iiinova — where your digital self keeps evolving

Reposted by Andrey Chetverikov

John Clithero @johnclithero.bsky.social · 12/10/2026

Congrats to Mike Woodford on winning the Nobel Prize! Some of his recent work is also quite exciting, working on how "cognitive noise" and "cognitive imprecision" play a foundational role in decision making: blogs.cuit.columbia.edu/mw2230/recen...

blogs.cuit.columbia.edu

Models of Imprecise Cognition – Michael Woodford

0145 0

Andrey Chetverikov @achetverikov.bsky.social · 12/10/2026

In all seriousness though, Woodford's more recent work on the relationship between decision-making in perception and in value-based decisions is something that I often mention in lectures. Like this paper, for example: www.annualreviews.org/content/jour...

annualreviews.org

Modeling Imprecision in Perception, Valuation, and Choice

Traditional decision theory assumes that people respond to the exact features of the options available to them, but observed behavior seems much less precise. This review considers ways of introducing...

030 0

Andrey Chetverikov @achetverikov.bsky.social · 12/10/2026

I'm very happy about today's Nobel Prize because it gave me an opportunity to casually mention to students, while discussing a paper they found, that "this guy got a Nobel prize today, but we still won't include him in our Excel sheet"(sorry Michael!).

120 0

Andrey Chetverikov @achetverikov.bsky.social · 11/10/2026

My circhelp package has been updated on CRAN. cran.r-project.org/web/packages... There have been multiple small fixes since the last update, plus the "density asymmetry" analyses - a potential replacement for mean analyses for biases like serial dependence. achetverikov.github.io/circhelp/art...

cran.r-project.org

circhelp: Circular Analyses Helper Functions

Light-weight functions for computing descriptive statistics in different circular spaces (e.g., 2pi, 180, or 360 degrees), to handle angle-dependent biases, pad circular data, and more. Specifically a...

020 0

Andrey Chetverikov @achetverikov.bsky.social · 09/10/2026

thanks for updating! anyway, if you want to check out the studies we used for cross-validation, I can share the links with you over email then

010 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

An unexpected consequence of that is that now I use bots to do quality tests for all new or modified study. No more cropped stimuli or miscalculated trial numbers! Plus gamifying a bit is also easier now, but that's more on the dev side, still have to get a grip on that.

120 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

I think this wouldn't be a remedy as long as these conditions are rendered in a way that a machine can easily interpret (or if there is enough time to make and parse screenshots).

010 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

I wanted to say, no it's only DOM/Canvas/source. But. I checked the logs: for two tasks they did pixel analyses. Simple ones, like template matching but still. For another task, they tried but screenshots did not work fast enough to capture a brief stimulus. Sources were always read though.

010 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

I'll send you some links in DM. Generally, bots did not have any info on the task. I can't speak for @talha-ozudogru.bsky.social's AGEA, but my off-the-shelfs tests (Clastra) had general guides written by LLM before seeing any task. A new agent then went through each task aiming for one-shot pass.

200 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

Sorry, was busy last couple of days, missed many questions! Yes, we did a try perceptual tasks, they are in the supplement. WebGL (like on pavlovia or with jspsychophysics), canvas, does not matter, they do them well.

010 0

Andrey Chetverikov @achetverikov.bsky.social · 08/10/2026

No, we didn't. And to be honest, we did not try to make them superhard to detect or see if they could bypass strong security measures (e.g., captcha or keystroke/cursor movements). The preprint "only" shows that bots can do any task with little effort. This (to me) was still very unexpected!

010 0

Reposted by Andrey Chetverikov

Richard Huskey @richardhuskey.bsky.social · 02/10/2026

🧵In April, I built an AI agent that could complete the Attention Network Task. The agent was task-specific and costly. I concluded that online behavioral experiments were safe from agentic infiltration. Six months later, my view has changed. Our new preprint explains why osf.io/preprints/ps...

osf.io

Generalizable experiment agents threaten online behavioral experiments

Online behavioral experiments rely on the assumption that participants are human. Advances in large language models (LLMs) challenge this assumption, but previous demonstrations have largely focused o...

2155 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Thanks a lot to @talha-ozudogru.bsky.social @cstrauch.bsky.social, @richardhuskey.bsky.social and Leendert Van Maanen for this collaboration. Personally, the the rapid progress of the bots, is astonishing. It doesn't mean for me "back to the lab" but something to consider for new studies. n/n

240 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

So where are we now? Bots pass most tasks, often in one shot, and are easy to develop. We don't know if they are present in the data. They don't have human performance, but behave human-like, making them hard to discard. So I believe that at least detection measures are needed. 9/n

192 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

When Clastra tried one task that @talha-ozudogru.bsky.social developed, with simple bot traps such as invisible text and AI instructions, it ignored them, responding like a human would. So basic countermeasures won't work (we haven't explored more advanced ones, and that's an arms race). 8/n

160 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Also, neither Clastra nor AGEA tried to really hide that they are bots. Both were told to identify as such. But this is an ethical decision, supported by model alignment, not something that a potential bot-running participant would do. 7/n

160 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Then I asked ChatGPT Astra to spawn agents to follow these instructions. One agent per task, naive. And it did do so, quite remarkably, with fewer attempts than AGEA. This isn't a really fair comparison, as I used a pricey Pro subscription, while AGEA relied on free ones to run the bots. 6/n

140 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Another twist: during the cross-validation, I also tested out-of-the-box solutions, which is Clastra in the plots. OpenAI Claude made instructions for agents to pass the tasks as a form of QA testing. The instructions didn't say anything about the tasks' specifics, just a general pipeline ... 5/n

140 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Here's the figure from the supplement with different tasks used for "cross-validation" of the bot. You can see that it often under- or overestimates humans. So sometimes it can be filtered out as a very poor subject, but would you automatically discard a very good one? 4/n

181 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

And AGEA did manage to go through them! That doesn't mean it performed as humans, but it would be hard to detect/disqualify it as non-human (posthoc, having a bot and a non-bot sample, for a given bot you likely can, but probably not for a variety of bots with differing "human models"). 3/n

140 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

@talha-ozudogru.bsky.social developed AGEA, an agent that passes a wide range of cognitive tasks. To ensure the tasks weren't too easy, I also put it through some of my actual studies it hadn't seen before. They weren't the textbook versions, but research-grade ones, with some twists and quirks. 2/n

140 0

Andrey Chetverikov @achetverikov.bsky.social · 02/10/2026

Some months ago, I wrote a skeptical reply to a commentary on bots in online studies in PNAS, arguing that no evidence yet exists for such bots. Together with the authors of the original commentary, we now show that bots can pass (almost?) all online cognitive studies with human-like performance 1/n

25129 0

Reposted by Andrey Chetverikov

Christopher DiPiazza @chrisdipiazza.bsky.social · 25/09/2026

Drexel University is choosing to shut down the country’s oldest natural history museum, The Academy of Natural Sciences, for “funding” reasons (even though it costs .6% of their annual budget and they spent WAY more on ai nobody wants this year!) Go to save-the-academy.org to help.

315696 0

Andrey Chetverikov @achetverikov.bsky.social · 25/09/2026

This is a regular editorial PSA: please do not ignore "Suggest reviewers" field when submitting, or add them to the cover letter if there is no special field. Also try to include not only the top names - they receive tons of invitations. Please also inform your PhD students about it.

111 0

Andrey Chetverikov @achetverikov.bsky.social · 23/09/2026

Ah, then it probably just parsed the source. Still might impressive. Have you tried source obfuscation? Apparently, you can also just vibecode the obfuscation algorithm implementation that would be quite strong against the same LLM www.elastic.co/security-lab...

elastic.co

The Cost of Understanding: LLM-Driven Reverse Engineering vs Iterative LLM Obfuscation

Elastic Security Labs explores the ongoing arms race between LLM-driven reverse engineering and obfuscation.

010 0

Andrey Chetverikov @achetverikov.bsky.social · 22/09/2026

Is it one-shot or are many trial and error attempts needed? Because if it is the latter, just block after 3 reloads.

100 0

Andrey Chetverikov @achetverikov.bsky.social · 21/09/2026

I think this also needs to be seen in context of decreasing national funding. But yes, AI likely decreases the application threshold significantly. Is it the core issue? Probably not, award rates were ridiculous even before.

010 0

Andrey Chetverikov @achetverikov.bsky.social · 21/09/2026

I made an interactive demo with different visual illusions to replace the old videos we used before. Feel free to play with it here illusions.andreychetverikov.org and to use it yourself (the code is on my Github; if I missed to include somebody in the credits - please let me know).

illusions.andreychetverikov.org

Visual Illusions Lab

Eleven live, adjustable visual-illusion demonstrations for outreach events.

0288 0

Andrey Chetverikov @achetverikov.bsky.social · 16/09/2026

I got it already! Impressive work, it's very nice to have a theoretical summary of EC. I also liked the controversies section, especially the flexibility note. Indeed, if flexibility is taken into account, many of the impressive fitting results become a bit less impressive.

110 0

Reposted by Andrey Chetverikov

Ivan Tomic @ivntmc.bsky.social · 10/06/2026

Feeling really proud of this one: a review of Bayesian efficient coding with Máté Lengyel and Paul Bays (@bayslab.org) is now available online in TICS! authors.elsevier.com/a/1nF3i4sIRv...

Screenshot of a Trends in Cognitive Sciences article header. Title: 'Bayesian efficient coding as a theory of perception: progress, controversies, and prospects.' Authors: Ivan Tomić, Máté Lengyel, Paul Bays. Available online 9 June 2026, In Press, Corrected Proof.

310937 0

Andrey Chetverikov @achetverikov.bsky.social · 16/09/2026

I think your authors link has became too popular and no longer works =)

110 0

Reposted by Andrey Chetverikov

Luke Sjulson @lukesjulson.bsky.social · 13/12/2024

Neuroscience students asked us to teach a PRACTICAL course on experimental methods, and it is now on YouTube! Please like and repost to help us get the word out! www.youtube.com/playlist?lis... Lecture 1: Signals and data acquisition Focusing on hardware, digital/analog I/O, synchronization 🧵

youtube.com

Neuroscience methods - YouTube

Nanocourse: Approaches to Study Neural Circuits This course was taught by Anita Autry, Tiago Gonçalves, and Luke Sjulson at Albert Einstein College of Medici...

19437215 0

Andrey Chetverikov @achetverikov.bsky.social · 15/09/2026

Nice! Reminded me of Levine work on reinforcement learning psycnet.apa.org/record/1971-...

040 0

Reposted by Andrey Chetverikov

Julia M. Rohrer @dingdingpeng.the100.ci · 11/09/2026

Another thought: Ruben is *not* a sloppy guy but rather paranoid about making mistakes. The models are still able to catch at least some legit issues. Now imagine if we were serious about the scientific record being accurate and implemented such checks before stuff gets published.

2121 0

Reposted by Andrey Chetverikov

Arthur Prat @arthurpr4t.bsky.social · 09/09/2026

We derived six laws of psychophysics from a single efficient-coding equation. The laws include Weber's law; scaling laws in visual working memory (which had been noted before, but with unclear theoretical justification); Wei & Stocker's law of human perception; arthurprat.com/pdfs/Prat-Ca...

arthurprat.com

49728 0

Reposted by Andrey Chetverikov

Icelandic Vision Lab @icevislab.bsky.social · 10/09/2026

🚨Paper alert🚨 Here we try to tackle the tricky problem of measuring the inherently private experiences of visual imagery. Now out and open access in Cortex: doi.org/10.1016/j.co... @heidasigurdar.bsky.social @arnig.bsky.social #imagery #visualimagery #visionscience #psychology #cogsci #aphantasia

doi.org

Redirecting

2198 0

Andrey Chetverikov @achetverikov.bsky.social · 02/09/2026

But isn't the end price different in that case? Deals often pay fixed price (right?) regardless of the numbers of papers published. And they are advertised as a way to cut costs. It would be great to get an uncertainty band on these graphs.

100 0

Reposted by Andrey Chetverikov

Igor Utochkin @utochkin.bsky.social · 27/08/2026

Our 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 0

Andrey Chetverikov @achetverikov.bsky.social · 23/08/2026

Happy #ECVP2026 everyone! We have two posters this year, come and see. The first (tomorrow, 9 am) is about eye and mouse trajectories leading to behavioral biases in a #VWM task. The second (Tuesday at 3 pm) is about how we (mis)treat correlations when estimating averages.

0150 0

Andrey Chetverikov @achetverikov.bsky.social · 18/08/2026

That moment when you look at the data analysis script and think "Claude wrote some shitty code again", before realizing it's your code 🤦

040 0

Andrey Chetverikov @achetverikov.bsky.social · 16/08/2026

Liked this book too. An unusual setting, great example of what makes sci-fi and fantasy so interesting, they deal with scenarios that are plainly impossible, but make the real.

020 0

Reposted by Andrey Chetverikov

Alice Schwarze @aliceschwarze.bsky.social · 14/07/2026

Today marks a first in that I have found a mildly entertaining post on LinkedIn

Photo of a page in an Agatha Christie novel with em-dashes circled. The comment on the photo: "Not you too, Agatha Christie! Not you too"

0587 0

Andrey Chetverikov @achetverikov.bsky.social · 10/07/2026

I'm curious are there ppl who actually pay for the papers like that? Most articles are available online nowadays through authors or some well-known websites

030 0

Reposted by Andrey Chetverikov

Grumpy Philosopher @stevecooke.org · 10/07/2026

What, and I cannot stress this enough, the actual fuck? Do you know what a security guard on a UK campus carries? A stab vest, a radio, and some training in mental health and deescalation. Assault rifles? Wow. Such a messed-up country. www.theguardian.com/us-news/2026...

theguardian.com

‘We will never use them’: the California universities stockpiling AR-15s, grenades and submachine guns

A 2021 state law allows campus police to own military equipment for civilian safety – students fear it may be used to quash dissent

53515 0

Andrey Chetverikov @achetverikov.bsky.social · 29/05/2026

I'm not sure I get how AI helped when your data shows that AI-assisted was at best at the same level as human-only, and lost in terms of number of errors detected.

140 0