Part Of: Links sequence

Biology

• Experts of mind control: Parasites are oddly effective at hacking nervous systems to behave against an animal’s best interest (for humans, c.f. Toxoplasma gondii).. If biological minds are so profoundly susceptible to neuroengineering, will not AGI be similarly vulnerable to pref-hacking? . 
• Protosterols discovered shed light on the “Boring Billion” – what eukaryotes were doing between 1.8 ga and 0.8 ga.

Cognitive Science

• A few months ago, a new ICN was discovered: the somato-cognitive action network (SCAN). This action-oriented SCAN is located in the interstices of M1. More recently, the cingulo-opercular network is strongly functionally linked to SCAN. Given its high-level position, the “cingulo-opercular” network has been renamed the Action Mode Network (AMN) to align with our new understanding.

AI

• The mechanistic interpretability crowd has finally discovered how to locate “grandmother neurons” in real (non-toy) LLMs. Big news for the future of alignment.
• A nice visual introduction to the Manifold Hypothesis. 
• Alpha Fold 3 released. 
• GPT-4 is able to predict the future (forecast events after its training cutoff) quite accurately… but only when asked to tell stories about what will happen. 
• In some tasks, AI seem to be asymptoting at ~110% human performance. Only in a couple categories are LLMs dramatically better (chess, persuasion, …)

• In the aftermath of Cambridge Analytica, I had concluded that big-data microtargeting (disinformation campaigns) are overblown, or not very effective in practice. But LLMs are consistently achieving superhuman results in persuasion. While microtargeting might not have been very effective in 2016, but that may change in 2028. 

Physics

• Roger Penrose long ago proposed that the human brain uses quantum effects in microtubules and that was the origin of consciousness (Orch-OR). For a long time, microtubules were thought to be “too warm and wet” to sustain quantum coherence. But they have been shown capable of sustaining quantum effects. Implications both for quantum computing, and also the Orch-OR hypothesis. 

Other

• AI, and automation more generally, seems to damage religious affiliation. Supporting data includes 1) correlations between- and within-nations, 2) longitudinal studies tracking change in religious attitudes for different automation-exposure, and 3) experimental support.  (This result confuses me, because it doesn’t align with my understanding of why people practice religion. The inverse correlation between welfare and religion is more explicable..)
• France began the demographic transition (aka fertility crisis) one century before the rest of Europe. There are two primary explanations for the transition: cultural evolution and life history models. The evidence from France supports the former (perhaps the early secularization played a role; the pro-natal Catholicism message lost clout).