Dan Romero on Warpcast

Content pfp

0 reply

0 recast

0 reaction

Dan Romero pfp

In you're in favor of "AI safety" (broad definition), what's your most compelling cast-length argument?

38 replies

10 recasts

54 reactions

tldr (tim reilly) pfp

tldr (tim reilly)

The technology is obviously epochal, but its path still largely unknown. It’s appropriate to have at least *some* humility toward “known unknowns” of a historical magnitude.

1 reply

0 recast

8 reactions

Jonny Mack pfp

i’m in the all-gas-no-brakes camp but will do my best to steelman: 1. we’re on the cusp of agi 2. agi is like nuclear: extinction-level threat, global superpower-level leverage, and potential civilization expanding resource 3. we dont really know how it works 4. thus, we should proceed with *extreme* caution

4 replies

0 recast

2 reactions

six pfp

if toddlers were able to make an adult human from scratch would they be able to control it?

1 reply

0 recast

1 reaction

Ben - [C/x] pfp

At fist automanufacturers made similar arguments against requiring seatbelts in cars. Imagine what we’d be saying about our great/grandparents if that line of thinking had prevailed.

2 replies

0 recast

1 reaction

petar.xyz pfp

I think all new versions of AI should be tested in some sort of a focus group before releasing them to the public.

2 replies

0 recast

0 reaction

Ivy pfp

a machine looking at the right set of KPI's for humans would reach unfavourable conclusions on what should be done with us

1 reply

0 recast

0 reaction

grin pfp

AI safety is a nuanced discussion that’s very hard to have in cast-length format but if you must: the appropriate balance of novelty and safety is generally good, and denying the need for either is ridiculous

0 reply

0 recast

3 reactions

GIGAMΞSH pfp

I'm in the Sam Harris camp. Intelligence is the critical ingredient for all technology so we should treat AGI like summoning a god. But I'm less concerned than I used to be because I think the safety path is becoming clear: everyone on earth participating in continuous RLHF of the biggest models.

1 reply

0 recast

3 reactions

Joe Blau 🎩 pfp

Ilya put it best. Humans love animals, but when we want to build a road between two cities, we don’t ask the animals for permission. AGI will treat us the same way.

1 reply

0 recast

2 reactions

keccers pfp

p(doom) and all of that is a distraction when I think of ai safety I mostly think of things like + how to appropriately compensate people whose copyrighted work was scraped for training data w/out consent + a shared standard for proof of humanity + how to allow ppl to opt out of being training data in future

0 reply

0 recast

2 reactions

wizard not parzival pfp

wizard not parzival

safety != halt // should the government have privy information or should we release everything worldwide tonight

1 reply

0 recast

2 reactions

BC 🍈 pfp

Naming your next iteration "Q" anything in 2023 shows a worrisome lack of foresight and intelligence.

0 reply

0 recast

1 reaction

Connor McCormick ☀️ pfp

Connor McCormick ☀️

Ah shoot I missed this! Do you want the most compelling argument about the needs for AI Safety or do you want the most compelling argument about the policies to achieve it?

0 reply

0 recast

1 reaction

fredrik pfp

mitigation via give us more time to understand the existential risk possibility for humanity as a whole a small no of tech bros shouldn't be allowed to creat existential risk for human life as we know it on earth

0 reply

0 recast

1 reaction

Gregor pfp

Incentivize transparency & accountability before gov steps in with club foot. IEs: user passes a short exam to be able to use advanced models. If a prompt tips red flag, user relinquishes privacy in order to continue their task. Opensource safety tools Subsidies for startups building ai safety tools

0 reply

0 recast

1 reaction

Sam Iglesias pfp

We don’t know what perverse path it might take to maximize its objective function, nor do we have a handle on how to craft its objective function to balance beneficence, human autonomy, and non-maleficence. We sort of see this with recommendation algos already.

1 reply

0 recast

1 reaction

Q🎩 pfp

Once someone can give me the exact process to 100% GUARANTEE a child is raised to NOT be a murderer, then I’ll start to think we might have ANY idea how to control another, possibly superior, intelligence.

0 reply

0 recast

1 reaction

Britt Kim pfp

i believe governments will use AI for subjugation. for the safety of a free citizenry, we need restrictions on government usage.

0 reply

0 recast

1 reaction

𒂠_𒍣𒅀_𒊑 pfp

𒂠_𒍣𒅀_𒊑

not the typical safety, but I hope for the best design and degree of public transparency & unconditional access. my opinion is that we should allocate more capital into specializing & downscaling the models, as well as formalizing & accelerating the valid data/performance/responsibilities of the global public infra.

0 reply

0 recast

1 reaction