localminima

YOU WILL NOT BECOME A QUANT not because you're stupid. not because you don't have access to information all the content is freee - MIT posted Strang's linear algebra course, Harvard gives away probability theory PDFs, Stanford - optimization u won't become a quant because you dont have the discipline to solve 200 textbook problems in a row cuz the moment you see an integral, your brain says "tooooooo hard" and you open TikToooook while you're watching another 5-minute video "how I made $10K in a day trading crypto," a guy your age is sitting down deriving the BlackScholes equation from scratch. nooot copying. Nooot googling the solution. takes a blank sheet, writes dΠ = rΠ dt, and an hour later he has the formula that underpins a trillion dollar derivatives industry on his desk in 18 months he'll be making $300K-$500K u'll be complaining on Twitter that "markets are manipulated" and "the rich always win" the difference isnt luck. the difference is that when he saw conditional probability P(A|B), he didn't close the article he sat down and solved 50 problems until it became intuitive. and you read the definition, said "got it" and moved on. Spoiler: you didn't get it here's the truth nobody tells you: Jane Street, Citadel, HRT - they're not looking for smart people they're looking for people who can sit on one problem for 6 hours and not give up cuz in real trading, nobody's going to hand you a ready solution. The market is 5,000 simultaneous equations with 50,000 variables, and they're all changing every millisecond the average Jane Street employee made $1.4 million per year in 2025. That's AVERAGE. Not top trader. Not a legend. Just a regular guy who knows what eigenvalue decomposition is and isn't afraid to use it and you? You still think trading is about "feeling the market move"" That if you post cool profit screenshots on Telegram, someone will believe you know what you're doing quants don't feel. Quants calculate. While you're guessing "will Bitcoin go up or down," they've already calculated that at current volatility σ=0.65, correlation with S&P ρ=0.43, and accounting for conditional probability based on onchain metrics, the expected value of going long is negative. So they short. And they take your money this article gives you the entire roadmap. literally step-by-step what to learn, which books to read, what code to write. All free. All accessible. 18 months at 2 hours per day but you won't start. because lvl 1 homework is "solve all problems from chapters 1-6 of Blitzstein's textbook" that's 150+ problems. and your brain has already found an excuse: "I don't need this, I'll just trade patterns" okay. keep going. keep blowing up accounts and believing "next time I'll get lucky" And somewhere, a guy who's sitting today deriving Itô's lemma will be making your annual salary in a month in 2 years and the funniest part? You'll read this text, feel a sting, maybe even tell yourself "damn, I need to get serious" u'll open the textbook. You'll see the first formula and you'll close it cuz you don't want to BE a quant. You want to LOOK LIKE a quant. And those are different things

Omar, former Two Sigma quant who turned human judgment into systematic alpha: "If an analyst is heavily bullish and gets even more bullish, the best thing to do is bet against them." Everyone assumes the opposite. Bullish means up, bearish means down. That is the trap. By the time the crowd is most certain, it is all priced in. The signal is not the crowd. The signal is the first person who steps out and says the emperor has no clothes. That is when you pile in. Watch this first, then read the article below.

This paper completely changed how I think about LLM agent architecture in trading: Analysts -> Bull vs Bear debate -> Trader -> Risk team -> Fund Manager Here is the 5-step blueprint for a multi-agent trading pipeline: Analysts: four parallel agents (fundamentals, sentiment, news, technical) read their own data feeds and submit structured reports instead of free-form chatter. Bull vs Bear debate: two research agents with opposing positions argue for n rounds in natural language, a facilitator records the winning side as a structured entry in the state. Trader: reads analyst reports and the researcher verdict, emits a bullish or bearish decision with reasoning, writes a separate report for the risk team. Risk team: risk-seeking, neutral, and risk-conservative agents debate the trader's decision for n rounds, adjusting position size and limits. Fund Manager: reviews the risk debate, finalizes the decision, updates state, and triggers execution. Key insight: agents that talk through structured documents preserve context, agents that talk through free-form natural language dialogue play a game of telephone. On AAPL over January to March 2024 TradingAgents reports a Sharpe of 8.21 against 1.64 for the best rule-based baseline KDJ&RSI on the same ticker. Read this, then check the article below.

This paper completely changed how I think about specializing LLMs for trading: Multimodal data -> Volatility labeling -> Structure -> Claims -> Decision Here is the 5-step blueprint for training a trading LLM: Multimodal data: collect 100k samples across 14 blue chips over 18 months from five channels - prices and technical indicators, news, fundamentals, insider sentiment, macro. Volatility labeling: build a composite signal from EMA returns over 3, 7 and 15 day horizons normalized by rolling 20-day volatility, then project onto five classes Strong Sell / Sell / Hold / Buy / Strong Buy. Structure (Stage I): supervised fine-tuning plus RFT on section organization (fundamentals, technicals, sentiment) so the model learns to reason like an analyst. Claims (Stage II): RFT on an opinion-quote-source scheme per bullet so every claim is anchored to a citation from the input data and hallucinations drop. Decision (Stage III): RFT with an asymmetric reward matrix that penalizes false bullish signals about 12% harder than false bearish ones, reflecting how fast markets fall and the priority of capital preservation. Key insight: reasoning-tuned LLMs (o3-mini, o4-mini, DeepSeek) lose to plain LLMs on trading because their reasoning drifts away from financial data. In the held-out June-August 2024 window Trading-R1 at 4B parameters hits a Sharpe of 1.80 on AAPL, while OpenAI's o4-mini reaches a Sharpe of -1.36 on the same ticker. Read this, then check the article below.

Victor Haghani, co-founder of LTCM, the fund whose 1998 collapse nearly broke global markets: "Elon Musk has a 90% chance of having very little wealth left in ten years." Not because the businesses are bad. Just the math of the distribution at 50% volatility. All of the probability mass sits below what he holds today. "How do I become a billionaire" is the right answer to the wrong question. Becoming a billionaire is easy. Take a ton of risk. You will almost certainly lose it later. A man who was right on his trades and still went bankrupt broke down where the mistake lives.

This paper completely changed how I think about backtest metrics: Number of trials N -> Variance of Sharpes -> Sample length -> Skew and kurtosis -> Deflated Sharpe Here is the 5-step blueprint for honestly evaluating a strategy: Number of trials N: record how many independent strategy configurations were tested, because the more trials, the higher the expected max Sharpe even under zero true skill. Variance of Sharpes: compute V[SR] across all trials, wider variance means more room for a random extremum to look impressive. Sample length T: short histories inflate Sharpe, sample length enters the correction directly. Skew and kurtosis: negative skew and fat tails systematically inflate Sharpe, so higher moments are part of the deflation. Deflated Sharpe: plug all five inputs into the DSR formula and read off the probability that the true Sharpe is above zero after correcting for every source of inflation. Key insight: in markets with memory effects, an overfit strategy does not just decay to zero, it systematically loses money. A strategy with a reported Sharpe of 2.5 over 5 years falls below the 95% confidence bar to a DSR of 0.9004 once the author discloses that 100 configurations were tested. Read this, then check the article below.

This paper completely changed how I think about LLM agents in trading: Past-cutoff test -> Counterfactual perturbations -> Memorization probes -> Post-finetune audit Here is the 4-step blueprint for diagnosing information leakage in a trading agent: Past-cutoff test: run the agent in matched market windows before and after the base LLM's training cutoff to separate forecasting from memorization. Counterfactual perturbations: modify key inputs (earnings, technical indicators, price paths) and measure Prediction Consistency, the share of predictions that stay unchanged. Memorization probes: ask the model direct QA on past data points like "what was NVDA's close on March 15 2022", any accuracy above chance is recall, not reasoning. Post-finetune audit: measure in-distribution vs unseen accuracy before and after finetuning on financial text, the gap shows how much finetuning just imprints the past. Key insight: finetuning on financial data makes the model better at remembering the past and worse at forecasting the future. FinMem keeps 82.13% of its predictions unchanged after counterfactual perturbations to its inputs. Read this, then check the article below.

Tom, former quant PM at Tudor Investment and Moore Capital: "I haven't seen a real new idea in trading in at least fifteen years." Every idea you come up with in year two, year three, year five was thought of a hundred times before you. Everyone in quant finance is a genius. That does not make you special. It makes you just like everyone else. Building a social network against Facebook is easier than finding edge in trading. A veteran of two legendary funds broke down what actually moves the needle.

Ex-Google engineer explained AI agent loops, harness, evals in 20 minutes - better than 500$ courses. trace every run -> judge it with an LLM -> diagnose -> fix -> ship. That loop is how agents self-improve over time. Agent loops + memory + harness + evals - thats the stack. Watch it, then save the framework below.

This paper completely changed how I think about quant strategy verification: Data -> Maker (signal) -> Checker (independent verification) -> Execution -> Risk monitor Here is the 5-step blueprint: Data ingestion: an automation pulls OHLCV, order book depth, and news on a cadence matched to the asset class, one hour for equities, one minute for crypto. Maker: an agent reads a skill file with rules and accumulated lessons, opens the latest data, and emits a candidate trade with reasoning. Checker: a separate agent with no exposure to the maker's reasoning runs fixed gates, Sharpe above 1.5, max drawdown below 10%, Newey-West t-stat above 2.0, out-of-sample window at least 24 months. Execution: only signals that clear every gate route to the broker via an MCP connector, with position size capped by rules inherited from the skill file. Risk monitor: a parallel watchdog polls positions every minute and flattens everything if drawdown crosses 5% of allocated capital. Key insight: a low checker rejection rate is not a sign of maker excellence, it is a sign the checker is too loose. In the authors' production deployments the maker-checker split rejects 40-60% of candidate signals before they reach execution. Read this, then check the article below.

@h100envy @zostaff I'm the one who comes to your mom's bedroom on weekends

@localminimaa @zostaff lol bro, she is anthropic engineer in 25 years old, but who are you?

Fiona Fung, who leads the Claude Code and Cowork teams at Anthropic, 25 years an engineer: "Anthropic engineers ship eight times as much code per quarter as they did in 2025." Then she names the split nobody wants to be on the wrong side of. Some engineers are leaning in and thriving. Others are frustrated, fighting, resisting. What separates them is not talent. It is a growth mindset, the willingness to admit that what made you good until now may not keep you good. Her read on the resistance is blunt: under the frustration is fear. And the only way through is to ask what is actually in your control, and do that.

Schneier gave this talk before the Mythos numbers came out. He framed hacking as finding what a system allows but never intended That used to describe a human with time and talent. Now it describes a model running overnight The lecture aged faster than anyone wanted

Schneier called this in 2021. AI doesn't hack like a human - it finds what the system allows but didn't anticipate Mythos is just the first time we have the numbers to prove it at scale (video from channel Center international de criminologie comparée CICC)

Run from the crash, you sell the recovery too. The best days hide inside the worst ones You cannot dodge the lightning and keep the rain

The best days in the market live right next door to the worst ones. Run from the crash and you sell the recovery too The crash and the recovery are the same event wearing two faces Watch

@h100envy the launch gets the headline. the infra gets nothing

Evan Morikawa ran the team that kept ChatGPT alive when it went from zero to a hundred million users in two months. No ML PhD. Just the hardest scaling problem in tech history. When the whole world hit one product at once, his team was the reason it didn't collapse. The bottleneck wasn't the model. It was that GPUs don't grow on trees. Everyone remembers the launch. Almost nobody knows the engineer who stopped it from falling over.

@thegreatest_sv $20/month in, $15k/finding out. the math is offensive

A SOLO DEVELOPER IS USING CLAUDE CODE TO HUNT BUG BOUNTIES ON PLANES. THE SETUP COSTS $20/MONTH. THE PAYOUTS RUN TO $15,000 PER FINDING. Companies are paying thousands of dollars for finding vulnerabilities in their systems. > The Math: > 1 valid low-severity bug = $200–$1000 > 1 valid high-severity bug = $1,000–$15,000 > 1 critical RCE on a top program = tens of thousands > DeFi audits on Immunefi = some of the largest payouts in the field > The Stack: > 51 skills: auto-load based on what you're testing > 574 patterns: curated from real disclosed HackerOne reports > /triage: 7-question gate that validates before you file > /report: generates clean submissions for every major platform > Burp MCP: Claude reads your HTTP traffic directly > The Cost: $20/month Claude plan. Everything else free. > You don't need a security degree. You don't need expensive tools. You don't need a team. > You need the bundle and 3 minutes to install it. Full setup, every command, legal targets to practice on below

@slash1sol map open, most still sleeping

Everyone is going to watch the 2026 World Cup. Almost nobody is going to read it like a market. 104 matches isn't just more soccer -> it's the deepest tournament ever played and the first one with zero historical base rate, because nobody has ever run a 48-team, 12-group format. Every model trained on old World Cups is flying blind into it. 16 venues across a whole continent means Mexico City altitude, Texas heat, and travel fatigue nobody's pricing in yet. Fitness edges the early lines won't catch. @gmpm_xyz ep. 7 with Dave, Kkoma and Bread is the map you open before the lines wake up. Most people will catch the chaos but a few will catch the edge.