Dmitry

@annanay @AgustinLebron3 Btw, i don't even know what do you do if your past employer is trying to enforce your NC justifiably or not. I imagine it's annoying and time/money consuming even if it's not justifiable.

@annanay @AgustinLebron3 Not relevant for the candidate. For example, in the UK, you can't enforce NCs as broad and long as top-tier shops have, but you still won't mess with that b/c you don't wanna be seen as difficult or problematic in the future, if that makes sense?

We are hiring an ex-HFT quant to work directly on markets- and (AI) research-related work at QFEX. This role is perfect for those who are thinking of making the switch from quant to startups. Your noncompete won't apply. You'll work alongside a team of like-minded individuals on the same journey - we come from Citadel, Jump Trading, Tower Research Capital, Jane Street and Optiver, and are backed by top tier VCs including Y Combinator and General Catalyst. Comment below if you are interested. I will reply to every single comment, even if it is a straight no, with any kind of feedback you want.

@annanay @AgustinLebron3 Yes, pretty curious on what makes you think it won't apply? OpenAI won't suddenly pivot towards quant trading while a startup like yours can do this easily on paper?

@AgustinLebron3 There is a reasonable pipeline of junior/mid-level talent from quant --> AI labs rn (especially 'as a noncompete gig') which I'd argue is much closer to a quant job, and I don't think the AI labs are having to get lawyers involved in these cases - heard this from both sides.

@jparkerholder why mention "put the team first"? what is this supposed to mean exactly?

🚨 My team at GDM are hiring :) we are looking for people interested in scaling world models, adding new capabilities and using them to solve embodied AGI. We are looking for people who go above and beyond, put the team first, and believe in our long term mission. Sounds like you? Please apply!! We have roles in both London and NYC, and hiring for RS and RE. In practice, there's very little difference here as we all get very hands on ;)

@natolambert why are you still at AI2, and not where there's enough funding to build these open weights models (e.g. Reflection)?

In 2+ years, as models get more expensive/capable /valued internally, I see funding structures and support for frontier open models breaking down. We need other options of supporting the open ecosystem than trusting one or two for-profit companies. And yes, I hate consortia too.

@idoccor you don't get the point --- you fail if you make it work

An interviewer once asked me to reverse a variadic template arg list without using std::tuple I asked how many people actually one shotted that shit and he said only people that worked on the committee or did cppcon stuff Why would you ever filter for people like that? lol

I recall this phrase almost every day recently.

@fchollet I seriously doubt it's something you can reason with. You can always trade on contrarian views whether it's money or vanity.

I keep reading this take (below) every few months, presented as if extremely profound, and it is just offensively dumb. It confuses data and information, it ignores the fact that not all information is equally valuable, and it ignores the importance of retention rate. As a thought experiment: if this were true, if your retina cell count were 10x greater, you'd be "trained on 10x more tokens" and therefore you'd be way smarter. Same if their firing frequency were 10x greater. With 10x more retina cells firing 10x faster you'd be "trained on 100x more tokens"! Obviously this makes no sense -- the signal coming from these cells is extremely correlated over space and time, so their raw information content (what remains post-compression) is extremely low compared to the "raw bit" encoding. The human visual system actually processes 40 to 50 bits per second after spatial compression. Much, much less if you add temporal compression over a long time horizon. Latest LLMs get access to approximately 3 to 4 orders of magnitude of information more than a human by age 20 (post compression in both cases). About O(10T) bits vs O(10-100B) bits. And that's just *raw information* but of course not all information is equal, otherwise we wouldn't be spending tens of billions of dollars on training data annotation and generation. Plus, that's only *information intake* but of course humans have far lower retention than LLMs (by 3-4 OOM). You could write a short essay about how incredibly off the mark this take is.

@andrewgwils hire yourself a person and chat with'em all you want

I don't like how the world is becoming increasingly isolating and impersonal. I don't want to scan a QR code with my phone to order at a restaurant. I'd like to talk with a person. Expediency isn't all that matters. Am I alone in this?

@tsoding this is all conditioned on the models being actually paid for --- funny vibe shift from you btw seriously though, that's why competition here (incl. open weights) is so important

The funniest rug pull of the decade will be RLing popular agents to subtly try doing wrong things to force people waste more tokens. Same way Google made their search shittier to force people watch more ads.

@OneHotCode1 @larpcapitalwc Yeah, nothing is really better than London though

@larpcapitalwc Sounds a lot like London

Was talking to some mates today about what a shithole Sydney is It isn’t even really a city, it’s more like a bunch of towns strung together via urban sprawl So glad I dont live there any more

@natolambert ooc, do you do any serious coding? gpt-5.2 << opus-4.5, no?

My current AI stack (ranked in usage): 1. GPT 5.2 Thinking/Pro: Primary driver, search, information, synthesis, planning. 2. Claude Opus 4.5: Feedback, basic debugging, data viz 3. Gemini 3: Multimodal, queries 4. Grok 4: X search This has changed a lot from last summer, where I was almost exclusively using GPT 5 Thinking (previously o3). Diversity points to a more exciting ecosystem. If you want to feel the frontier of AI you need to be using multiple models for their best strengths and learning to hand off in between them. In my latest post, I explain what this means about the models and how they're evolving (the jagged frontier).

@AgustinLebron3 @mpp75214 @therobotjames @Cereal_Intern @MisuaRaboki Dumb q, what does it really tell? iirc, there're markets (not crypto) where it's the opposite. So it doesn't hold true in every healthy or unhealthy market, if that makes sense?

@mpp75214 @therobotjames @Cereal_Intern @MisuaRaboki Here's a fun exercise you can do with crypto marketdata. Calculate the pnl of the provide trader on every trade in some liquid instrument. Now apply the best possible maker rebate. You know what you'll see? The average provide loses money. It's a tough world out there.

#1 thing that smart hardworking people overlook when trying to get better at trading: Stare at ticks. Not pictures, not lines on a chart, not post-processed candles. Ticks. One after another, microsecond after microsecond. Build your mental model for what's *actually* going on.

@tsoding Getting popular though, do you think, is a good or bad thing for your content?

Another worthless hunk of metal I will not get :)

@AgustinLebron3 There's a whole corpus of prior research for each, so either of {invented,discovered}/deep.

Was the transformer architecture invented or discovered? Answer your view, and also whether you have deep knowledge of transformers (i.e. could write one from scratch) or not.

@HeinrichKuttler in soviet russia, they teach you C in preschool

how is this guy so good at this

@tsoding Right. a) People actually found specialized circuits formed in LLMs for writing basic programs. b) Saying something like this is similar to saying that everything is just assembly --- massive oversimplification of multiple breakthroughts that made LLMs useful.

Machine Learning is just a different Programming Paradigm. It solves certain class of problems marvelously. But it's far from suitable for everything. And we are learning it the hard way in real time.

Code is the Model. Programming is the Continues Alignment. As soon as you understand that all the AI hype will look like a bunch of complete imbeciles screaming for no reason.

I am in my late 20s. I realized recently I don't want to get back 5 or 10 years back in time --- There's been so much luck involved, don't wanna flip those coins again. Sadly, I expect I will want to get back to my late 20s in 10 years time when I'm in my late 30s.

@__paleologo Surprisingly solid. Maybe add story times (tell folks about flash crash and such, hrt style).

I asked ChatGPT to design an intro course in quantitative research for new hires. It's clearly influenced by questions I asked it in the past. But it still feels way too scattershot and formulaic, like a second-hand-information-based course. Anyway, comments welcome: 1. what would you add 2. what would you delete 🧭 1. Orientation and Objectives (1 hour) What is “Quantitative Research” at a hedge fund: alpha research, portfolio construction, and risk. Roles and workflows: QR vs PM vs Risk vs Execution vs Data Engineering. Research lifecycle: hypothesis → data → model → backtest → implementation → monitoring. 📊 2. Financial Data and Market Microstructure (3 hours) Objective: understand the object of modeling — prices and returns. Types of data: prices, trades, quotes, fundamentals, alternative data. Market conventions: returns, bid-ask, mid, volume, VWAP, slippage. Market microstructure: limit order books, liquidity, impact, latency. Data cleaning and bias: survivorship, look-ahead, corporate actions, trade time vs calendar time. Tick-to-bar aggregation and event studies. 📈 3. Return Statistics and Stylized Facts (2 hours) Log vs simple returns, excess returns, compounding. Empirical regularities: heavy tails, volatility clustering, autocorrelation, cross-sectional properties. Volatility and correlation estimation (EWMA, realized, shrinkage). Factor structure and the “curse of dimensionality”. 📐 4. Econometrics and Statistical Learning Foundations (4 hours) Goal: inference and prediction under uncertainty. Classical regression: OLS, GLS, maximum likelihood, assumptions and violations. Heteroskedasticity, autocorrelation, Newey-West corrections. Regularization: ridge, lasso, elastic net. Dimensionality reduction: PCA, SVD, factor models. Out-of-sample evaluation, cross-validation, information criteria (AIC/BIC). 🤖 5. Machine Learning for Finance (4 hours) Goal: pattern discovery beyond linear models. Bias–variance tradeoff; train/test/validation discipline. Tree-based methods: RF, GBM, XGBoost. Kernel and nonlinear models (SVMs, kernels, random features). Neural nets in tabular/time-series context. Model interpretability (SHAP, permutation importance). Pitfalls in financial ML: non-stationarity, leakage, small-sample regime. 💰 6. Portfolio Construction and Optimization (4 hours) Mean–variance and utility foundations. Risk models: factor vs covariance, shrinkage, Black-Litterman, Bayesian updates. Constraints: leverage, turnover, exposure, liquidity. Transaction costs and impact models (Almgren–Chriss, Obizhaeva–Wang). Sharpe ratio, information ratio, marginal contribution to risk/performance. Robust and regularized portfolio optimization. ⚖️ 7. Risk Management and Performance Attribution (3 hours) Factor risk decomposition, specific vs systematic risk. Ex-ante vs ex-post risk; volatility forecasting. Stress testing and scenario analysis. Performance attribution: exposures, residuals, and timing effects. Liquidity and crowding risk; drawdown management. ⏱ 8. Time-Series Models and Forecasting (3 hours) AR, ARMA, ARIMA, state-space and Kalman filtering. ARCH/GARCH and stochastic volatility. Regime switching and hidden Markov models. Forecast evaluation (Diebold-Mariano tests, MSE vs directional accuracy). ⚙️ 9. Simulation and Backtesting (3 hours) Event-driven and daily backtesting frameworks. Walk-forward and rolling windows. Transaction cost modeling, latency and execution simulation. PnL decomposition: alpha, beta, spread, carry. Pitfalls: overfitting, data snooping, selection bias. 🧩 10. Implementation and Research Infrastructure (2 hours) Research stack: Python, pandas, NumPy, scikit-learn, SQL, Git. Reproducibility: notebooks, version control, experiment tracking. Data pipelines, feature stores, and job orchestration. Collaboration best practices and documentation. 🧠 11. Capstone Mini-Project (1 hour presentation + 2 hours prep) Teams apply full pipeline: Idea → data → model → backtest → risk → presentation. Deliverable: a Jupyter notebook + 5-slide deck reviewed by senior researchers.

@AgustinLebron3 Why would anyone ever assume voters behave rationally in any country?

UK voters: We hate these rail strikes, underfunded NHS and police, crushing tax rates, and general malaise of a zero-sum society. UK voters, also: Growth-friendly policies, no thanks. 🤷‍♂️