Wocean Man (@dot_products) header
Wocean Man (@dot_products) profile picture

It's a fiasco bratan

New Zealand
Joined June 2026
Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)Photo by Wocean Man (@dot_products)
📁 gatinho miando alto com efeito de explosão
43
3,106
88
20,195
239,447
chabibiti where should I put the hospital
> one of the most hurricane prone places on the planet > puts hospital on an island with one road
30
i bet it feels good as hell to eat a whole pumpkin as a bison
25
348
30
3,734
56,978
The greatest rightist minds are using AI to make theo von say “nigga it’s over.” Constant dooming. The greatest minds are using AI to conjure a promethean age of progress and abundance of knowledge. We’re going to know things we didn’t even know we wanted to know.
1
57
The sick freak I am, I was hoping for RH to be false and for the counterexample to be just some ordinary number with no explanation.
1
38
Beast mode
BREAKING: Jensen Huang reportedly threatened to forcibly remove an Nvidia employee’s skullcap, exposing their brain, and then urinate on it. — NYT
10
do you ever feel like a frenchie eating corn on the cob loud as hell
10
329
13
3,532
73,247
sometimes i trick this sea lion with the cup game and he scream loud as fuuuuuck
5
136
6
1,719
36,979
Wocean Man retweeted
It will be alright
44
648
45
7,287
124,701
Where were you when the Riemann Hypothesis was proven true?
1
38
8 is a really awful number if you think about it.
1
10
Using a variational calculus of computations to find new algorithms like a boss.
3
68
𝗥𝘂𝘀𝘁 𝘃𝘀 𝗖# We see a lot of rewrites to Rust recently. So, let's see how they are different. Both are running native machine code, but they get there at very different times. 𝗥𝘂𝘀𝘁 (𝗿𝘂𝘀𝘁𝗰 + 𝗟𝗟𝗩𝗠): 🔹 Source code → rustc → LLVM → linker → one native binary (app / app.exe). 🔹 Cargo downloads crates from crates.io and compiles them from source. 🔹 The 𝗯𝗼𝗿𝗿𝗼𝘄 𝗰𝗵𝗲𝗰𝗸𝗲𝗿 statically analyzes ownership and lifetimes before any machine code exists. 🔹 std and other crates are statically linked into the binary by default. libc / ucrt are typically linked dynamically. 🔹 The executable doesn't need a runtime or JIT. The OS loads it into memory and main() starts executing. 🔹 Memory is freed deterministically by drop() glue code the compiler inserts when the owner goes out of scope, so there is 𝗻𝗼 𝗴𝗮𝗿𝗯𝗮𝗴𝗲 𝗰𝗼𝗹𝗹𝗲𝗰𝘁𝗼𝗿. 𝗖# (.𝗡𝗘𝗧 𝗥𝘂𝗻𝘁𝗶𝗺𝗲): 🔸 Source code → Roslyn → IL → straight into the dumpster. Unfortunately, someone usually retrieves it. 🔸 NuGet packages are individually wrapped bags of garbage, carefully versioned so you can reproduce the exact same smell across environments. 🔸 The CLR is an industrial waste-processing facility that takes your garbage, loads it into memory, and pretends it's software. 🔸 Methods are JIT-compiled because apparently even the CPU refuses to touch C# until the last possible moment. Tier 0 is fresh garbage. Tier 1 is optimized garbage. Dynamic PGO studies which garbage you use most frequently so it can serve it faster. 🔸 The 𝗚𝗖 is the only honest component in the entire ecosystem. Even the runtime acknowledges that everything you're creating is trash and periodically sends a dedicated worker to clean it up. 🔸 Native AOT lets you compile the entire landfill ahead of time, producing a standalone executable so you can distribute your garbage without requiring everyone to install the garbage-processing facility. ⚖️ The trade-off: • In Rust you answer "who's going to release this memory?" 𝗮𝘁 𝗰𝗼𝗺𝗽𝗶𝗹𝗲 𝘁𝗶𝗺𝗲. The compiler can be strict, and you get a binary with no warm-up and no GC pauses. • In C# the garbage is so deeply embedded in the language that Microsoft had to build an entire runtime whose primary responsibility is periodically taking out the trash. And somehow, after decades of engineering, the garbage still needs warming up. Back to you, which one would you choose for your next service, and why?
7
4
2
51
2,678
Full batch Adam on a quadratic settles into an exact w -> -w 2-cycle. Whether it's stable was left open (Bock & Weiss). Closed form: at eps=0, stable iff b2 < (1-3b1)/(b1(3-b1)), so never for b1 >= 1/3. At PyTorch defaults it's stable only in a 0.28% LR window.
1
1
48
AdamW fits the same stability test: replace b1 with b1(1-lr*wd) and add a factor of (2-lr*wd). Realistic weight decay barely moves it: at PyTorch defaults with lr*wd = 1e-3, the 0.28% window only grows to 0.29%.
1
24
Oral dose that keeps a drug above MIC for exactly T hours (1 compartment, 1st order absorption) has a closed form: r=ka/ke, s=exp(-ke*T), u=((1-s)/(1-s^r))^(1/(r-1)) Dose = V*(ka-ke)*MIC / (F*ka*(u-u^r)) Drug crosses MIC at t=-ln(u)/ke, stays above until +T.
2
85
The ITAE of a 2nd order step response has an exact closed form. With s=sqrt(1-z^2), p=acos(z), q=exp(-pi*z/s): ITAE*w^2 = 3z^2-s^2 + (2*exp(z*p/s)/s)*(pi*q/(1-q)^2 + (2*z*s-p)*q/(1-q)) Optimal z = 0.752444..., not the textbook 0.7
1
32
man, who should i trust? a bunch of nerds who made linux run on apple silicon, reverse engineering that pesky gpu isa among a gazillion other things, all pre-ai. or the guy who thinks having your agent rewrite shitty rust he never looked at in x86_64 is excellence?
133
352
23
6,565
287,386
Amazing stuff. Will have to express the paper's k-SAT pipeline in Stairwell and see if it can check or improve the cost.
1
1
127
Checks on small cases. Not proof, but it checks out within what is feasibly computable.
1
42
All my exams done for the year. Real analysis was a bitch.
3
135