Post-Dunbar

What to do when we run out of memory

In the past 365 days, I’ve Zoomed 2313 unique people (if you’d like a copy of my script, leave a comment and I’ll shoot it over). Add the IRL connections from dinners, conferences, and happy hours, and subtract the people I already knew, and it comes out to about 2000 net new people. The last time I came close to that scale of memorization was when I had to cram 43 presidents, 118 periodic table elements, and 1021 Pokemon during a particularly busy stretch in middle school.

I started thinking about all this because of something embarrassing that happened last week. I was fresh off of a Europe trip and decided to go to a happy hour my investor was hosting. Jetlagged, I made my way over to the bar and was stopped by a very excited “Long time no see!” I looked up and tried to figure out who this guy was — A customer? A former classmate? Someone I’d met at a trade show years ago?

I tried to buy myself some time and blurted out the classic “How’ve you been?,” as I continued to frantically search my memory. Empty response body. As I stalled, someone else came over to introduce themselves and I heard the name of the person in question. It didn’t ring any bells. Luckily I got out of the conversation with minimal awkwardness, but hours afterward, I was still agonizing over that interaction.

Decades ago, a guy named Dunbar came up with his infamous number, 150. He said this was the cognitive limit for how many people we can remember and therefore have meaningful relationships with.

The threshold for a meaningful relationship, according to Dunbar, is if you know someone “well enough to greet without feeling awkward if you ran into them in an airport lounge.” It’s laughable how far we are from that world. It feels like every time I’m at SFO or JFK, I bump into someone I know. Generally speaking, I would probably say hello to any of my 5000+ connections on LinkedIn. Maybe Zoom and social media have made relationships more casual than they used to be, and maybe the definition of meaningful needs to be updated. But regardless, the sheer number of people we’re meeting has gone up, and looser relationships aside, we need to mentally maintain a broader set of contacts these days.

More recent studies suggest that, even if they’re not deep relationships per se, humans are capable of remembering a lot of faces, 5000 on average. It’s important to note some nuance here: Recognizing faces is a task of recognition, compared to remembering names, which requires recall. The number can be reasonably high for this.

Those researchers also found that variation in recall ability is large; some humans are just way better at remembering other people. To explain those differences, the paper suggests that “individual differences could reflect heterogeneity in visual cognition.” To put it another way, some people have better encoding algorithms for facial / visual data than others do. It’s reminiscent of how LLMs rely on embeddings, and perhaps it’s proof for how some types of embeddings algorithms might work better for certain datasets (ie human faces), or even sub-datasets (ie Asian faces vs rest of world). Empirically, it seems like memory for both human brains and AI relies on scoring algorithms that vary across people and models respectively.

Another study hypothesizes that variation in memory abilities "could reflect different social environments-some participants may have grown up in more densely populated places with more social input." Not unlike AI models, humans that have accumulated larger, more diverse training datasets seem to have an advantage.

All this is to say, humans have a wide range of “hardware specs” when it comes to memory, but even on the high end of the spectrum, our biological hardware is already failing us. Unless you have photographic memory, you’re probably struggling under the weight of all these new virtual connections. Digital socialization has outpaced our brains’ ability to store new data.

It’s becoming obvious that prior to Neuralink ubiquity and/or the emergence of other biotech hacks, LLMs will play a big role in memory augmentation. The limitations aren’t in the technology either at this point; they’re in the user experience. The key design considerations will be:

• Whether the solution leans more towards software or hardware — Initially I thought a player like Rewind.com might provide a fully cloud-based solution, but it’s looking like everyone wants to build a sexy device now. Bring on the AI chains…

• How the product will maximize inputs, both visual and auditory (ie textual)

• How the tool, or the go-to-market, will knock over peoples’ current attitudes towards monitoring / data collection / privacy

• How to rebrand this away from the trite “personal CRM” problem space

There are certainly more criteria I’m missing, so I’m keen on watching the space evolve and updating my opinions. With that said, if you’ve been thinking about this or have a cool prototype you’re working on, drop me a line.