Elfarouk Harb 🌱 (@ElfaroukHarb) profile picture

Quant Researcher. Previously Software Engineer @Google, Quant Trader @citsecurities, and PhD in Theoretical Computer Science @UofIllinois

Chicago, IL
Joined September 2017
Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)Photo by Elfarouk Harb 🌱 (@ElfaroukHarb)
Elfarouk Harb 🌱 retweeted
Hedging nat gas on Kalshi is worse than futures in basically every possible way I can think of. Size: E-mini Nat Gas (QG) is 1 contract min, ~$8500 notional, and you only post a fraction of that as margin. Micro Henry Hub (MNG) is even smaller. Delivery: Both are cash settled. You’re not delivering/recieving tankers of gas… UI: Most futures brokers look pretty much like a stock app. I think even Robinhood lets you do it by just searching the ticker. Margin calls: The laundromat hedge here is long, so fund the full notional and you're fully collateralized and covered with no Margin calls. Fees: 2-3 bps vs Kalshi's absurd 100-500 bps fee depending on odds.
1
1
10
723
Anyone who’s dug into tax law (which is arguably easier to formalize than civil or criminal or immigration law) knows how insanely subjective it can be and how much room it leaves for interpretation and discretion. Like take the “simple” short term rental tax loophole. You can believe you’ve ticked every box in the wording of the law and still face an audit over whether your rental qualifies for the tax benefit. You can lay out the same facts “axiom” by “axiom”, and different auditors or judges can and have reached different conclusions. You can’t formalize that subjectivity away just because the system expects something resembling mathematical rigor in how you construct an argument.
OK, some mathematicians just need to shut the fuck up
2
9
2,075
This can't be stressed enough. Absolutely, and I mean absolutely nobody, has thought to train a foundation model on raw L3 market data. Every firm is still in the Stone Age, click trading and fitting linear regressions. This is truly revolutionary! /s
Today, @bayesiankalki and I are launching Mnemonic Labs (YC F26). We're a neolab training a foundation model for quant trading that learns directly from market activity. In <6 weeks we’ve built our training pipeline and model architecture. Our current training run uses over 25TB of tokens. We met 28 years ago shortly after I was born (we’re brothers!). We’re Olympiad finalists (5x USACO combined) and MIT graduates. Kalki made $20M in annual PnL at 22 at Tower Research and I was the youngest quant at Citadel at 19. Over the past ten years, we’ve been part of quant trading’s evolution from linear models to trees and now neural networks. Learn more about us at ycombinator.com/launches/XRK… or reach out at [email protected]!
12
4
435
33,765
I think we need to have a moratorium on hearing from economists on this issue until we solve the broader issue of what to do with that field
Ok, last thing, and I know it's unpopular, but on "grad students make this place run" and are "workers": when you become a professor, you will learn this is completely untrue. Love our students, but they are *students* and incredibly costly ones. 1/3
3
45
3
536
7,958
This definitely overcomplicates it. The intuition is actually quite simple. Let μ be the uniform measure on the deck 𝒟. A binary yes/no question corresponds to a set A ⊆ 𝒟, and what you learn from the answer of the question is the σ-algebra 𝒢 = {∅, A, Aᶜ, 𝒟}. Any 𝒢-measurable guess g: 𝒟 → 𝒟 is constant on the atoms (say g ≡ a on A and g ≡ b on Aᶜ). But we have, μ(g = id) = μ({a} ∩ A) + μ({b} ∩ Aᶜ) ≤ μ{a} + μ{b} = 2/52, with equality iff a ∈ A, b ∈ Aᶜ. The bound is independent of A. ∎ Hope this helps build intuition.
Replying to @LongFormMath
Many computed it correctly. Well done! Harder question: What is good intuition for why they are the same? Here’s my attempt: If you decided ahead of time that you were either going to guess the ace of spades or the ace of hearts, then both of your questions are equally good at determining which of these two you should guess. Moreover it doesn’t matter which two cards I pick ahead of time. All selections are equally likely to contain the card. Now, if the chosen card is not one of these two cards, I’m guaranteed to get the answer wrong. But what happens if it is one of those two cards? Then I’m guaranteed to get it correct! For both questions, I will get it wrong 50/52 of the time and correct 2/52 of the time. What basically happened is I divided all cards into two halves (red/black), and the question’s answer tells me which half to guess in. But I’m picking it not at random, I pick with my own chosen distribution (100% chance choose the ace of hearts/spades). Since all 26 cards in your half are equally likely to be the chosen card, you can pick your own distribution.
1
5
758
I burned my ChatGPT tokens so you don’t have to. stoc-2027-submission-atlas.e… ~75% of submissions report AI involvement in developing proofs or results. Some highlights: 1) ~91.2% disclose some AI use. ~11 explicitly deny AI, LLM or generative-AI use; another ~128 have no disclosure located. 2) ~74.7% report AI involvement in proof/result development. That includes supporting arguments, proof repairs, algorithms and constructions. In algorithmic game theory, that figure reaches ~85.9%. The model comparison, across all submissions: — ~61.1% report ChatGPT/GPT but no Claude Opus. — ~1.8% report Opus but no ChatGPT/GPT. — ~15.6% report both. — ~Gemini is a lot sparser (and mostly used in conjunction with other models) at ~6.5%. 3) The newcomer question might be the most interesting: ~730 of 2,988 author names, or roughly 24.4%, have no earlier indexed paper until end of 2025 in any of 253 screened theory/foundations conference/ workshop series. I will be very curious about the acceptance outcomes for newcomer papers. [Disclaimer: These numbers are best-effort estimates. They come from scraping Google Scholar and dblp and then reconciling author names, for example recognizing that “E. H.” is Elfarouk Harb. Anyone who has dealt with name matching knows it’s a cursed problem and very hard to get right. You can check the matches, the guesses, the supporting evidence and each author’s 2024/2025 publication activity on the site.] 4) Lots of overlapping claims/competing papers on the same problem: • 3 submissions claim the Matrix Spencer conjecture. • 3 claim overlapping refutations of Hyperclique hypotheses. • 2 improve Multiway Cut’s CKR integrality-gap lower bound. • 2 matroid-secretary submissions claim optimal 1/e guarantees. One of the matroid-secretary submission ALREADY merges three predecessor preprints... Final Disclaimer: this is AI-generated. Counts, classifications and identity matches can be wrong. Feel free to point your agent at the website/data and get it to refine the results for you.
STOC 2027 received 1665 submissions. In a bold new experiment, all submissions (and author identities) are immediately made public. This is the most immediate and revealing posting of any conference I am aware of. Take a look and see what you find! stoc2027.sigact-ec.workers.d…
3
10
2
44
9,966
The shamelessness of a lot of these profs is truly mind blowing.
Replying to @ElfaroukHarb
Of course students should personalize their emails. That's the whole point. Generic emails get ignored. But if you are really interested, you personalize your email yourself. Letting ChatGPT do it is a strong sign that you are not really interested. Costly signaling, man.
16
1,056
$115 for a PhD application btw. Sometimes the only reason to apply to a school is one professor. Gee, I wonder why students personalize these emails. Could it be that the app costs 115 fucking dollars? That they have no idea if the one person they want to work with in the school is even taking students? That any generic email gets ignored? Nope. Must be scheming students trying to con you.
my question is, if you think I'm dumb enough to not realize your personalized email was tailored by ChatGPT, do you really want me to be your PhD advisor?
12
7
3
511
62,798
The “HRT intern class is the new PayPal mafia” thing is so forced. It was a 3 week winter program in 2015. The PayPal mafia built a company together for years. If you want an actual mafia look at Tsinghua's Yao Class: Kimi, OpenAI, DeepSeek and xAI
1
3
204
Elfarouk Harb 🌱 retweeted
Son has developed a math proof: There are two ways to make an even number. (even + even, odd + odd). There has one way to make an odd number (odd + even). Therefore, there are fewer odd numbers than even numbers.
364
131
78
16,370
1,503,658
Bullshit. It's happened to me plenty of times in quant and tech interviews, and whenever I've said upfront that I'd seen the problem before, the interviewer just asked me to quickly sketch the solution to confirm I actually understood it. If I were interviewing someone who did that and couldn't even explain the core idea, I'd reject them on the spot.
most of the interviews at big tech is about luck
5
1
119
26,248
In my experience, being honest that you've seen the problem and then explaining the solution well almost always works in your favor. I've been in that situation four times and passed every one, despite not performing that strongly in two or three of them.
16
1,861
If I had to guess underlying tech, 3D Gaussian Splatting + High precision Lidar/Slam scanners —> feed it into a modeling agent/harness like VIGA/ChatGPT to reproduce references from the render —> Final touches by human modelers. Main issue is anyone who can afford the scanners can probably afford to do the rest on their own…
Moving Atoms (YC S26): Scan to Sim is here. Send us a LiDAR or SLAM scan of your factory floor, warehouse or yard, from any scanner. Upload multi-gigabyte scans straight from your browser (up to 200 GB per file). Within 48 hours you get back a simulation-ready digital twin of your real site. Every object is separated and editable, and accurate robot embodiments are placed right on your floor: autonomous forklifts, pallet AMRs, tuggers, tote robots and palletizing arms. Test the workflow before you build it. Try layouts, fleet sizes and routes, then watch the whole operation run sped up, before a single robot arrives on site.
214
Elfarouk Harb 🌱 retweeted
Let me try to state this differently - AI now performs the transitive closure of what we know (or very close to knowing). So, in many cases, it would feel like we were just a blip away from doing it ourselves. I have experienced this feeling several times by now.
> "Neither Fable 5.1 nor Opus 5.5 could refute the conjectures without the hint of our approach …. Now this tells me that our 2023 write-up was useful" Or it tells you that the models you used are weaker in Math. Here is ChatGPT Astra refuting all 3 conjectures in ~45 minutes given only Coppersmiths’ 1982 paper, and van den Brand, Nanongkai, and Saranurak's 2019 paper: chatgpt.com/share/6acee44d-d… The underlying impulse is very human, I understand why having the right idea on the shelf since 2023 and abandoning it is frustrating. But hinting that the models couldn't have done it without our "human insight" is just pure copium at this point and does no service for anyone.
4
4
38
4,654
> "Neither Fable 5.1 nor Opus 5.5 could refute the conjectures without the hint of our approach …. Now this tells me that our 2023 write-up was useful" Or it tells you that the models you used are weaker in Math. Here is ChatGPT Astra refuting all 3 conjectures in ~45 minutes given only Coppersmiths’ 1982 paper, and van den Brand, Nanongkai, and Saranurak's 2019 paper: chatgpt.com/share/6acee44d-d… The underlying impulse is very human, I understand why having the right idea on the shelf since 2023 and abandoning it is frustrating. But hinting that the models couldn't have done it without our "human insight" is just pure copium at this point and does no service for anyone.
I rarely post here these days but I wanted to share my reflections: people.csail.mit.edu/virgi/t…
2
4
2
243
23,431
“When a true genius appears in the world, you may know him by this sign, that the dunces are all in confederacy against him.” — Jonathan Swift The leopards ate my face syndrome needs to be studied. If you think it stops at math, wait till you find out what leopards eat…
This quoted post is unavailable.
2
440
Literally saw an econ prof from @HarvardHBS gloating about this on Twitter 3 days ago, then deleted it a few minutes later. Math got hit first because AI labs aimed at it first, not because it’s easier. Point a tiny fraction of the same money at econ and see what happens. What exactly are people schadenfreude-ing about?
Man I understand the hatred towards mathematicians coming from people who genuinely struggled at math and other subjects, but seeing it from literal physics and philosophy professors is unreal. "Maybe their job shouldn't be subsidized" bro all of our jobs are for the same reason
2
2
5
712
Related: when I TA'd algorithms in Fall 2022, I polled the class on P vs NP, framing it with Avi Wigderson’s question of whether intuition can be “mechanized”, or whether solving a problem feels "polynomially close" in effort to verifying a solution. Exactly one student believed P=NP. I asked again when I taught in early and late 2025. Nearly a third of the class raised their hands. I suspect many students growing up with AI will eventually wonder why we were ever stuck on math problems that AI now solves in seconds.
With sub-n log n FFT (130), subcubic APSP & subquadratic 3SUM, today has shaken my belief that humans were ever good at algorithms. At this rate, I've updated to maybe P = NP after all, via some clever SAT algorithm GPT-8 discovers running in n^c time for c ~ 10^6. 🙈
1
3
1
18
3,019
Holy fuck. Just based on a cursory look, major ones seem: Unique Games Conjecture proved L = RL = BPL Matrix multiplication with ω ≤ 9/4, insane jump from the current best ~2.37 Ω(n³) permanent vs. determinant lower bound FFT and integer multiplication faster than nlogn Almost-linear time maximum matching and edit distance approximation
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
1
11
95
10,750