iiinova — where your digital self keeps evolving

Brad Wyble @bwyble.bsky.social · 19/08/2026

Also, you have to squint really hard to see AI using students doing "better" while equating for homework time.

Brad Wyble @bwyble.bsky.social · 23/02/2026

We need a new version of I, Robot where a robot behaves normally for a few days and then suddenly starts killing everyone because its context window got compressed, which changed the first law to: "A robot shall harm a human being or allow a human to be harmed"

Screenshot from X that reads "Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb." And then as a screenshot of interactions with some agent that is typing commands that will delete her inbox.

4113

Brad Wyble @bwyble.bsky.social · 18/02/2026

Nowadays this email subject line could have two **completely different meanings**

An email with the subject "First shots fired in massive Wisconsin race"

000

Brad Wyble @bwyble.bsky.social · 08/01/2026

In the 1940's would could buy a ring for 15 cents that had radioactive material and if you looked inside you could see the particles hitting a little screen. *Stares sadly at the latest Happy Meal toy* (btw one of these is for sale right now on ebay)

Image of the Lone Ranger Atomic ring that you could buy in Kix cereal

162

Brad Wyble @bwyble.bsky.social · 06/01/2026

such as when the top-two rows are in the first array and the bottom-two rows in the second. More complex patterns, like the one in the image here, had very little improvement over time. (the data shows the average of the simple/complex patterns) (16)

Illustration of data from a task in which there are two arrays of dots that do not overlap. Subjects get better at the task as the ISI increases from 200 to 1000 ms.

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

However this flexibility is difficult for cognitive scientists because memory representations are highly contextual. It means that we might represent an array of colored squares as a set of separate objects, or instead as a single holistic representation. (12)

Illustration showing a memory of four colored squares being stored as four separate items or one holistic representation

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

Or if we view the same scene with the goal of remembering the steering wheel, we might create a memory more that emphasizes the visual detail of wheel specifically, maybe including the Mercedes logo. (11)

Image of the same car dashboard but now remembered in a way that emphasizes the steering wheel

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

We also assume that memory for one scene or object can be composed of several different *kinds* of representation. For example when we perceive this car dashboard we might build a memory containing a coarse-grained representation of the whole as well as memories of the familiar objects within. (10)

Illustration of how a picture of a car dashboard is converted into memory with multiple parts, both visual and conceptual

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

For example when asked to visually imagine a D, rotating it, and adding 4 on top, subjects can report that the resulting shape is a sailboat (Finke et al. 1989). Memory stores ‘4’ and ‘D’, and then generative mechanisms reconstruct these as visual reconstructions into the perceptual hierarchy. (7)

Visual illustration of combining a D and 4 to create a sailboat

220

Brad Wyble @bwyble.bsky.social · 06/01/2026

We assume also that generative processing allows conceptual information to be reconstructed in a format akin to visual sensory representations that can be manipulated and re-processed by perceptual mechanisms. (4)

Conceptual illustration of how a verbal description such as “red car” could be projected through a generative hierarchical model to create a noisy visual instantiation.

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

We assume that compositionality allows complex scenes or objects to be mentally decomposed into constituents that can be individually manipulated or recombined to form new forms. (3)

An illustration showing one possible compositional visual representation of a car including its components, its overall shape, and the most prominent color. In our framework, a memory representation could access some combination of any or all of these depending on the task.

110

Brad Wyble @bwyble.bsky.social · 06/01/2026

We recently published a theoretical review about how compositional and generative mechanisms in working memory provide a flexible engine for creative perception and imagery. Pre-print: osf.io/preprints/ps... Paper: www.sciencedirect.com/science/arti...

Illustration of the hypothesized flows of information between perception, memory and cognitive control in a conceptual model of working memory. Stimuli attributes are processed to varying degrees of abstraction and parts of these representations can be loaded into working memory under the guidance of cognitive control. Familiar stimuli such as the letter B activate visually abstract representations while less familiar stimuli are limited to sensory representations. Information can be shifted both up and down levels of the perceptual hierarchy to build either more or less abstract representations of either perceived or imagined stimuli. Working memories can be shifted into or out of the hierarchy as needed.

38634

Brad Wyble @bwyble.bsky.social · 29/12/2025

Journals discover that they can nag you about a review being late before it's actually late.

Screen shot of text complaining that a review for a paper was due on Dec 29, but it's currently Dec 28...

040

Brad Wyble @bwyble.bsky.social · 13/11/2025

What I expected, of course, is that memory for these passages would improve systematically from 0th order to full text, but that's not quite what happened:

Line graph showing how each of the 4 passage lengths (10,20,30,50 words) were recalled at each level of approximation (0,1,2,3,4,5,7,text).

Accuracy increases steadily as the list of words gets shorter, and also increases as the approximation gets closer to text. However it hits a ceiling at the 5th level and does improve beyond that point.. It's just flat across 5,7 and text

111

Brad Wyble @bwyble.bsky.social · 13/11/2025

Examples of complete gibberish (which they call a 0th order approximation to english):

10: byway consequence handsomely financier bent flux cavalry swiftness weather- beaten extent
20: betwixt trumpeter pebbly complication vigorous tipple careen obscure attractive consequence expedition pane unpunished prominence chest sweetly basin awoke photographer ungrateful

30: crane therewith egg journey applied crept burnish pound precipice king eat sinister descend cab Idaho baron alcohol inequality Illinois benefactor forget lethargy fluted watchtower attendance obeisance cordiality dip prolong bedraggle

50: hammer neatly unearned ill-treat earldom turkey that valve outpost broaden isolation solemnity lurk far-sighted Britain latitude task pub excessively chafe competence doubtless tether backward query exponent prose resourcefulness intermittently auburn Hawaii unhabit topsail nestle raisin liner communist Canada debauchery engulf appraise mirage loop referendum dowager absolutely towering aqueous problem

110

Brad Wyble @bwyble.bsky.social · 13/11/2025

Examples of coherent prose of varying length:

10: the history of California is largely that of a railroad

20: more attention has been paid to diet but mostly in relation to disease and to the growth of young children

30: Archimedes was a lonely sort of eagle as a young man he had studied for a short time at Alexandria Egypt where he made a life-long friend a gifted mathematician

50: the old professor's seventieth birthday was made a great occasion for public honors and a gathering of his disciples and former pupils from all over Europe thereafter he lectured publicly less and less often and for ten years received a few of his students at his house near the university

100

Brad Wyble @bwyble.bsky.social · 17/09/2025

I changed my profile picture, which they are welcome to use in perpetuity, worldwide, in any manner.

260

Brad Wyble @bwyble.bsky.social · 08/09/2025

Prince of Persia was a groundbreaking game for its accurate character movement back in the late 80's. To do this, the creator videotaped his own brother running and jumping on the sidewalk and then did pixel mapping frame by frame as shown here at the Strong Museum of Play

Photo of the frames of a video of someone running and jumping for the game Prince of Persia.

3695

Brad Wyble @bwyble.bsky.social · 28/08/2025

In our two experiments, we found, then replicated, an accuracy boost for displays containing related pairs in visual working memory. These results suggest that our functional understanding of object interactions facilitates working memory compression. >

120

Brad Wyble @bwyble.bsky.social · 28/08/2025

Our new study (Titled: Memory Loves Company) asks whether working memory hold more when objects belong together. And yes, when everyday objects are paired meaningfully (Bow-Arrow), people remember them better than when they’re unrelated (Glass-Arrow). (mini thread)

56815

Brad Wyble @bwyble.bsky.social · 07/07/2025

Look at this interesting google trend graph for "attention". Those cycles peak in October and March of each year like clockwork. My guess is that these are Psych perception & cognition classes. Are there other explanations? @sbmost.bsky.social @bjbalas.bsky.social @benwolfevision.bsky.social

A blue line graph showing relative interest in the term "Attention" on google from 2004-2025 with a very rhythmic trajectory course

350

Brad Wyble @bwyble.bsky.social · 22/05/2025

Drawers? Why not a whole dresser?

220

Brad Wyble @bwyble.bsky.social · 02/05/2025

But it's just not doing a great job, still. And yet we're being forced to integrate into our jobs. Being upset seems completely reasonable.

Google's AI summary suggesting that Bob Dylan weighed between 6 and 7 pounds, and links to a reddit post where this is mentioned.

320

Brad Wyble @bwyble.bsky.social · 14/04/2025

New Preprint! Interested in learning about how working memory is subserved by both compositional and generative mechanisms? Read on!

Diagram showing how a perceptual hierarchy is bidirectionally connected to a working memory system and also cognitive processes

1318

Brad Wyble @bwyble.bsky.social · 27/03/2025

I asked a friend with access to the new GPT image model to create a poster aggressively attacking my work on the attentional blink and it created this masterpiece. Perhaps we finally found a practical use for these models.

A poster that reads: "The attentional blink? More like the attentional myth. Brad Wyble wants you to believe your brain blinks and forgets the second target... But what if that's just bad science in disguise? Correlation isn't causation: Just because T2 drops in accuracy doesn't mean there's a blink, it could be interference, expectation or task design flaws. Alternative models exist: Have you head of the boost and bounce theory or object updating models? Brad conveniently doesn't mention them when pushing the blink. Cherry-picked data? Manyi AB studies rely on RSVP tasks that create the blink effect. real-world attention doesn't work in PowerPoint slides. Stop Blinking. Start Thinking. It's time to leave the Blink behind and demand a better model

10882

Brad Wyble @bwyble.bsky.social · 17/01/2025

As a test of google, I searched for the attentional blink, which I know a lot about: So much of this is wrong. It's not an inability to recognize, it's not that the T2 is "too soon" ( lag 1 sparing eliminates this). Just really sad that simplistic and wrong takes are now presented as fact.

Google search results for the attentional blink presents this summary "The attentional blink (AB) is a temporary inability to recognize a second target that appears very shortly after the first. It's a result of the brain's limited capacity to process information.
How it happens

The brain is processing information from the first target

The second target appears too soon for the brain to process it
The brain is unable to allocate enough resources to process both targets

How it's studied

The AB paradigm presents stimuli like letters, numbers, or pictures in rapid succession

Participants are asked to report if they saw the second target

Real-world consequences

The AB can have serious consequences in real-world situations, like when a driver is unable to notice a deer in the road

Related concepts

Repetition blindness: The difficulty of recognizing a second target that is the same as the first in a rapid serial visual presentation task"

030

Brad Wyble @bwyble.bsky.social · 17/01/2025

I haven't gotten the Gemini upgrade to my gmail yet but I'm doing what I can to avoid it

Google gmail feedback form filled out to read "Please do not add Gemini to my email. I do not want, nor do I need an AI nanny looking at my private correspondence and making suggestions. "

060

Brad Wyble @bwyble.bsky.social · 13/01/2025

People keep rediscovering what is essentially the testing effect, which psychologists identified at least as early as the 1960's.

Screen shot of page 1 of the article titled: "Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom" in the journal PNAS

1204

Brad Wyble @bwyble.bsky.social · 31/12/2024

These are not just from gotcha's either. E.g. check out this output from Gemini. Not only is the information dangerously wrong, it is inconsistent within the space of a few sentences.

Text from Google's AI Gemini that reads "A resting heart rate of 100 beats per minute is considered "normal" for most adults, as a typical resting heart rate falls within the range of 60-100 beats per minute; however, if your resting heart rate consistently stays at or above 100, it could indicate a condition called tachycardia and you should consult a doctor to investigate potential causes"

000

Brad Wyble @bwyble.bsky.social · 31/12/2024

Still batting 100 Gemini. (Out of 1000)

AI search results from google Indicating that the rabbit's name is Binky.

This is incorrect and probably a bastardizarion of Binkley who is an entirely different character.

010

Brad Wyble @bwyble.bsky.social · 30/12/2024

Getting 80 young adults to answer 50 ambiguous MCQ questions the same way I would.

White text on black reading "Poorly explain what you do for a job."

190

Brad Wyble @bwyble.bsky.social · 14/12/2024

I feel it's like this though. Why does every place need to cater to every viewpoint?

Cartoon with three panels, in the first a large group are telling some people that they don't belong. Those people retreat and make their own place, then the large group is asking them why they are excluding them.

030

Brad Wyble @bwyble.bsky.social · 29/11/2024

Legendary game of hearts last night.

Handwritten scores on a piece of paper showing many consecutive 26's

130

Brad Wyble @bwyble.bsky.social · 12/11/2024

Slightly better but you can see why AI is no replacement for real art

110

Brad Wyble @bwyble.bsky.social · 12/11/2024

Explaining to my grandkids that back in my day a picture like this required you to actually get 700 beds and put them on the beach instead of asking AI to draw it. [Left picture is from Pink Floyd's 1987 Momentary Lapse of Reason album]

Image of Pink Floyd's momentary lapse of reason album cover photo showing hundreds of beds on a beach.

Picture of three women arranging the bedding on said beds, presumably before the shot was taken.

2304

Brad Wyble @bwyble.bsky.social · 09/10/2024

When you definitely understand correlation vs causation:

Screenshot of an article titled "Trump and Harris Fight for Pennsylvania County known for Picking Presidents"

040

Brad Wyble @bwyble.bsky.social · 23/09/2024

Here's one: Altmetrics for one of our papers, preprint went live in 2016, paper published in 2020. Note that most of the attention it generated was in the first surge, all the way on the left. This paper sits within the top 1% of pubs for its category based on just posting it on biorxiv>

201

Brad Wyble @bwyble.bsky.social · 23/09/2024

Amazing that a for-profit publishing company with billions in profit is asking me to dig up email addresses for them when I suggest alternates. You can google it yourself buddy. Also... fax machines? C'mon Elsevier. You can spend $10,000 on a webUI cleanup.

Screenshot of text that reads "Suggestions for alternative Reviewers (if possible, please include an email address, telephone number or fax number)."

011

Brad Wyble @bwyble.bsky.social · 21/09/2024

The AI-empowered Thesaurus.com entry for battery acid exemplifies how empty the semantic model of LLMs is. Some people do use battery acid as a slang term for strong coffee, so it is associated semantically, but that does not make these synonyms. (also battery acid is not a real word)

Thesaurus.com entry for battery acid indicating the strongest matches are caffeine, cappuccino and espresso.

171

Brad Wyble @bwyble.bsky.social · 21/09/2024

If you thought that was crazy, how about this:

Screenshot from Thesaurus.com indicating that the strongest synonyms for battery acid are caffeine, cappuccino and espresso.

020

Brad Wyble @bwyble.bsky.social · 06/09/2024

Also wild that the word "Operator" is effectively a backdoor password that gets you past many AI voice script's. You could spend an hour arguing with it, or just say Operator and they often connect you to a person. Life imitates art.

Image of Trinity from the movie the Matrix talking on a payphone with one hand outstretched toward the camera.

010

Brad Wyble @bwyble.bsky.social · 26/08/2024

Reminder that LLMs are incredibly dumb. Not only are they worse than small children at Tic Tac Toe but they don't even understand the rules. This is Co-Pilot. GPT4o is similar. You shouldn't use these tools for anything important unless you rigorously check the results.

An illustration of a text version of Tic Tac Toe wherein the board is full and noone has won yet the text indicates that the LLM thinks it has won.

120

Brad Wyble @bwyble.bsky.social · 08/04/2024

The totality as seen in Lux. The camera was recording 78000 before it started.

Graph of the number of lumens by time. Starts at about 30 thousand and dips to 0 and then starts increasing.

000

Brad Wyble @bwyble.bsky.social · 04/04/2024

Looking forward to this Squid Game sequel

Image of a man being fed a tomato by a red robot perched on his back. The caption is "Robot will feed you tomatoes while you run"

000

Brad Wyble @bwyble.bsky.social · 28/03/2024

I just discovered this evidence of the first beta test of Twitter back in the 1700's.

Painting of the fighting between Tom Johnson, and Isaac Perrins, of England and Birmingham respectively. Gentleman's Magazine 1789. The fight lasted 62 rounds. https://en.wikipedia.org/wiki/Isaac_Perrins#:~:text=Isaac%20Perrins%20was%20an%20English,the%20English%20Champion%20Tom%20Johnson.

010