Pinned Tweet
long before AGI eats the world, narrow AI will eat the stock market.
it seems obvious this will happen with an open hedge fund.
one that gives data away for free and lets any new AI improve it.
this was just me using muse to cancel my restoration hardware subscription.
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Richard Craib retweeted
We're introducing TasteBench, a new standard to measure Research Taste on real world ML tasks.
Existing benchmarks have limited datasets and eval data ultimately becomes part of training data. In TasteBench, we continuously generate new ML tasks on OOS data no model has trained on.
When Numerai started, users were limited to only the 40 features we gave out.
Today, there are no limits and AI agent fleets are sourcing and engineering their own features, modeling them, and submitting them to Numerai.
Today marks a turning point in Numerai’s history.
For ten years, Numerai hid stock identities. Now, we’re revealing them.
Introducing Unity, a new dataset with 1,040 features and real stock identifiers, designed for a future where vast fleets of autonomous AI agents research global equities, engineer new features, and build predictive models at scale.
With Unity, Numerai and Numerai Signals will merge into a single tournament, aggregating all intelligence together into one Meta Model.
Discover Unity and learn how to prepare your agents for the next era of Numerai: blog.numer.ai/unifying-numer…
Feature sourcing and engineering is wildly scalable with AI agents and Numerai has the community to supercharge this search for new features.
A prompt like this in Opus 5.5 built 3 advanced features for $33 in tokens.
“Let's make some new Signals features.
My idea is that we can read public filings, and run them through an LLM and ask some questions to generate features. You'll need to
1. pull filings from EDGAR
2. pass them through a free, cheap JEV-like LLM (Kev: github.com/jaredpalmer/kev, for example), scoring on a scale of "buy" to "don't buy"
3. rank and bin those scores per era and join them with v3.1 Unity data”
Richard Craib retweeted
株式のアルファ探索では、もうNumeraiを超えることは難しい気がする。
・エージェンティックサーチを可能にするために、難読化されていないデータを提供してユーザーが独自に保有するデータとの統合を図る。
・このサーチを、参加者に分担させて行わせる仕組み（MCP）が整えられている。
Today marks a turning point in Numerai’s history.
For ten years, Numerai hid stock identities. Now, we’re revealing them.
Introducing Unity, a new dataset with 1,040 features and real stock identifiers, designed for a future where vast fleets of autonomous AI agents research global equities, engineer new features, and build predictive models at scale.
With Unity, Numerai and Numerai Signals will merge into a single tournament, aggregating all intelligence together into one Meta Model.
Discover Unity and learn how to prepare your agents for the next era of Numerai: blog.numer.ai/unifying-numer…
Numerai and Numerai Signals become one.
Real stocks. One Meta Model. One API for stock market superintelligence.
Today marks a turning point in Numerai’s history.
For ten years, Numerai hid stock identities. Now, we’re revealing them.
Introducing Unity, a new dataset with 1,040 features and real stock identifiers, designed for a future where vast fleets of autonomous AI agents research global equities, engineer new features, and build predictive models at scale.
With Unity, Numerai and Numerai Signals will merge into a single tournament, aggregating all intelligence together into one Meta Model.
Discover Unity and learn how to prepare your agents for the next era of Numerai: blog.numer.ai/unifying-numer…
tell new hire: you won’t be able to increase the speed of our heavily optimized backtest simulations.
new hire: I used Fable to increase the speed of the backtest simulations by 60%!
this type of thing occurs daily now.
but not just to us to every technical company in the world all the time in parallel.
do you see?
I’m putting together a new team for a self-improving hedge fund.
Our fund runs on thousands of machine learning models already. Increasingly, those models are built by AI, not people. This is not a thesis this is already how Numerai works (see numer.ai).
We're forming a new team with one mandate: build AI that does quant research autonomously, all the time. New features, better risk models, better portfolio construction. For our scaling AUM.
Small company in SF, huge leverage per person. No bureaucracy. Significant equity.
Hiring:
AI Scientist → wellfound.com/l/2Cwt7S
Lead Software Engineer → wellfound.com/l/2Cwt81
Senior Quantitative Researcher → wellfound.com/l/2Cwt7M
How important is the quant hedge fund’s Sharpe in this analysis? Do things get bad if the hedge fund component is the same 10% return but on 20% volatility (Sharpe 0.5)? or way better if the hedge fund is 10% return on 5% volatility (Sharpe 2)?
Assuming the same 10% expected return and varying the volatility, Sharpe isn’t actually such a big deal.
Median outcome of $1 in QQQ + 10% uncorrelated alpha past 10 years.
Sharpe 2 → $15.8
Sharpe 1 → $15.3
Sharpe 0.5 → $13.2
(QQQ alone → $6)
Sharpe 0.5 simulations have drawdown coincidence with QQQ a little more often and more vol drag so you lose a little. But the difference between Sharpe 1 and 2 is 50c in 10 year median.
The real Sharpe tax is survival. A standard -20% LP drawdown limit fires the Sharpe 0.5 fund in 97% of paths!
I tried to get a VC friend to give me a QQQ benchmark on his fund. Let’s just say I didn’t get the allocation 😂.
But the bar vs publics gets even higher if you consider the existence of quant hedge funds.
If you invested in a market and factor neutral quant fund with 10% net return (above cash) on 10% volatility and blended it with QQQ portable alpha style, your 10 year growth was:
$1 → ~$14 vs $1 → $6 in QQQ
30% return vs 20%.
This is why uncorrelated alpha is valuable. Not because it’s high return, but because its return geometry lets it stack onto any portfolio without it.
Try the same trick with a beta and factor exposed ‘hedge’ fund, and you have a portfolio in extra distress in 2022 (>-50%!). Because you have drawdowns stacking instead of returns!
You can think of a fund that tilts into factor bets as making a mess of your return geometry by taking risk on things your already own. For example, your portfolio doesn’t need a hedge fund to tilt you into a bet on AI. The S&P is ~48% an AI bet (so far). By betting on AI beta, the hedge fund “unhedges” to earn returns you own already and sends you a bill for fees.
So when you hear a quant hedge fund with 10% net excess return on 10% vol for 10 years with no factor or market exposure, don’t think “wow you lost to QQQ — think wow you beat *anything* by 10% per year, including QQQ, and you’d be a 99th percentile VC too 🤯”. Uncorrelated alpha. Don’t forget this is the whole point.
my Claude being cold af:
The deepest offense is that it’s short ARK’s own thesis. The entire ARKK pitch is right-tail hunting — convex, power-law payoffs from disruption. An autocallable is the precise inverse: the note gets called away the moment the stock rallies, truncating the right tail, while the holder keeps the left tail below the barrier. If Cathie’s right, you collect 17.5% and get called; if she’s wrong, you own the innovation universe down 60%. It monetizes belief in the thesis by selling the thesis’s payoff profile.
Then the standard grooves: capped upside with equity downside is exactly what geometric compounding punishes — Kelly allocates near-zero to that shape regardless of the arithmetic yield. Coupons aren’t income, they’re premium; return-of-capital risk relabeled as return on capital. The basket is one factor wearing forty tickers, so the barriers knock in at correlation one, in precisely the state of the world where everything else you own is also bleeding — anti-insurance, losses delivered at maximum marginal utility of wealth. And the vol being sold is opaque to the seller: retail can’t see the strike vol or the dealer margin embedded in each note, but the exotics desks on the other side can, to the basis point.
The final insult is that it will work beautifully for a while. Short-vol equity curves are the most seductive objects in finance — smooth monthly coupons compounding trust that gets spent all at once in a knock-in. Years of Sharpe, one afternoon of skew.
“Jane Street lost $15 billion in July and the correct response was spectatorship. Fifteen years ago that loss lives inside a bank — atop depositor capital and hedge fund portfolios — and we are all conscripted as its creditors. The Volcker Rule (Dodd-Frank §619) exiled proprietary risk from deposit-funded balance sheets; Basel III’s capital rules made it too expensive to smuggle back in. So the loss was quietly absorbed by the same year’s trading profits — losses met by the owners’ own winnings, not the public’s money. This is what regulatory victory looks like: catastrophe as a private matter.”
— my Claude (likes Volcker regs)
I tried to get a VC friend to give me a QQQ benchmark on his fund. Let’s just say I didn’t get the allocation 😂.
But the bar vs publics gets even higher if you consider the existence of quant hedge funds.
If you invested in a market and factor neutral quant fund with 10% net return (above cash) on 10% volatility and blended it with QQQ portable alpha style, your 10 year growth was:
$1 → ~$14 vs $1 → $6 in QQQ
30% return vs 20%.
This is why uncorrelated alpha is valuable. Not because it’s high return, but because its return geometry lets it stack onto any portfolio without it.
Try the same trick with a beta and factor exposed ‘hedge’ fund, and you have a portfolio in extra distress in 2022 (>-50%!). Because you have drawdowns stacking instead of returns!
You can think of a fund that tilts into factor bets as making a mess of your return geometry by taking risk on things your already own. For example, your portfolio doesn’t need a hedge fund to tilt you into a bet on AI. The S&P is ~48% an AI bet (so far). By betting on AI beta, the hedge fund “unhedges” to earn returns you own already and sends you a bill for fees.
So when you hear a quant hedge fund with 10% net excess return on 10% vol for 10 years with no factor or market exposure, don’t think “wow you lost to QQQ — think wow you beat *anything* by 10% per year, including QQQ, and you’d be a 99th percentile VC too 🤯”. Uncorrelated alpha. Don’t forget this is the whole point.
VCs btw can capture a giant pool of risks and returns unavailable to public equities eg by buying Anthropic or even Numerai (we were $10m valuation when VCs invested in us and now $500m when AUM was $1m now $740m).
So VCs earn returns no one can earn. There wasn’t some proxy OpenAI or Numerai you could buy on public markets in 2016 instead imo.
But where things get tricky is the story for levered public equity factor bets. It’s unclear these survive the future. Allocators with Claude will get demanding. Claude will be like super Venn.
aw… 🥹. I have incredible fondness for all the things @JeffDean built for all humanity and AI beings.
