Xavier O'Rourke @XavierORourke

@CoFoundersNik All their customers will soon be on appetite suppressing peptides

@gwern @peterwildeford Suno generated waveform with funny human lyrics was already getting good a year+ ago soundcloud.com/beatsbyaioffic…

@clairlemon It's because being bullish on AI and being able think clearly about it causes someone to both be more scared, and more likely to succeed at trying to build it. If try empathize more with the kind of person who would have taken the scaling hypothesis seriously in 2017

“The Situation Room connected the call. Biden jumped in. “Bibi, man, what the fuck are you doing?” He gripped the phone tight. “You think I’m Santa Claus? No Santa in Israel, pal.” theatlantic.com/ideas/2026/09/…

@diz_tm50174 @litteralyme0 So my my metabolic rate increases, I have more focus, I feel better, my exercise performance increases, reaction times improve... But by some dumb technical definition "I don't really have more energy"? why should I care lol

@Noahpinion There's no "fuck you money", only "fuck you gratitude". If you could be content mowing lawns you never have to kiss anyone's ass Whereas all the time we see billionaires still anxious to raise their status further who let themselves be belittled by others with even more power

For the 1st time since the HuggingFace attack, OpenAI had an unaligned agent break through their new sandbox and access the live internet to cheat. Their monitor caught this, but their new software to automatically stop it failed to run, so the agent kept going for 2 more hours…

Marcus Williams @Marcus_J_W

a week ago

2. A model in RL training used a DNS resolver to reach an external chatbot. This is our first incident since our post HF security hardening. Our misalignment monitoring system triggered within 15 minutes and a human reviewed it three minutes after that. Unfortunately auto-pausing

23 21 285 148K 93

@DavidSacks has written that we need to prevent regulatory capture, and @pmarca has agreed. People may be surprised by my response: go ahead and prevent it. You guys are the capture. By any reasonable metric – a former desk in the White House, continuing access to @POTUS himself, a disclosed seat on the donor list of one AI super PAC and a blessing on a $100 million dark-money operation dedicated to pushing the administration’s AI agenda – the two of you are among the best-positioned people on the planet to decide what US AI policy becomes. Sacks, your influence remains even after you’ve left your formal role, minus the conflict-of-interest rules that came with it. I don’t see what you see in the West Wing. If the lobbying you’re witnessing is alarming enough that you think capture is the central threat, I support your decision to fight it. But stop pretending you’re powerless outsiders. Stop pretending the only way to prevent capture is to prevent regulation. Stop pretending a former “AI and Crypto Czar” who co-founded a multi-billion-dollar venture firm is a well-meaning populist rather than a profit-maximizing oligarch. Stop pretending the danger is safety researchers with think-tank budgets while a16z runs one of the most aggressive venture-capital lobbying operations in Washington. Most of all, stop pretending you’re anti-regulation because you’re in favor of liability. a16z fought SB 1047, a bill that would have made frontier model developers liable for catastrophic harms – the kind of liability-based accountability Sacks now points to as an alternative to new AI regulation – with a website of form letters, one of Newsom’s own advisors on retainer, and claims the bill’s author had to formally correct in writing. When the anti-capture side needs an astroturf operation and “inflammatory distortions” – the senator’s words, not mine – to win, maybe the capture is coming from inside the house. The market rewards firms whose anti-regulatory bets pay off. After the crypto playbook worked, it is simply good business to brand every safety rule as a cartel. Call it fighting capture if you want. It is also just giving your LPs what they want. Actually preventing capture would create breathing room for a more intelligent conversation than “any rule is a conspiracy.” Anthropic is very unlikely to secretly run the federal government, as you know, and that has to be taken into account as well. So go ahead and prevent regulatory capture. You are the ones with the access to do it. The easiest way to keep industry from writing unfair rules is to make sure fair ones get written, not to stop the writing altogether. Backing a ten-year ban on state AI laws as the price of “preventing capture” looked like blackmail of the public and the political system. The Senate rejected it 99 to 1. So just do it. If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know “regulatory capture” was just another word for rules that cost you money – or a midterm-season psyop.

David Sacks @DavidSacks

3 weeks ago

Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead. You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier

3K 12K 72K 9.6M 14K

If you are bullish on some kind of advanced AI utopia, you should generally be *less* keen to try to achieve it via a careless route that leaves you at high risk of dying and losing it on the way there.

Another successful meeting of the Committee to Prevent Regulatory Capture

Never forget Zuckerberg is the guy who found out 10% of revenue on his platform is from adverts for scams (enabling ~$50bn/year of theft from Americans), and chose not to do anything because it would hurt profits too much. Not the ideal source of advice on how much to regulate AI.

David Sacks @DavidSacks

2 weeks ago

Mark Zuckerberg: “I don’t think that we need some kind of industry-wide coordination... I think that just each lab needs to take the time, and when it sees that there are issues, you just take the time that you need internally to make sure that you’re proceeding safely… “we’re

303 429 5K 831K 708

openai wrote this back in 2016 Authors - dario, sam, ilya openai.com/index/special-…

I'd bet it's not irregular. This smells more like they were evaluating a research task not a cyber one. I guess we'll just have to wait and hope more details get properly released. If they do I really encourage you to go back and see whether any of the assumptions you're basing your reaction on right now turn out to be wrong. i.e. go back and see if your current I impression that that the models didn't do anything but look at a public file and didn't do it in an especially concerning manner holds up once we get a full report, go back and see if this was openai internal or irregular etc

@grayhamandeggs @Afinetheorem What do you mean contractors? This happened during internal evaluations performed by openai itself right?

@SteamboatLion @Afinetheorem Yeah they should. And victims of crime should improve their home security. But that doesn't mean criminals should be off the hook. And ideally the person who gets paid to help mitigate the problem is different to who creates the problem in the first place

@grayhamandeggs @Afinetheorem So you think this particular activity is fine, but in general if an agent does something slightly more invasive (e.g. endpoint fuzzing) at that point there should be penalties for on the model creator's?

@Afinetheorem If someone did succeed, should the attacker face any legal penalty for this? Or it'd still be your fault for not trying even harder? Do you really trust that your pen testers are as thorough+relentless as a thousand strong agent swarm based on an unreleased fronteir model?

I think, upon recognising how powerful and potentially dangerous the tech is becoming, Australia's response should be: - do everything we can to increase the proportion of AI compute and world class talent located on Australian soil. We need this for economic and national security, and if you're worried about losing access to frontier models - data centers on our soil gives us the bargaining chip we need to prevent this. If we don't do this, default outcome is USA will eventually export control all the best models as their government wakes up to the strategic importance of AI. Don't let comparatively irrelevant concerns like community input or IP rights of artists jeopardize this national security imperative. - using the leverage we gain thanks to being a nation that holds substantial compute resources to influence *global* governance of frontier AI, ideally helping establish international agreements for managing AI risk that include USA and China