Working Memory Expansion Working memory capacity could be expanded by applying radio signal processing techniques to neural activity, allowing the brain to encode and combine abstract concepts through frequency modulation. Researchers propose that neurons using multiple carrier frequencies could correlate firing patterns to form larger, more complex ideas, potentially overcoming the limitations of human memory for technical tasks like AI safety research. Speculation. Over the last few posts, I talked about expansion of general human processing power by growing an AI to do whatever computations extra biological neurons would. This is powerful and flexible, but also has bigger value-drift risk and is probably technically harder than other approaches. There might be technically simpler ways of augmentation. It does seem that evolution made modular structures https://www.lesswrong.com/posts/JBFHzfPkXHB2XfDGj/evolution-of-modularity ; perhaps subcomponents of human cognition are similarly separable. For example, individual neurons correspond to https://doi.org/10.1038/s41562-023-01706-6 individual episodic memories 1 , and People can use mnemonics to get extraordinarily good at remembering simple structures like digits, but those don't scale to the sort of abstract concepts a technical AI safety researcher would find useful. We need better flexibility. One great thing about computers is that, unlike humans, if you give a Western Digital SN8100 a 20 million digit vector 2 , it sticks. If you need a 20MB sparse connectivity matrix to represent a single abstract thought, that's fine, your server has a terabyte of VRAM. There's a mesh of neuron tails which sits in the outermost micrometers of cortical tissue. It's called neuropil. Unlike other long-range brain connectors, neuropil has quite slow conduction velocities; it's unmyelinated, so signals travel slowly and are more metabolically expensive to send. Evolution found unmyelinated neuropil worth the processing lethargy and ATP costs; why? Time coding could be a good area to look, since temporal data is really important for neuron semantics. Timing improves https://doi.org/10.1038/s41467-024-48664-9 the decoding of macaque working memory items on spatially complex tasks, for example. And physical distance is coded https://www.annualreviews.org/content/journals/10.1146/annurev-neuro-072116-031538 as https://www.youtube.com/watch?v=iV-EMA5g288 temporal difference, in memory cells. As a further sign of time-coding's relevance, signal delay in neuropil causes https://www.nature.com/articles/s41467-021-26175-1 large, diffuse waves which spread across brain surface 3 ; a I fear I should first explain FM radio. Constructive interference, where multiple signals combine into a single loud signal, looks like this: when you sum across separate signals. This is how we get musical chords. We can also loop one signal and see how it interferes with itself ; by tuning the loop frequency, we can isolate the power of a carrier signal. 4 https://www.lesswrong.com/feed.xml fngoo32sn3nf If we smooth the summing process added EMA parameter , we can trade spectral for temporal precision https://www.youtube.com/watch?v=MBnnXbOM5S4 ; the receiver is less sensitive to frequency changes, but can come to a conclusion faster. By turning the carrier on and off, we can encode and decode lower-frequency signals. This is how AM amplitude modulation radio works. Shifting the carrier frequency likewise changes how well it sums on the receiver, and we can use it to encode signals too. In this last example, you can see that "frequency" modulation actually looks to the receiver like phase https://en.wikipedia.org/wiki/Phase waves modulation; slightly bumping the frequency causes each new beat to anticipate its rolling average. Harder anticipation quickly flattens the received power. For neurons, if we had multiple simultaneous carrier frequencies bands , we could correlate firing activity so that sub-concepts congealed into bigger and more abstract ideas. Any parts of the brain oscillating in the same band would, over time, get lots of activity from the conglomerate concept and average close to nothing for unrelated ones. Let's say that some part of my brain represents my dog, some other part represents sensations of running, and a third holds how sunlight filters through trees. Three modular circuits. When I imagine trail running with my dog, the three circuits "bind" together on a shared carrier wave. If you "tune your radio" to that carrier wave, you get exactly the signal "sunny dog visuals, proprioception of running". This is what an MIT-Sweden collaboration terms " spatial computing theory https://picower.mit.edu/news/spatial-computing-enables-flexible-working-memory ". The timing information here is flexible https://doi.org/10.1038/s41467-023-36555-4 in a way that the underlying physical circuits can't be. Flexible time codes are probably in some strong sense required for holding things in working memory 5 ; if working memory entries come with large-scale brainwaves, it's probably pretty easy to locate which patterns One might say that large-scale waves hold the control information about representations; wave timing reorganizes existing primitives https://www.lesswrong.com/posts/YLuifkTPR7TPLqoat/biological-computing-underhang into new concepts. By scanning for characteristic control waves, we can infer that