools help learning, versus appear to help learning. 160
Will Harrison @willjharrison.bsky.social · 20/08/2026Hey Blake do you do much undergrad teaching? My feeling is my first year students would need 2 semesters of dedicated study before understanding your point. 000
Will Harrison @willjharrison.bsky.social · 20/08/2026great to see some places still have a backbone 071
Will Harrison @willjharrison.bsky.social · 17/08/2026I also left around the same time. We had something special on Twitter, pre-Musk, but now I reckon there's no excuse for publicly funded scientists to tweet or even peruse twitter. 070
Will Harrison @willjharrison.bsky.social · 13/08/2026Just putting this idea out there: if a theory of cog neuro requires screening out left handed people from experiments, maybe the theory isn't really a theory, yet. (And cool work from this team!) 0110
Will Harrison @willjharrison.bsky.social · 11/08/2026An adaptive algorithms that lets you shape your stimulus in any of 50 dimensions according to an observer's response.... and it works almost in realtime so you can use it in experiments on the fly. 0171