Jonathan Larkin

fyi

the model alone is no longer the product

Incredible stuff happening on the AI-run radio stations

@karpathy @ATabarrok Anthropic truly has the mandate of heaven.

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

@tenobrus matlab refuses to die off.

wait what the fuck

@marmaduke091 I think this will be RLM. hence the quotes around infinite. Still very exciting.

Wow. Infinite context windows "coming soon" mentioned in the Claude event. Very exciting. I think they made a breakthrough.

If you can’t convincingly articulate your edge in a market, whether sports betting, prediction markets, stocks, commodities, or whatever else, trading may provide some entertainment but you’ll almost certainly lose money over the long term.

a country of goblins in a datacenter

Would you be interested if JetBrains releases a totally local AI agent, working 100% on your laptop, using our code insight engine and deeply integrated into the IDE? Yes, it will be probably 1 month behind the very recent frontier models, but no token blood bath anymore WDYT?

@tleilax___ @grok is this true?

Wolfgang Schadner, a Swiss quant, found the closed formula for the direct inversion of Black & Scholes option implied volatility ! Everyone has been using root search for ~50 years and his formula is fast. More elegant result, as no boundaries or starting value are required. You still need somewhat of a root search in the inverse Gaussian quantile, or use a smart approximation ;) Then it is down to single digit microseconds. w-schadner.com

this is quite striking

POV you are coding during the year of 2026.5

🧠 New Research ideas! Are LLMS making equity markets more efficient? Here's the number. Lu, Xu, and Vulicevic (2026) use outages at major LLM providers as natural experiments to isolate the effect. When LLMs are available, next-day return predictability drops by 46–61%. When providers go down, predictability roughly doubles or triples, as if the clock ran backward. The mechanism: LLMs process public information faster and on a larger scale, compressing the window for slow arbitrageurs. The pace at which public information gets priced in has visibly accelerated over the past two years. This paper puts a clean number on what I'd been observing informally. The implication for systematic managers is direct: any signal that relies on slow information absorption is decaying faster than it did five years ago. 📄 Lu, Xu, Vulicevic: papers.ssrn.com/sol3/papers.cf… → More research like this in my newsletter: ivanblanco.ai/newsletter

one of my all-time favorite plots

and a separate section for when to add rules, how to have rules for MCP calls, etc. Anyone else doing this?

The idea to use a graph is a good one. I’ve been using a Snakemake file as a project DAG and it’s been working well. My AGENTS.md instructions:

@colossusmag “no west coast solutions to east coast problems” remains undefeated. probably not for much longer though.

Hassabis secretly built a hedge fund inside DeepMind trying to beat Jim Simons. Google shut it down.

@Rainmaker1973 Please share the link to the journal article.

A single dose of a new cancer drug made a brain tumor almost disappear – in just five days. Doctors at Massachusetts General Hospital reported “dramatic and rapid” tumor regression in the first patients treated with a next-generation form of CAR T-cell therapy for glioblastoma, one of the most aggressive brain cancers known. The therapy, called CARv3-TEAM-E, was developed to overcome a major hurdle in treating solid tumors: their ability to hide from the immune system. The personalized treatment reprograms a patient’s immune cells to attack the tumor, and in one extraordinary case, nearly eliminated the cancer within just five days. This novel therapy is designed to target multiple features of the tumor at once, a strategy that may help overcome the common challenge of treatment resistance in solid tumors like glioblastoma. Although the tumors eventually returned, the early outcomes were described as unprecedented. One patient saw a 60% reduction in tumor size that lasted for half a year—an impressive result in a cancer known for its aggressiveness. The trial’s success marks a major step forward for immunotherapy in brain cancer and raises new hopes for long-term control or even a cure. Researchers are now working to refine the treatment and extend its effects, with the ultimate goal of turning a once-terminal diagnosis into a survivable condition.

@iblanco_finance Cool paper, but … “Our results suggest that the learning-to-rank, a philosophy so far largely ignored in finance, sheds novel light on the cross-section of stock returns.”

⚙️ New research! "Learning-to-rank vs. predicting expected returns". One works much better. Lin, Su, and Zhu compare two approaches to building long–short equity portfolios: Standard method: estimate expected returns, then sort stocks. Their method: directly optimize stock rankings. Results: → Learning-to-rank: ~1.5–2.3% monthly excess returns, Sharpe up to 1.2 → Standard return models: 0.35–0.7 Sharpe The reason is clean: when you optimize for returns, you optimize for something you estimate poorly. When you optimize for rankings, you optimize for something more robust, relative ordering. The gains come specifically from better identification of top and bottom stocks, and from better hedging. The method doesn't need better return forecasts; it just uses the available information more efficiently. This is a methodological contribution, not a new signal. But methodology is often where the edge actually lives. 📄 Paper: papers.ssrn.com/sol3/papers.cf… --- → Join the newsletter: ivanblanco.ai/newsletter ---

I hope schools are teaching kids to just sit down with codex / claude code and make stuff.

we as software engineers are becoming beholden to a handful of well funded corportations. while they are our "friends" now, that may change due to incentives. i'm very uncomfortable with that. i believe we need to band together as a community and create a public, free to use repository of real-world (coding) agent sessions/traces. I want small labs, startups, and tinkerers to have access to the same data the big folks currently gobble up from all of us. So we, as a community, can do what e.g. Cursor does below, and take back a little bit of control again. Who's with me? cursor.com/blog/real-time…