AI Predicts Neuroscience Study Results Better Than Experts New study shows AI large language models (LLMs) outperform human neuroscientists in predicting neuroscience study outcomes.

🧪 An AI model outperformed experts in predicting neuroscience study outcomes, per Nature Human Behaviour.

BrainGPT lead with 86% accuracy, surpassing human accuracy of 63.4% 🩺🖥️ 🧠

So hypothetically, we could ensure that critical functioning of society is resilient to a broad array of otherwise-catastrophic attacks.

Automated vehicles could be responsible for all food and item deliveries.
I assume we'd want to automate critical parts of the supply chain. Power plants, crucial food processing, telecommunications equipment, etc.

Heavy automation (self-driving cars, humanoid robots) would increase cyber-risk, but I assume it could greatly reduce bio-risk and other critical risks.
In theory, it could be fairly low-cost to ensure that most humans are highly isolated into discrete clusters.

There are other things like "understanding the reasons for things" and "knowing how to hunt down useful data sources."

If you want to imagine much better forecasting platforms, don't get too focused on the forecasting part. Empirical forecasting is a great way to anchor intellectual work, but it makes up the minority of such work.

I think prediction markets aren't really about prediction. They're about good epistemics. The `prediction` part is a way of forcing researchers to be meritocratic, honest, candid, and truth-seeking.

I think there's often a stronger case for decentralizing retrieval than synthesis, as information is often spread out throughout an organization.

My main point is that the value of "getting information from people in an organization" is often more about retrieval than synthesis, for a broad definitely of retrieval. You can often do the synthesis in one cluster, either with a few humans or with AI.

It's complicated. You'd want to give negative reward to opinion-heavy content that's proved false. You'd still probably want to give positive rewards for people who gave objective stats that just happened to promote the wrong view, if it seems like the direction was unintentional.

If you wait long enough (after the events happen), you could go back and update estimates of impact - so that people who gave the most useful information in support of the event that actually happened get extra rewarded, or something.

I feel like one better-than-internal-forecasting platform, very soon, would be a system that uses AI forecasts but allows individuals to give opinions+information. The AI forecasts would respond accordingly. People could then be provided with a score for how valuable the information is.

Yea, but it's difficult to refactor them that way. Good if you don't do much editing.

Do people have favorite apps to post semi-long-form to Bluesky?

Sad that messages have a limited length. Really don't like the interface for writing long messages by splitting them up into a bunch of tweets. Would prefer something that at least automates this.

Happy that more effective altruists are joining BlueSky. I really want to use BlueSky, but previously my colleagues weren't on it (it still seems slow). And when I searched for "effective altruism" here... I wasn't very impressed.