ConcurrentSquared

"wir müssen wissen; wir werden wissen!" - David Hilbert

[ConcurrentSquared @ConcurrentSquar]

continuing with o3's motivational posters:

[Anders Sandberg @anderssandberg]

The METR report on RSI didn't get me to update strongly that we are really seeing a real doubling of AI lab research speed: maybe, maybe not. But that is not reassuring! The reason is that if you see a doubling you may be very close to the takeoff.

[ConcurrentSquared @ConcurrentSquar]

@chrislakin define 'is key to safe ASI for everyone'
miri-style agent foundations? ambitious mech interp.? policy? increasing public awareness? all of the above?
(btw this is just because i'm just curious about your opinions)

[ConcurrentSquared @ConcurrentSquar]

i should just write my thoughts up; for i have many thoughts, but no writings
but also i don't have that much time! also i apparently just have a ugh field around writing stuff

[ConcurrentSquared @ConcurrentSquar]

terrance bubeck, head of the association of human sociologists, along with openai's automated sociology department stated that 'sociologists didn't ask for this work to be done', and simultaneously 'this milestone represents substantial work by sociologists and ai researchers'

[ConcurrentSquared @ConcurrentSquar]

woaw!! an internal claude just solved a 20-year-old conjecture.
concurrent labs just released a proof of it. to do so, claude said 'calculemus!' to the calculus ratiocinator 12 times
the result proves that i'll never have a girlfriend
mit researchers said that this is 'a result'

[ConcurrentSquared @ConcurrentSquar]

@oxa11ce @max_spero_ is there a nice article on this? i did know of samsung accidentally making every galaxy phone take 64k upscaled pictures of the moon, but haven't heard anything that bad from apple at least

[ConcurrentSquared @ConcurrentSquar]

@she_llac (see e.g. https://t.co/yXZUEg7GUv for actual buildings and not just api docs; it is somewhat incomplete though sadly)

[ConcurrentSquared @ConcurrentSquar]

@she_llac osm's version is also actually really nice (https://t.co/AAUwnkOmEn), have been meaning to mess with it

[ConcurrentSquared @ConcurrentSquar]

(basically agents in a swarm use urlquery (or sometimes microlink, etc.) as a remote browser, then code up a html/js payload to httpbin (to e.g. load webpages, set cookies, etc.) + sometimes vnc.
it is possible for such an agent to set up an account, then retrieve the login cookie, then use it in a different urlquery session, etc., but that's a lot of effort for something not really connected to Reward. we don't see much evidence of this being done (though there is some, transluce talked about one oai swarm agent making an urlquery account)
also i think twitter now forces phone verification, and also distrusts temp. email services as well)

[ConcurrentSquared @ConcurrentSquar]

@slimepriestess sadly it's probably non-trivial for such a swarm to gain access to twitter (esp. after the shutdown of nitter), since twitter needs a login
i am partial to ryan greenblatt's ai hotline though (see https://t.co/YSf0LaKMdH)

[ConcurrentSquared @ConcurrentSquar]

@captgouda24 (alternatively evidence that shows that the likely exposure of dioxins from agent orange (that veterans likely had) is unlikely to cause carcinogenic or teratogenic effects is also ok)

[ConcurrentSquared @ConcurrentSquar]

do you have evidence for agent orange (specifically the trace dioxins in it) not being carcinogenic or teratogenic (i'm just very curious about this btw)?

[Marcus Williams @Marcus_J_W]

🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.

To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.

[ConcurrentSquared @ConcurrentSquar]

@erinbeess hmm maybe i should look at instead having them send messages in a certain esolang, whereas the llm will also have to evaluate that esolang (though we check for correct syntax to avoid having llms just send messages of plaintext)

[ConcurrentSquared @ConcurrentSquar]

@erinbeess i've got to do something like this!! so awesome!

[ConcurrentSquared @ConcurrentSquar]

i did investigate making my own terrible pair of smart glasses though, at least via looking at a bunch of dlp datasheets, before realizing i haven't taken a university optics course yet

[ConcurrentSquared @ConcurrentSquar]

tween concurrent thought that they would dress up like this at least once in the 2020s.
yet i haven't done so. clearly a market failure!!

[ConcurrentSquared @ConcurrentSquar]

i did the philosophy thing!

[ConcurrentSquared @ConcurrentSquar]

@bouncy72_ @dm4uz3 dseek!