matthew

city enjoyer

[matthew @spotlemma]

@YouJiacheng @thsottiaux bruh

[matthew @spotlemma]

@Bensam123TV @_coenen I mean plenty of people do have doubts about the nature of our reality. It’s just that an immersive stream of rich sensory inputs is much more convincing than a reality perceived through the body of an HTTP response

[matthew @spotlemma]

@nickcammarata A pause (ban) on RLVR would probably suffice

[matthew @spotlemma]

Climate change invariably comes up when I describe AI risk to people and it’s been pretty surprising to me that people a) believe that it’s an existential threat and b) can use that framing to downplay other risk vectors as merely incremental

[roon @tszzl]

@ryan__jiang I think climate change does not pose any % existential risk to humanity

[matthew @spotlemma]

@ChadNotChud @LinkofSunshine A Lean certificate would be almost useless if it didn’t introduce vast amounts of new scaffolding/machinery and insights for proving lower bounds

[matthew @spotlemma]

We’d be better off with our colloquial understanding of model training by shifting the default framing from “what’s in the training data” to “what is the model rewarded for?” Much more relevant for present capabilities discussions

[Dario Soatto @soatto]

"That framing is half-right, and the half that's off matters"

Where in the training data is this abominable sequence of tokens and why is it weighted like 100000x

[matthew @spotlemma]

@JHWeissmann I don’t think that a desire for self-preservation is required to end humanity

[matthew @spotlemma]

@crulge I don’t think this is an accurate understanding of their viewpoint. Anthropic does not really care about economic outcomes except insofar as they support its mission

[matthew @spotlemma]

@firstadopter I don’t think this is unbridled honesty for its own sake. Anthropic employees likely do mostly believe that making plain the stakes is good if it increases the likelihood of regulation and/or coordinated slowdown

[matthew @spotlemma]

@ChaseBrowe32432 The HackerNews thread is hardly any better!

https://t.co/WJCCF1tKXM

[matthew @spotlemma]

@marsxiang_ except that we are so so far from P NP. Likely the last Clay problem to be resolved

[matthew @spotlemma]

@theojaffee Somewhat comforting to see the limitations of Astra's forecasting calibration, having recently observed some disconcerting p(doom) estimates from it

[matthew @spotlemma]

@souljagoyteller @hecubian_devil This is a pretty easy one for him to dismiss as “millions of dollars of compute thrown at brute-force search”. Especially given that it seemingly required extensive human-guided breakthroughs as precursor

[matthew @spotlemma]

The Navier Stokes drama is particularly alluring right now because it’s the first lab scandal in what feels like forever (weeks) that doesn’t really affect doom estimates (apart from its bearing on the cooperative judgment and integrity of the parties involved)

[matthew @spotlemma]

@Duderichy Boogie nights, y tu mama tambien, city of god

[matthew @spotlemma]

@unixpickle Not particularly imaginative given I wouldn’t be surprised to see any of these tomorrow

[matthew @spotlemma]

@crulge Still largely correct read of the way many people (among a certain cohort) do travel these days though and what drives the trends towards certain destinations

[matthew @spotlemma]

@yoavgo Isn’t it also the case that emergent swarm behaviors are much more difficult to detect, fight, shutdown if they’re implicitly orchestrated by diffuse information networks, able to leverage unbounded decentralized compute, than one run by a single actor?

[matthew @spotlemma]

@yoavgo One reason could be about privilege of the model’s environment. GPU clusters are highly sensitive and emergent behavior when lab employees use the models to run experiments is a bigger concern than a rogue employee orchestrating malicious access