Sign in

nic 🔹

@nmk.wtf
480 followers 471 following 1.3K posts

mathematician, quant intern, reader of things nmk.wtf I give 10% of my income to GiveWell and so should you

PostsRepliesMedia
nic 🔹 @nmk.wtf · 20/09/2026
Also worth noting that the post argues against the trade with ants analogy.
120
nic 🔹 @nmk.wtf · 17/09/2026
It still does!! It is just a big group, and the malaria people dont get the most attention. Also USAID cuts hurt the global health cause area
160
nic 🔹 @nmk.wtf · 29/08/2026
nice
010
nic 🔹 @nmk.wtf · 19/08/2026
011
nic 🔹 @nmk.wtf · 05/08/2026
000
nic 🔹 @nmk.wtf · 21/07/2026
How GPT found the polynomial
091
nic 🔹 @nmk.wtf · 19/07/2026
Thanks claude
010
nic 🔹 @nmk.wtf · 29/06/2026
These are pretty fun
000
nic 🔹 @nmk.wtf · 29/06/2026
My podcast listening in 2026 so far. I apologize in advance
120
nic 🔹 @nmk.wtf · 29/06/2026
Claude has had enough of my shit
1401
nic 🔹 @nmk.wtf · 18/06/2026
You can also track rivalries in my new game. Right now I am my greatest enemy
100
nic 🔹 @nmk.wtf · 17/06/2026
I added a history of blue v red for my new game: at-trenchline.fly.dev
000
nic 🔹 @nmk.wtf · 17/06/2026
smh
020
nic 🔹 @nmk.wtf · 17/06/2026
I made an at proto game at-trenchline.fly.dev
021
nic 🔹 @nmk.wtf · 16/06/2026
lmao
From Money Stuff:
We talk from time to time around here about my favorite postmodern comedians, the people who design user interfaces for Citigroup Inc.’s internal software. If you work at Citi, you come in every day and open up some proprietary software to do trades and send money and stuff, and that software has a bunch of fields you can fill in and buttons you can press to do things, and some of them are pranks. Like there was allegedly a field in some payments screen that “came pre-populated with 15 zeros, which the person inputting a transaction needed to delete”; if you forgot, apparently typing “$1” in that field would send the customer $1,000,000,000,000,000? Just hilarious Dada-esque comedy.

Citi appears to do this stuff out of genuine commitment to the bit, but pitfalls like this exist at every bank. Every bank has software with fields that you can fill in and buttons you can press, and you are tired and distracted and you might, say, put the dollar amount of your trade in the “shares” field and sell 7,548 times as much stock as you intended to sell. That really could happen anywhere, though in fact it happened at Citi.The next paragraph:
More generally, it is the nature of software that any piece of software will contain a button you should not push, just to make life exciting. (For me it is the “Publish” button in this column’s content management system. Have I ever published a half-written draft by accident? No. Do I think about it every day? Yes. Have I jinxed myself by writing this? Of course.) Probably most banks do not give their employees software with a big red button labeled “LIQUIDATE BANK,” which when pressed initiates a liquidation of the bank, though Citi might. (No, I’m kidding, at Citi the liquidation is probably triggered by a small green button labeled “Help.”)
1170
nic 🔹 @nmk.wtf · 14/06/2026
020
nic 🔹 @nmk.wtf · 13/06/2026
While its a good meme, I do think its important to point out that anthropic has a more precise policy position that this.
1130
nic 🔹 @nmk.wtf · 10/06/2026
I maybe found a cool new optimizer
100
nic 🔹 @nmk.wtf · 09/06/2026
"My undergraduates can not only prove √2 and √3 are irrational but also that √4 is irrational."
120
nic 🔹 @nmk.wtf · 08/06/2026
In this model the cars are allowed only to dodge against the flow of the other traffic.
120
nic 🔹 @nmk.wtf · 08/06/2026
If you also allow occasional sideways dodges you get much smoother shapes
130
nic 🔹 @nmk.wtf · 08/06/2026
Switching between the deterministic and randomized update rule is extremely mesmerizing!
172
nic 🔹 @nmk.wtf · 08/06/2026
Woah, just played around with both, and it turns out their behavior is shockingly different!
120
nic 🔹 @nmk.wtf · 08/06/2026
You should add the randomized update rule!
Wikipedia:
Randomization
A randomized variant of the BML traffic model, called BML-R, was studied in 2010.[9] Under periodic boundaries, instead of updating all cars of the same colour at once during each step, the randomized model performs 
L
2
{\displaystyle L^{2}} updates (where 
L
{\displaystyle L} is the side length of the presumably square lattice): each time, a random cell is selected and, if it contains a car, it is moved to the next cell if possible. In this case, the intermediate state observed in the usual BML traffic model does not exist, due to the non-deterministic nature of the randomized model; instead the transition from the jammed phase to the free flowing phase is sharp.
140
nic 🔹 @nmk.wtf · 08/06/2026
This is so cool!
150
nic 🔹 @nmk.wtf · 07/06/2026
Unlikely showing up in unlikely events again
010
nic 🔹 @nmk.wtf · 01/06/2026
Okay now explain this intuitively
In quantum mechanics, spin is an intrinsic property of all elementary particles. All known fermions, the particles that constitute ordinary matter, have a spin of ⁠
1
/
2
⁠.[1][2][3] The spin number describes how many symmetrical facets a particle has in one full rotation; a spin of ⁠
1
/
2
⁠ means that the particle must be rotated by two full turns (through 720°) before it has the same configuration as when it started.
110
nic 🔹 @nmk.wtf · 28/05/2026
It is also not just underpowered models. I trained Neural Networks and they exhibit the same failure mode.
120
nic 🔹 @nmk.wtf · 28/05/2026
This is also robust to the training period of the data. (Note that with the larger training period we don't have House YoY, so you will notice that mortgage becomes more important)
120
nic 🔹 @nmk.wtf · 28/05/2026
So what changes? First, and most striking, the baseline vibe is just way lower. This could be some macroeconomic factor I haven't yet considered. But seems unlikely. Next, people seemingly care a lot more about unemployment and change in home prices.
120
nic 🔹 @nmk.wtf · 28/05/2026
This is robust to leaving out each of the individual predictor variables.
A 2x3 grid of time-series line charts titled "R2 - Leave-one-predictor-out (train 2000-2019)," comparing actual versus predicted values from 2010 to 2026. Each panel shows a black solid "Actual" line and a blue dashed "Predicted" line, with a green-shaded training region before 2020 and an orange-shaded post-2020 test region split by a red dotted vertical line. The six panels show a full base model and five ablations, each dropping one predictor: full base set (R2=0.81), drop cpi_yoy (0.70), drop unemployment (0.49), drop gas_yoy (0.74), drop home_price_yoy (0.73), and drop mortgage_rate (0.80). In every panel the model tracks actual values closely during training but diverges after 2020, with predictions running persistently above the declining actual series. Post-2021 gap annotations range from about -13 to -29, largest when cpi_yoy or gas_yoy is dropped, indicating those predictors matter most for post-2020 fit.
110
nic 🔹 @nmk.wtf · 28/05/2026
Recreating some vibecession research. You can clearly see a structural change after covid.
140
nic 🔹 @nmk.wtf · 26/05/2026
Built a little app with claude that can optimize polyomino arrangements. Here each piece wants to touch 3 of the other type.
141
nic 🔹 @nmk.wtf · 26/05/2026
i'm doing my part
010
nic 🔹 @nmk.wtf · 14/05/2026
I figured it out. Just adding a line like "which blocks which" or something would probably be enough. Anyways, very pretty!
010
nic 🔹 @nmk.wtf · 14/05/2026
Or do I want this arrangement?
010
nic 🔹 @nmk.wtf · 14/05/2026
Is this the "only interact with self" setup?
210
nic 🔹 @nmk.wtf · 12/05/2026
Screenshot of a tweet by user "rain" (@__ghostfail) posted 18 hours ago, reading: "taking 35 tabs of acid and listening to this on loop every night until im claude". The tweet quote-retweets an Anthropic (@AnthropicAI) post from 22 hours earlier announcing that Claude's Constitution is now an audiobook read by two of its authors, Amanda Askell and Joe Carlsmith, including a Q&A on the writing process and the philosophies that shaped the document. The quoted post embeds a video (paused at 0:38) showing a title card that reads "Claude's Constitution" in orange, with the partial quote: "Models are learning about their past selves on the internet. It's kind of a fascinating and very hard situation for them to be in, actually," with "be in, actually," highlighted in black. The tweet has 12 replies, 40 reposts, 761 likes, and 34K views.
0120
nic 🔹 @nmk.wtf · 12/05/2026
A different view.
010
nic 🔹 @nmk.wtf · 12/05/2026
I had claude pull from yahoo to calculate this for a bunch of different indices. The Chinese ones are the only ones where intra-day has higher returns.
100
nic 🔹 @nmk.wtf · 12/05/2026
This trend holds for other indices as well, including in Europe and Asia (except for China for some reason).
000
nic 🔹 @nmk.wtf · 12/05/2026
This is true of all stocks!
380
nic 🔹 @nmk.wtf · 12/05/2026
So instead we can ask what is the effect of inequality on the happiness relative to your neighbors. Here we do see a persistent negative coefficient. Even if you only compare countries to themselves in the past and future you keep seeing this effect.
100
nic 🔹 @nmk.wtf · 12/05/2026
You can just use CC to do economics research. I wanted to look at the effects of inequality on happiness. If you do a naive regression you don't find any effects. This is because happiness has large regional components (things like culture).
100
nic 🔹 @nmk.wtf · 12/05/2026
I need to post more
150
nic 🔹 @nmk.wtf · 05/05/2026
context
Jarred 7 hours ago | unvote | next [–]

I work on Bun and this is my branch
This whole thread is an overreaction. 302 comments about code that does not work. We haven’t committed to rewriting. There’s a very high chance all this code gets thrown out completely.
I’m curious to see what a working version of this looks, what it feels like, how it performs and if/how hard it’d be to get it to pass Bun’s test suite and be maintainable. I’d like to be able to compare a viable Rust version and a Zig version side by side.
reply
120
nic 🔹 @nmk.wtf · 05/05/2026
Crazy, it seems just projecting weights on a hypersphere improves training? arxiv.org/pdf/2509.25206
Screenshot of a tweet by Keller Jordan (@kellerjordan0) posted May 1. Text reads: "New modded-NanoGPT optimization benchmark result: @wen_kaiyue has improved upon both the Muon and AdamW baselines, by replacing their weight decay with hyperball optimization. The new record is 3325 steps." Below the text is a line chart titled "Modded-NanoGPT Optimization Benchmark as of 2026/04/30" plotting validation loss (y-axis, 3.2 to 4.0) against training step (x-axis, 0 to 6000) for four optimizers, with a dashed horizontal target line at 3.28. Best step counts to reach the target: Muon 3500 (orange), AdamW 5625 (blue), MuonH 3325 (green), AdamH 4875 (purple). MuonH reaches the target first; AdamH outperforms AdamW. Tweet has 7 replies, 48 reposts, 421 likes, 55K views.
000
nic 🔹 @nmk.wtf · 02/05/2026
¯\_(ツ)_/¯
010
nic 🔹 @nmk.wtf · 29/04/2026
It is still a little janky, but now I have reasonable bots you can play against: oathbroken.online
000
nic 🔹 @nmk.wtf · 28/04/2026
You'll never guess what event this is
460