Nick Bostrom: "[Yudkowsky] has this recent book with Nate Soares, 'If Anyone Builds It, Everyone Dies.' Now, my view is that if nobody builds it, everyone dies. In fact, most people are already dead who have lived, and the rest of us look set to follow within a few short

From my POV, I am trying to *save everyone's lives* by accelerating AI. My view is that AI will accelerate medical cures that could save the lives of billions. Delaying AI therefore risks killing billions of people. I am on the side of life, not death. I want us all to live.

An unfortunate reality of the internet is that when you say something brief or ambiguous, people who disagree with you will often read in silly interpretations you didn't intend. They'll refuse to take your words at face value and will assume you implied things you didn't say.

Should we create agents that fully take over people's jobs, or create AIs that merely assist human workers? This is a false choice. Full automation is inevitable, whether we choose to participate or not. The only real choice is whether to hasten the inevitable, or to sit it out.

I'm sure there are people who really think humanity should be replaced by AIs, but that's not what I think. I just think we should be happy to let advanced AIs get legal rights and social influence, just as we are for human children who grow up. I want coexistence, not war.

If you didn't know, @IBDinvestors publishes a quarterly Stock Guide in a free PDF. The latest one titled "The ETF Playbook: Strategies for Growth Investors" just dropped couple days back. There is always something to take away from that publication ⏬⏬ https://t.co/ij8GrYWNw4

I wonder how much higher birth rates would be if everyone were familiar with these research results.

Twin and adoption studies consistently show that parenting choices have minimal effects on a kid's eventual intelligence, personality, or happiness (except in cases of extreme neglect or abuse). This should revolutionize how we raise children, yet almost nobody knows or cares.

A big caveat here is that the concept of "thinking speed" is quite vague. In this thread, I'm referring to something like "the serial speed at which major, top-level decisions are made" or, in the context of LLMs, the sequential forward passes per second (_not_ total throughput).

It's important not to confuse what I'm saying. I'm not saying computer hardware will slow down or algorithms will get worse. For a fixed level of AI capability, "thinking speeds" will increase over time. But the thinking speed of the most intelligent model at any time may slow.

⚡ Bimergen Energy $BESS: Pioneering the Future of U.S. Grid Stability Join us Tuesday, Nov. 11 at 4:15 PM ET for a live investor webinar with the Bimergen management team. Learn how $BESS is advancing 2.0 GW of battery energy storage projects across major U.S. markets: a $400M

This is one reason to expect that future superintelligent AIs will not be run at very high "thinking speeds". This may sound paradoxical, since computer clocks are so fast. But it makes sense if these extremely smart AIs are mainly being used to make very high-quality decisions.

There's a trade-off between thinking quality and thinking speed. This is apparent in LLMs, where larger LLMs are run at slower tokens/s than smaller ones. But the phenomenon is more general: institutions are slow and deliberative if they care about quality (e.g., courts).

I continue to think that a benign AI takeover is likely both inevitable and desirable. As AIs become more agentic and capable, they will gradually assume more responsibilities, gain legal rights, and earn social influence. There's no need for a dramatic coup, or treacherous turn.

Here's an example of the type of claim I'm talking about: "If engineers are growing AIs that they don’t understand, then they have far less ability to shape how those AIs are going to behave. Lack of understanding constrains engineering." This is from

People often suggest we need to understand how AIs work to shape their behavior. I disagree. We don't understand how human brains work, yet we have many tools to shape human behavior. By contrast, we understand the motion of planets almost perfectly, yet we can't control them.

People often seem to think Mechanize is simply building clones of existing apps for AIs to navigate, like Doordash or Slack. That's not true. We're actually developing novel SWE problems and test suites to advance AI coding abilities. This is different and a lot more exciting.

We are looking for an ML engineer to lead our new initiative to fully automate AI R&D as soon as possible. Making AIs smart enough to build the next generation of AIs has been a long-term goal of the field. Join us to make this happen.

I personally work on automating labor because I expect the benefits of vast abundance and new product variety created from AI automation will outweigh the negative effects on human employment. Cost-benefit reasoning is unlikely to win you many friends, but it's far more honest.

There's no reason to expect automation to always complement labor rather than substitute for it. Nor has automation always raised employment: male labor force participation has fallen since at least 1950. Rather than bullshitting about this, AI advocates should embrace reality.

I have nothing against open source in principle, just as I have nothing against donating to charity. But people keep being more optimistic about the feasibility of open source AI than the evidence warrants. Until people wake up to this fact, I suggest betting against open source.

Yet open source makes even less sense for AI than traditional OSS. Efficient AI inference depends on massive returns to scale from large data centers. Almost no one can run large open source models on their personal machines, so people end up relying on commercial vendors anyway.