also the credit assignment kicks in. even if you succeed, how much information does any single decision in a sprawling maze really contribute? you turn right - environment is identical. okay turn left - environment still looks the same. how do you reward a solution?

suppose you are training an agent to navigate a maze. each turn contributes almost no feedback, and success is highly path dependent. how do you solve problems like this with RL? the outcome is verifiable, but that doesn't help much given how low SNR the setting is

⚙️ Outdated software = open door for attackers. Before you shop online: 1. Update your devices 2. Turn on automatic updates 3. Delete apps you don’t use Check out our Safe Shopping Guide and get more safety tips from @StaySafeOnline

> see a new Lean proof of Navier Stokes > ask author if it's axiomatized or not > he doesn't understand > explain difference between decide and native decide > he laughs, "it's a marvelous proof, sir" > look inside > it's axiomatized

if you wanted to empower a bunch of nontechnical people to use something like claude code, how would you do it? a tremendous amount of agentic capability is locked behind an intimidating and unfamiliar form factor. how can I make it easier?

almost all people who interact with LLMs today are systematically unaware of where the frontier is because they use the chat UI. if you're not on the ground using an agent that has total control over (and context) of your machine, you're staring at the world through a keyhole.

https://t.co/V4NEQEh8m2

[me on a date trying to impress her] "yah so I'm basically in the NBA...no I didn't play sports, what I mean is I work with excel 12 hours a day and don't have holidays"

this guy just doesn't want anyone to have the secret alpha

don't listen to this weird squid anon, he probably doesn't even trade. you're just one more newsletter away from real alpha

AMAC: sign up for a 90-day trial today to find out why Scott Jennings and millions of Americans are taking a stand for their values and joining AMAC.

On a positive note, a whole lot of my alerting appears to work

Nice, I vibe coded my first outage

It's not hard to envision a future in which every business has a constantly running fleet of spider-like agents "cleaning the hull": doing everything from ideating new opportunities every N hours to alerting infrastructure problems more holistically than Grafana/Prometheus.

The primary interface of LLMs is still fundamentally chat-based, even for Claude Code/Codex CLI agents. But that's weird! Agents are capable of doing so much more with self-governance. Something that works extremely well: set up an agent loop with a "role" prompt (optionally with

Me designing a cool new experiment: Haha fuck yeah! Yes!! I'm so smart! Me seeing the results of the experiment: Well this fucking sucks. What the fuck

@systematicls @quant_arb For an example of why bounding features can be bad: suppose you're trying to forecast a weather system using sensor data, and one of your sensors is continuously emitting wrong data an order of magnitude larger than the rest of the sensors in your fleet. You can't assume the true

☭ Another celeb endorses socialism Mean Girls star Amanda Seyfriend says socialism sounds “gorgeous.” But her argument and reasoning - that "socialism" just means helping each other - suggests she doesn’t actually understand what socialism actually is.

@systematicls @quant_arb I wouldn't use the word "extrapolation", but you're basically right. Suppose you have a regression Y = \alpha + \beta X + \epsilon. If you're willing to assume a normal distribution on Y, then you can estimate how "surprised" you should be by any prediction Y^ = \alpha + \beta X

@systematicls The left looks normal, dense around 0, gradually tapering off to be more diffuse at the edges. The right looks uniform throughout and sharply winsorized at the edges, so possibly percentiled and rescaled to [-1, 1].

@experquisite @therobotjames @Robot_Wealth @systematicls @__paleologo @AgustinLebron3 Two of my articles survive: https://t.co/5ThUoIFhGg and https://t.co/uzoAsIUroX. Those cover data cleaning and modeling returns. One of @systematicls articles is the most concise, useful first intro to stat arb signal research I've seen online. The rest is just doing.

@bennpeifert Instead of a book, I would suggest William Benter's "Computer Based Horse Race Handicapping and Wagering System: A Report". It's a short, 16 page paper. It's a better first course in the spirit and practice of quant trading than almost all books about actual quant trading.

Rio Grande LNG is committed to hiring locally in the Rio Grande Valley, supporting the employment of residents from 40+ communities around the region. ​ Data is from July 2023 – September 2025.

@systematicls Trend and reversion are the same up to sign, so we can unify the LR analysis. From first principles: you have an autocorrelated process m_t = \beta m_{t-1} + \epsilon_t, \epsilon ~ N(0, \sigma^2). The OLS estimate of \beta is COV(m, lag(1, m)) / VAR(lag(1, m)). Expanding out that