cd /news/artificial-intelligence/biologically-plausible-sgd-is-hard · home topics artificial-intelligence article
[ARTICLE · art-15688] src=lesswrong.com pub= topic=artificial-intelligence verified=true sentiment=· neutral

Biologically Plausible SGD Is Hard

A biologically plausible stochastic gradient descent (SGD) algorithm is required to simulate the effect of additional brain mass for human intelligence augmentation, but the correct local learning rule remains unidentified. Researchers lack consensus on how neurons implement local updates, making it difficult to train AI to replicate the patterns of extra neurons and close a feedback loop in milliseconds. Without this rule, efforts to boost human intelligence through simulated brain mass will fail to produce superintelligent humans.

read1 min publishedMay 27, 2026

This is the hardest challenge I foresee with BCI intelligence augmentation.

The simplest architectural prior for improving human intelligence is to simulate what a brain would do given more compute. This is usually more time ("what conclusion would I come to given 100 years of thought"), but more brain mass seems more tractable given current technology.

Consider that models of long-term cognition get very sparse feedback. You can't finetune an LLM on your reflective trajectory across a century because that would take an entire lifetime. With high-resolution neural connections (and a robust understanding of how neurons learn), we could close a mass-based feedback loop in milliseconds-to-minutes.

Given better genetic engineering tech, we might give people a virus which 10x'ed their neuron count without terrible physiological side effects. We very probably can't do that this decade because genetic edits are hard.

But we do have a lot of GPUs.

If we train an AI to send similar patterns as lots of extra (developmentally integrated) neurons would, we've effectively boosted someone's brain mass. I strongly suspect that the relevant algorithms for training such an AI are compactly specifiable, since it seems like neurons learn from simple local firing statistics. It's like extending brain maturation beyond 2-to-3-decade typical timelines. Existing biological circuits specify some signal similar to a gradient error, and the computronium updates commensurately, just as any section of the universally-learning parts of the brain do.

An incorrect learning rule almost certainly doesn't cause competent agentic misalignment, but it won't make a superintelligent human either. Locating the right rule seems very hard; there does not seem to be consensus on how local updates are implemented.

── more in #artificial-intelligence 4 stories · sorted by recency
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/biologically-plausib…] indexed:0 read:1min 2026-05-27 ·