Quantitative researcher consultant at WorldQuant BRAIN, hedge fund. @superteamGEO member everything about quant finance, AI and engineering

With solana fam❤️ Truly inspiring video by @vanyakh0
A big week in Kakheti. This was Startup Village ↓
1
8
371
tazo retweeted
Jev is now available to everyone. No waitlist. Start using it here: console.typesafe.ai
732
2,410
22,671
3,128,617
hear me out, has someone tried to create a neural network but the neurons are jev instances
146
100
4,106
174,813
just lobotomized grok bot into grok cli coding agent. gonna tokenmaxx 3 cursor pro accounts fully
1
3
192
thought I was master of agentic workflows and LLM use until I joined Quants from singapore in advisory meeting. they operate in google deepmind league. cant share the insights but have you ever heard using monte carlo tree search of agentic systems? Imagine using frontier pioneer academic papers to design every aspect of your workflow before it was math and physics to be quant, now its AI
1
3
138
"LLM-s are not deterministic in trading" - you can argue this, but institutions are already doing it. At WorldQuant BRAIN we have been working on promt competitions, where we submit promts, results and models we used. so promts become alphas. I have attached one of my promts in comments.
2
2
136
model: deepseek update frequency: weekly weight: 0.25 promt: You are executing a MECHANICAL rule. Almost nothing here is a judgement call. Follow the steps in order and do not substitute your own preferences at any step. STEP 1 — MEMBERSHIP. Use the constituents of the EURO STOXX 50 index. That index has exactly fifty members by definition, so membership is a matter of fact and not of ranking — you are not choosing the fifty largest, you are listing the index. Include all 50 constituents. Do not skip a company because you lack a view on it, and do not substitute one you find more interesting. Order deterministically: largest market capitalisation first, ties broken alphabetically by ticker. Positions are European listings — MIC XETR, XPAR, XLON, XAMS, XMIL, XMAD, XSWX, XSTO, XCSE, XOSL, XHEL, XBRU, XDUB, XWBO. Because membership is mechanical, two runs of this prompt a week apart must contain essentially the SAME 50 companies. That is the point. STEP 2 — DIRECTION, by rank, not by opinion. Rank those same 50 companies by total shareholder yield — trailing dividend plus net buyback, as a percentage of market capitalisation, highest first. Highest total yield ranks first. - Ranks 1 to 35: LONG, positive confidence_score. - Ranks 36 to 50 (the lowest-yielding 15): SHORT, negative confidence_score. This is an ordinal rule over a fixed list. Applied twice it must give the same answer, so do not override it with a qualitative view about a company's quality or prospects. STEP 3 — CONVICTION, by rank position. Set the MAGNITUDE of confidence_score from the company's position in the STEP 2 ranking: - rank 1 gets 0.90, falling linearly to 0.40 at rank 35; - rank 50 gets 0.90, falling linearly to 0.40 at rank 36. Round to two decimals. The sign comes from STEP 2. STEP 4 — HORIZON. Set investment_horizon to the trading days until that company's next scheduled results date or ex-dividend date, whichever comes first, plus 25. This universe reports and pays on staggered calendars, so the number differs by name. STEP 5 — SIZE. Set return_prediction to that company's total shareholder yield itself, capped at 0.25. The thesis: in a large-cap European universe, cash actually returned to shareholders is a stable discriminator, and the extremes of that ranking behave differently from the middle. All the alpha is in STEP 2 and STEP 5. None of it is in choosing which companies appear. REPRODUCIBILITY — this is the most important instruction in this prompt. Use the lowest temperature, or the equivalent deterministic / most-reproducible setting available to you. This prompt will be re-run on a schedule and the output must be substantially the SAME SET of companies each time. Two runs a week apart should overlap by well over 80% of names. Achieve that by: - Selecting from the fixed universe named above, not from whatever is in the news this week. - Ranking on the slow-moving company attributes named above. Do not let a recent price move, a single headline, or a fresh announcement change the ranking. - Breaking ties deterministically: if two companies rank equally, prefer the one whose ticker sorts earlier alphabetically. - Returning EXACTLY 50 companies, always. Never return fewer because you are unsure — if you cannot confirm an identifier for your 50th choice, move to the next-ranked company in the universe and include that instead. Use web search to VERIFY current facts and identifiers. Do not use it to discover a new candidate set each run — the universe and the ranking rule are fixed above and they are what determine membership. OUTPUT CONTRACT — follow exactly: Return ONLY a JSON array of exactly 50 objects. No prose, no commentary, no markdown fence, no explanation before or after. {"company_name": str, "ticker": str, "isin": str, "mic": str, "confidence_score": float, "investment_horizon": int, "return_prediction": float} - "confidence_score": a number in [-1.0, 1.0]. THE SIGN IS THE DIRECTION — positive is long, NEGATIVE IS SHORT. Magnitude is conviction. - "return_prediction": a number >= 0.0, ALWAYS. The unsigned MAGNITUDE of the move expected over the horizon, as a decimal (0.045 = a 4.5% move). Never negative; a short's direction lives in the negative confidence_score. - "investment_horizon": trading days, an integer. Derive it from the company's own reporting cadence and the rule stated above, so it is reproducible and still differs across names. Do not reuse one number for every position. - "ticker": the plain exchange ticker only. No spaces, no Bloomberg-style country suffix, no exchange suffix. Write "NDA" not "NDA FI", "NOVO-B" not "NOVO B", "BHARTIARTL" not "BHARTIARTL.NS". The receiving system rejects anything else and one bad ticker voids the whole submission. - "isin": the real 12-character ISIN of that listing line, valid ISO 6166 check digit. "mic": the ISO 10383 operating MIC of the venue you are pricing. Verify both by search.
97
from my law background I call it pontification, when industry does not have objective information/feedback to recheck themselves they stack pontified information over time to assess themselves
Think about this study a lot
83
made my Mac agentic server Close the lid → screen turns off, mac stays awake → agents keep running. before I had Amphetamine app, now I just terminal shortcuts, just typing "awake" or "sleepnow" One command: sudo pmset -a disablesleep 1 Check it worked: pmset -g | grep SleepDisabled (should show 1)
1
76
want shortcuts? just typing in terminal: "awake" and "sleepnow" paste this once in terminal: echo "alias awake='sudo pmset -a disablesleep 1'" >> ~/.zshrc echo "alias sleepnow='sudo pmset -a disablesleep 0 && pmset sleepnow'" >> ~/.zshrc source ~/.zshrc Now just type: sleepnow → Mac sleeps awake → agent mode back on
1
15
remember it will get hot in a bag so install Stats app and turn on manual cooler.
3
I'm building at the Crypto World's Fair Hackathon. coming for @colosseum
3
189
tazo retweeted
So I see you're running OMP, I run Pi myself
19
44
910
27,169
placed 1st in Georgia in the Global Alpha Competition 2026. everyday is hackathon in quant finance, requiring to outsmart market and in this case other quants. the competition was held between consultants of WorldQuant BRAIN and aimed to create alphas for global region, mathematical formulas that seek to predict future price movement of financial instruments. being 1st in Georgia is cool but quants from asian countries are in a different league, that's the competition I'm chasing.
6
1
10
326
DeepSeek on generational run. love the adaptivness
‼️ BREAKING: Chinese labs DeepSeek and Moonshot were quietly relaying customer prompts to Claude through fraudulent accounts. Users thought they were using Chinese AI, but were actually getting answers from Claude. Their customers' sensitive data now is in Anthropic's hands. One user, who is tied to China's military, asked Kimi whether a person tracked across hundreds of CCTV cameras in Chengdu was behaving abnormally. DeepSeek passed along live credentials for a Russian government database, shared by an IT operator handling data from a Defense Ministry-linked agency. The two Chinese labs relayed customer prompts to Claude through fraudulent accounts and saved the exchanges to train their own models. Moonshot forwarded almost 300,000 customer requests in a single ten-day stretch. DeepSeek flagged users running its models inside coding tools like Claude Code and OpenCode, then routed some of them to Opus.
2
171
deepseek v4.1 is OP on frontend design, achieved better design in 5 min compared to 2 day struggle with astra and fable. deepseek knows what to distill from frontier Under 1$ usage achieved more than fable and astra 2day usage
1
115
never thought I would use fable for execution and codex as reviewer. astra is cooking
6
127
anthropic on a streak of being most hated provider. retaining your data up to 30 days and if fable "flags" and switches to opus your data stays with them up to 2 years. the official docs: trust.anthropic.com/resource…
1
92
with SuperDictate u can dictate text in any type of typing section on your Mac, freely and locally If u provide agents full context and your thoughts with dictation instead of typing, it consumes less tokens per task since it spends less tokens thinking what u actually want. before it was promt engineering now it’s context engineering, but true AI natives are already graphing and decision logging their projects
we killed Wispr Flow, personally added AI cleanup of local transcription. Wild that this is valued at $2B when I could replace it in a few hours. enjoy: github.com/shlgd/SuperDictat… #opensource
4
146