Daniel Janes

[Daniel Janes @dtjanes]

@jon_stokes @brianchau57 that were supposed to be unsolvable - hallucinations, prompt injections, LLMs not knowing where to stop. They're all getting better though. The main thing keeping me on the doomer train is just that I've been burned betting against the labs one too many times. 4/

[Daniel Janes @dtjanes]

@jon_stokes @brianchau57 mentioned them in a post. That ignorance is a strike against me, but for what it's worth, covering those would be very helpful.

For all that, though, the main strength of doomerism is that the labs really have powered through a lot. There were so many things 3/

[Daniel Janes @dtjanes]

@germansdiaz @GregoryConti19 What sort of data would you like to see? AI safety relies a lot on theory, it's true, but I think that's kind of an unfair knock when the whole argument is over the emergence of a yet-to-be-seen phenomenon. Consistent capabilities advancement counts as data, no?

[Daniel Janes @dtjanes]

@ChombaBupe Has there ever been "technical or physical backing" for the AI abilities before it is demonstrated to have them? The discourse is littered with predictions of things AI could never do and subsequently has done. How can you be so confident?

[Daniel Janes @dtjanes]

@JasonBakerJB @hlntnr Eventually all of the evals we are capable of making will saturate. At that point "better" will be undefined. Will we just make bigger and bigger models anyway? Let the RSI rip? What does "winning" this actually look like? /2

[Daniel Janes @dtjanes]

@JasonBakerJB @hlntnr What does a win condition look like for this? China isn't far behind us and we clearly don't have any kind of durable moat - unless you're willing to use an ASI to launch a crippling strike they'll just get right there in a few months. /1

[Daniel Janes @dtjanes]

To everybody who thinks that AI biorisk is no problem because wetlab access and materials will be hard to come by: apparently we're going to have Biolabs-as-a-Service. Frankly I would like to ban this, but at a minimum we need laws requiring whitelisting and licensing.

[Phil Jacobson @philjacobson]

Spent some time digging into Ginkgo Bioworks ($DNA) over the last few weeks.

I think Ginkgo is very well positioned to be the autonomous labs platform for the AI revolution in bio.

Captured my thinking in this thesis and thought I'd share it publicly. https://t.co/Kmr2KZH3wV

[Daniel Janes @dtjanes]

@NateWitkin You make an excellent point in the last paragraph, but it is completely at odds with a world where AI contracts the share of available employment and therefore status. We are going to have to restore volunteering, hobbyism, and other third-place organizations as status ladders.

[Daniel Janes @dtjanes]

@christopherrufo True but unfortunately catastrophic. Prudent AI policy is going to be massively set back by the negative polarization.

[Daniel Janes @dtjanes]

@RobertMSterling Why are you clowning on the Europeans, we're all ending up in the permanent underclass in this scenario.

[Daniel Janes @dtjanes]

@MaciejGorgol @RobertMSterling @Bjornsund Nope. Even if we get RSI first, the Chinese will probably get it within 3-6 months like they get everything else. This is a fast-follower technology which means you can't build enduring advantages with it, unless you commit to a total war.

[Daniel Janes @dtjanes]

@ramez In the short term, sure. But over the long term? Not at all. AI facilitates every step of black-market hazardous materials acquisition and deployment for all kinds of threats, to say nothing of the blanket social dislocations caused by massive job loss and redundancy.

[Daniel Janes @dtjanes]

@HadleyMcintosh Most anxiety about open-weight models centers around bio-risks and its logistical assistance for acts of terror and nihilistic violence. Cyber is virtually the only domain where open-weights might help and only if you give it write access to basically the whole system.

[Senator Todd Young @SenToddYoung]

I am calling for the National Security Council to hold regular meetings on the emerging national security threats posed by advanced AI systems.

America must continue to lead the world in AI while staying ahead of the security challenges posed by this rapidly advancing technology.

[Daniel Janes @dtjanes]

@GWHayduke97 Important to note that this effect only goes in the direction of attack. An AI can't tell you that there will be a school shooter without surveillance. It can't stop pandemics without societal-scale equipment deployment. There can be no "good guy with a gun" outside maybe cyber.

[Hayduke ⏹️ @GWHayduke97]

Aside from x-risk, the 'alignment' problem, and the fact that the 'best case' of solving alignment is a Wall-E dystopia, one of the biggest issues with increasingly powerful AI is that it helps give every individual on earth the ability to do things that previously not everyone could do. It's a moral leverage extender, and there's a very real minority of extremely maladjusted people who would shoot up a school or create a pandemic if they had the means to do so, but didn't have those means before AI.

[Daniel Janes @dtjanes]

@theammind @RedemTheTimes This is a fast-follower race, there's nothing to win at the end of it unless you're willing to start WWIII to cement your gains. And "winning", especially by proliferating open-weight models, is just creating a rival power to your own. These narratives are poorly thought out.