penlu

human computer interface. naturally occurring feature of your environment

[penlu @penlume]

4397328654844826923795068102505872571721883526553349659561256924505973939597593482272505698004801207988043088656411102133523080581 divides RSA-260

[penlu @penlume]

@Free2Read4Me I guessed and checked 1-digit prime numbers by hand

[penlu @penlume]

2 divides 4

[HD Moore @hdmoore]

https://t.co/fgE40dHDiy

The private keys for these certs are literally the numbers 4, 56, 66, 115, 155, 156, 230, and 237 (life imitates XKCD, again).

[penlu @penlume]

@cudalescent it appears to have been a busy couple of weeks

[penlu @penlume]

tfw we must ensoul the microwave first so that it isn't instead ensouled by some malevolent other

[w̸͕͂͂a̷͔̗͐t̴̙͗e̵̬̔̕r̴̰̓̊m̵͙͖̓̽a̵̢̗̓͒r̸̲̽ķ̷͔́͝ @anthrupad]

I think of the Claude worms across all our Internet of things like the healthy gut microbiome of post post post modern civilization

[penlu @penlume]

@KamaDevGroup I think it requires some discipline, often implicit, to ensure that any nontrivial program remains possible to reason about. common programming patterns perhaps correspond vaguely to proof techniques

https://t.co/hFWDzrEsxl

[penlu @penlume]

kicking around the assertion that well-factored code is identically that which is easy to prove correct (even if not so proven). counterexamples?

[penlu @penlume]

add some code, the docstrings and tests get longer. remove some code, the docstrings and tests get longer. "please improve", the code, docstrings, and tests get longer. "please patch", all longer. "please deduplicate", somehow code still gets longer. what is going on

[penlu @penlume]

@KamaDevGroup this is a deep question. I think line count is a decent one-glance metric of complexity. yes there's a lot more to consider. code golf exists.

best I've got rn: path count and state size are roughly exponential in line count. reasoning about a large program is therefore hard.

[penlu @penlume]

@utsav_khairnar gradient descent ✅
handling data ✅
running good experiments ✅
kernel optimizations ✅
also CPU and other perf stuff ✅
frontend (for tooling) ✅
database wrangling ✅
distributed systems ✅
networking ✅
security ✅
OS stuff (mostly containers) ✅

[penlu @penlume]

every problem in computer science appears in ML infra

[zek @zekramu]

genuinely there are so many interesting problems in the ai infra world. it just so happens that now inference is trendy. but there are a ton of other ones. you aren’t bullish enough of infrastructure jobs.

[penlu @penlume]

@JuspreetS the stakes were a bit lower to be fair, but I believe this has happened. throwback to spin class

[penlu @penlume]

@DimitrisPapail variously bad analogy here but how does typical machinist pay compare to material and equipment costs

[penlu @penlume]

@VesselOfSpirit = 2^4 * 3^3 * 331 * 9356747 * 3688147577 * 2794660986650546064239 * 125737289904614103224893323541304305012411445433167991095076253. this cost 25 GPU-seconds

[penlu @penlume]

The Feeling of Power by Isaac Asimov looking more prescient than I'd initially expected also

[penlu @penlume]

@natolambert we saw a similar phenomenon for opus 5 and gave further details:

https://t.co/CtpmcjmG1d

[Cheng-Yuan (Sam) Lee @cl571128]

We've received several questions about the Opus 5 FrontierCode results, where scores decline as reasoning effort increases. In fact, the behavior is expected under the benchmark design. FrontierCode evaluates merge-ability rather than correctness alone, incorporating criteria that reflect user experience.

One such verifier is a scope criterion, which penalizes modifications to the codebase beyond what the task requires. We observe this effect across all frontier models we evaluated, though it is most pronounced in Opus 5: at higher reasoning efforts, the model shows a stronger tendency to refactor code unprompted.

[penlu @penlume]

I have already begun writing out 1024-bit primes to try. (really only need to fill in 1022 bits, since the first and last are always 1)

[JP Aumasson @veorq]

thanks to @penlume and @sweis we now know when RSA-2048 will be broken without a quantum computer

[penlu @penlume]

@sweis awesome!!!

[Steve Weis @sweis]

636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919
is a factor of RSA-896
https://t.co/yGWyEC8RpZ

[penlu @penlume]

@syzygetics @chobiboni2 "a fire upon the deep" opens with a slightly similar scenario

[penlu @penlume]

@olegakbarov @viemccoy or like less glibly: many persistent things are the result of some autocatalysis with exponentially growing rate, and every time/place the ingredients for this particular loop exist, there soon arises a mini version of it going there too

[penlu @penlume]

@olegakbarov @viemccoy the hand has many fingers