On July 11, JD Vance argued that Big Tech CEOs are scaring people about AI’s dangers to secure their position at the top of the industry. Maybe the craziest thing about the AI risk debates is the strange bedfellows it generates. Versions of this argument have been made by...

“And also, nobody knows how planes really work, we just grow them in a large puddle and pick the ones that seem to fly. Impossible to know right now what they'll ultimately do though. Especially the ones which are really good at flying. Those always surprise us the most.“

One thing that is still misunderstood about advanced AI is that we really don't know how AI works If anyone should know how AI works, it's so-called "Godfather of AI" Geoffrey Hinton. But here he is in 60 minutes saying "we don't know exactly how [AI] works":

My favorite way of saying this recently is as follows: we know how iPhones work, because engineers designed every part of an iPhone. But we don’t know how AI works, because AI is trained/evolved and not designed. We feed in data and get out a model that does stuff, very different

In conventional software, many eyeballs make all bugs shallow. But in gigascale neural networks, all the eyeballs in the field haven’t even made 0.1% of the *features* shallow.

I really appreciate how straightforwardly Stuart Russell describes this situation. Often people discuss such things with euphemisms, or technical language that doesn't quite describe the actual consequences; imo clear speech is surprisingly helpful here. x.com/ai_ctrl/status…

#OTD 1953: Crick sent Schrödinger a letter thanking him for his book "What is Life?" that inspired him and Watson to enter molecular biology.

The next thing to jump out is that the idea that we’re really in the dark on how to build AI systems safely doesn’t strike me as a particularly strong argument for not regulating them! …

Occasionally, people used to claim that an event like this would trigger some sort of societal response!

Rob Miles (✈️ SF)

Randolph Carter Nuclear weapons didn't need to have precursor tech in chemical explosives that exhibit the same risk/benefit characteristics and was slightly less powerful. There's allowed to be a discontinuity in moving from TNT to nukes, because nukes aren't physically just "TNT but bigger".

Potential violations: - If the former nonprofit's executives and Board acted not in accordance with the nonprofit's mission by transforming OpenAI into a for-profit, including the downstream effect of permitting outside shareholders in that for-profit to which the for-profit

70% epistemic confidence: People will talk about Lighthaven in Berkeley in the future the same way they talk about IAS at Princeton or Bell Labs.

AI companies currently pay their staff in charge of deciding whether models are safe to release (e.g. leadership, red-teamers) largely in equity, the value of which seems likely to decline if models aren't released. I suggest paying them in cash instead. lesswrong.com/posts/sMBjsfNd…

roon This suggests a tension which is the opposite of reality. Doom isn't what we risk to get to heaven - doom is what stands between us and heaven.

What are people doing with their minds when they claim future AI "can't" do stuff? The answer is rarely «reasoning» in the sense of natural language augmented with logic (case analysis) and probability. I don't know if Eliezer's guesses are correct about what most scientists

had an emotional moment staring at a running faucet this week. we really don't appreciate civilizational infrastructure. literally "pay as you go" unlimited clean drinkable water delivered to everyone, physically piped straight to you and whisked away as soon as you dont need it

Always be on the look out for the pejorative-just, where someone says something is just something else. Usually if you take the word “just” out of the comparison, it becomes simple and anodyne truth.

Jack Clark Jeffrey Ladish So, here are my main objections: 1. "There isn't real evidence for X", and a large amount of what you wrote in the GPT-2 retrospective, is ambiguous between 'there isn't enough evidence for X to convince me of X', and 'there isn't enough evidence of X to convince [a group I have

davidad 🎇 Researchers with more gravitas usually invoke the no free lunch theorem to peddle the same idiotic nonsense. osf.io/4cbuv