Taking the test underwater, Most vital time to write, Déjà vu is not weaker memory
Weekly I/O #89: Context of Memory Recall and Formation, Vital Time to Write, Perirhinal Cortex and Deja Vu, Zero-One-Infinity Rule, All buildings are Predictions
Hi friends,
Greetings from San Francisco!
For this week’s email, I’m experimenting with removing “Weekly I/O” from the title because it takes too much real estate there. I apologize if this affects your email filter, and I would appreciate your feedback :)
And here are five things I learned this week and 1 piece I published last week. I hope you enjoy it!
Output
Here's what I've published since the last time we met.
1. On Taiwanese Names or: Call me Cheng-Wei not Cheng, thank you
A rant about people calling me Cheng, an introduction to Taiwanese names, and some cool naming rules.
Input
Here's a list of what I'm exploring and pondering on this week.
1. Memory retrieval is easier when the context of recall matches the context when the memory was formed. If you learn the content underwater, you will remember the content better when tested underwater than on the beach.
Paper: Context-dependent memory in two natural environments: on land or underwater
This is one of my favorite studies on memory (because it's quite funny, lol).
The researchers asked half of the participants to memorize a list of unrelated words while diving twenty feet underwater and the other half to memorize the same list on the beach. Later, they test these participants to recall these words underwater or on the beach. So essentially, there are four groups:
Learned the content UNDERWATER and recalled it UNDERWATER
Learned the content UNDERWATER and recalled it ON THE BEACH
Learned the content ON THE BEACH and recalled it ON THE BEACH
Learned the content ON THE BEACH and recalled it UNDERWATER
They found that if divers learned the words underwater, they remembered more words if tested underwater than on the beach. Likewise, if divers learned the words on the beach, they tested better on the beach than underwater. In other words, they can recall more when the test conditions match the learning conditions.
This context dependency is true for other kinds of memory too. No matter whether it's prospective memories (what you plan to do), episodic memories (what happened), semantic memories (information you know), or muscle memories (how to do things), memory retrieval is easier when the context of recall matches the context when the memory was formed.
Moreover, this context dependency isn't only about the environment. As I noted in Mood and Memory Recall, a sad mood can make sad memories easier to recall. Our emotion is part of the context.
This also relates to why changing study environments can significantly improve how well you remember information. If you want to recall a piece of information better anytime, anywhere, you should try to encode that memory with multiple different contexts, similar to how dual-encoding theory works.
2. There has never been a more vital time to write because your words online are voting on what values future AI systems will adopt. Write now because early data have a compound impact on AI.
Podcast: Gwern Branwen - How an Anonymous Researcher Predicted AI's Trajectory
From Gwern in this comment:
Much of the value of writing done recently or now is simply to get stuff into LLMs. I would, in fact, pay money to ensure Gwern.net is in training corpuses.
This is one of my beliefs: there has never been a more vital hinge-y time to write, it’s just that the threats are upfront and the payoff delayed, and so short-sighted or risk-averse people are increasingly opting-out and going dark.
Now is a crucial time to share your thoughts and ideas online because they will shape how artificial intelligence develops.
Large language models learn by studying text from the internet. If your values, sentiments, and perspectives aren't written down somewhere online, then to the AI they don't exist. Therefore, if you believe AI will be THE future, having your voice included subjects more than ever.
From Gwern in the podcast:
So if you aren't writing, you are abdicating the future or your role in it. If you think it's enough to just be a good citizen, to vote for your favorite politician, to pick up litter and recycle, the future doesn't care about you.
By writing and sharing your thoughts online, you're essentially voting on what values future AI systems might adopt and, by extension, what values future society might hold.
The timing is critical because early writings may have a compound impact, as future AI systems might learn from earlier AI outputs. If you don't write now, your perspective might be permanently left out.
Also see the discussion on Marginal Revolution.
3. Déjà vu is not simply a weak memory, but rather from a separate memory system in our brains. Perirhinal cortex creates familiarity while hippocampus handles the detail.
Paper: Episodic memory, amnesia, and the hippocampal–anterior thalamic axis
Scientists used to think our memories were simply strong or weak, like points on a line. There, the feeling of déjà vu, that peculiar sense of familiarity, was often explained as a "weak" memory with the assumption that our memories fall along a continuum. However, in the late 1990s, researchers Mishkin, Aggleton, and Brown discovered that memory actually has different subcomponents, and each can be either strong or weak.
Among these subcomponents, the perirhinal cortex supports familiarity-based recognition (knowing), while the hippocampus supports recollective-based recognition (remembering).
For example, the perirhinal cortex is like seeing someone and thinking, "I know that face," even if you can't recall anything else about them. On the other hand, the hippocampus helps you remember exactly where and when you met them and what happened during that meeting.
Research on people with damaged hippocampus shows this difference clearly. These people can still feel that things are familiar, even though they can't remember specific details about their experiences. And even without specific recall, this sense of familiarity can be so strong that it leads to a sense of certainty. They might be absolutely certain they've seen something before, but this familiarity does not provide access to the "where" and "when" of that event because of their damaged hippocampus.
In essence, déjà vu is not a weaker form of memory, but rather a different type of memory altogether. So when you feel the sense of familiarity but don't remember the details, the perirhinal cortex in your brain is activated without using the hippocampus to gather the temporal and spatial context.
4. Zero-One-Infinity Rule: When designing software systems, the only reasonable numerical limits are zero, one, and infinity.
Article: Zero One Infinity Rule
Zero-One-Infinity (ZOI) is a software design rule of thumb. It states that a programmer should not place random limits on how many instances a given entity should have. The only reasonable limits are zero, one, and infinity. In other words, you will either need zero of a thing, one of a thing, or an arbitrary number of the thing. Any other arbitrary limit increases the cognitive workload for both the users and the software designers.
A simple example is the file system's folders (directories). The topmost folder must have "zero" parent folder. Each subfolder must have exactly "one" parent folder. Each folder can have "any number" of subfolders.
Some people also use this as a heuristic outside of software design. For example, if you're going to believe in more than one universe, you might as well believe in an infinite number of them. Same for estimating how many alien planets there are.
5. All buildings are predictions, and all predictions are wrong. Architectural design should create adaptable spaces rather than attempting to perfectly predict and design for future needs.
Book: How Buildings Learn
Buildings are designed based on assumptions about how they will be used, which essentially makes them architect's predictions about future needs and functionalities.
However, the predictions are almost always wrong because the future is largely unpredictable. The needs and desires of building occupants change over time, often in unexpected ways. For example, an office building designed in 1990 couldn't have anticipated today's hybrid work culture and different workspace needs.
Therefore, an architect should focus on core structure and flexibility. They should prioritize a strong and adaptable structure over intricate details, allowing for easier modifications and renovations as needs change.
This also relates to piecemeal growth from Christopher Alexander, who suggests continuous incremental improvements and adjustments to buildings throughout their lifespan. This idea in architecture design is similar to premature Optimization Is the Root of All Evil in software design.
Two other architecture books that I might write about in the future are A Pattern Language and The Timeless Way of Building. They are often referenced in software and product design. Welcome to recommend more books in this field!
Recap
Try answering these five simple questions to review and reinforce what you've learned:
That's it. Thanks for reading. Please share which input you found the most helpful or intriguing. Just reply to this email with a number—it's quick and easy!
And as always, feel free to send me any interesting ideas you came across recently!
Looking forward to learning from you.
Cheers,
Cheng-Wei
Subscribe to Cheng-Wei’s Update | Subscribe to 程維的中文更新 | Subscribe to Weekly I/O | Facebook | Twitter