Morven

building weird things at the intersection of AI, quant & markets

[Morven @xmorvenn]

@RitOnchain appreciate venus

[Morven @xmorvenn]

Jane Street gave 3,757 teams 12.3 GB of its own trading data and $120,000 to prove they could predict a market that hadn’t happened yet.

the winner took $50,000.

but the money is probably the least interesting part.

Jane Street released anonymized data derived from its production systems, then built the competition so the usual backtesting tricks were almost useless.

inside the dataset:

1. 79 anonymized features and 9 responders, with one row for each instrument at each timestamp

2. the target was responder_6, scored with sample-weighted R² around zero. predict nothing and you score exactly 0. build something worse than nothing and you go negative

3. roughly 4.5 million rows in the training/test data alone

4. new symbol_ids appeared in the future test set, meaning your model had to predict instruments it had literally never seen before

5. submissions were frozen on January 13, 2025. Jane Street then evaluated them for another six months on market data that did not exist when anyone wrote their model

6. every notebook had a hard runtime limit. 8 hours during development, 9 during forecasting. a brilliant model that cannot finish is worthless

25,446 people entered. 3,757 teams competed. 4,294 submissions survived the constraints.

and the useful part is still sitting there for free.

the dataset, evaluation API and winning notebooks effectively give you a blueprint for testing whether a trading model found an actual signal or just learned how to explain the past.

the $120,000 is gone.

the experiment is worth more.

[Rossst.03 @Rossst_03]

https://t.co/zeqJCFDNYZ

[Morven @xmorvenn]

@xdurdenn thanks, bro

[Morven @xmorvenn]

@scream_crypto thanks scream

[Morven @xmorvenn]

@RitOnchain this is real alpha on reddit

[Mr Green @themrgreenn]

In 1997, a developer publicly humiliated Steve Jobs in front of a packed room.

Jobs had just returned to Apple after 11 years away. Instead of fighting back, he admitted the guy was partly right — then explained the principle he believed could turn Apple into an $8–10 billion-a-year business.

And his answer had almost nothing to do with technology.

Most companies, Jobs explained, work backwards. They build something technically impressive, fall in love with what they created, and only then ask how they are going to sell it.

He had made that mistake himself so many times that he said he had the “scar tissue” to prove it.

His rule was the opposite: start with the customer experience and work backwards to the technology.

Jobs used the LaserWriter as an example. Inside the printer were a Canon laser engine, an Apple-designed controller, Adobe PostScript and AppleTalk. It was an extraordinary piece of engineering, but none of that was the reason he believed people would buy it.

When the first page came out, Jobs looked at it and realized they didn't need to explain what was inside the machine. They could simply hold up the result and ask one question:

Do you want this?

That was the entire point.

Customers don't pay you for how difficult your product was to build. They pay for what it does for them.

A developer tried to expose what Steve Jobs didn't know about technology.

Jobs responded by explaining something far more valuable about business.

[Morven @xmorvenn]

@RitOnchain alpha info

[Morven @xmorvenn]

@xdurdenn the compute dial changes everything

[Morven @xmorvenn]

GPT-6 Astra may have just collapsed the most expensive moat in quantitative finance

not execution, strategy discovery

i mapped the entire system into a 10-page technical paper:

> 300 parallel agents watch order books, options flow, SEC filings and macro events

> 8 specialist bots turn raw anomalies into executable hypotheses without letting one bot grade its own work

> 4 alpha engines hunt statistical arbitrage, volatility dislocations, factor residuals and insider clusters

> every strategy is forced through point-in-time data, purged walk-forward testing, Deflated Sharpe and probability of backtest overfitting

> capital moves through shadow → paper → canary → live, with an independent risk bot holding the kill switch

> the repository structure, hypothesis contract, alert schema and position-sizing logic are all inside

most trading bots are built to place trades

this one is built to kill bad strategies before they touch capital

GPT-6 ASTRA: THE AUTONOMOUS ALPHA FACTORY

10 pages. the entire blueprint.

[Morven @xmorvenn]

@marfinxx valuable research, thanks

[Morven @xmorvenn]

@0xzlato appreciate it, this is brutal

[Morven @xmorvenn]

GPT-6 Astra may have just collapsed the most expensive moat in quantitative finance

not execution, strategy discovery

i mapped the entire system into a 10-page technical paper:

> 300 parallel agents watch order books, options flow, SEC filings and macro events

> 8 specialist bots turn raw anomalies into executable hypotheses without letting one bot grade its own work

> 4 alpha engines hunt statistical arbitrage, volatility dislocations, factor residuals and insider clusters

> every strategy is forced through point-in-time data, purged walk-forward testing, Deflated Sharpe and probability of backtest overfitting

> capital moves through shadow → paper → canary → live, with an independent risk bot holding the kill switch

> the repository structure, hypothesis contract, alert schema and position-sizing logic are all inside

most trading bots are built to place trades

this one is built to kill bad strategies before they touch capital

GPT-6 ASTRA: THE AUTONOMOUS ALPHA FACTORY

10 pages. the entire blueprint.

[Roan @RohOnChain]

GPT-6 Astra is the most dangerous AI model right now.

It gives you AGI-adjacent reasoning.

That can discover new profitable trading strategies for you 24/7.

If you set it up correctly, you gain a personal hedge fund. https://t.co/Sam7AYOut3

[Morven @xmorvenn]

@themrgreenn thanks, Mr Green

[Morven @xmorvenn]

@xdurdenn appreciate you

[Morven @xmorvenn]

Access the research paper here - https://t.co/aAJXHD0jAC

[Morven @xmorvenn]

@themrgreenn only hard work will lead to this

[Morven @xmorvenn]

@RitOnchain thanks, man

[Morven @xmorvenn]

GPT-6 Astra may have quietly killed the idea that an AI model has one fixed level of intelligence

i went through the recurrent-depth research behind Astra and found something much bigger than “it thinks longer”

the same weights can potentially run at different computational depths

4 loops for something easy
16 for reasoning
32 for hard code
64 when the problem refuses to converge
same model
same parameters
different amount of internal computation
which means the next scaling law may not be:
more parameters → more intelligence

it may be:
more useful computation → more intelligence
I mapped the entire architecture into a technical paper because the implications are f*cking wild

GPT-6 ASTRA: THE COMPUTE DIAL

the next frontier may not be building a bigger model

it may be teaching the same model how deep to think

[Mikadzyki🌙 @Mikadzyki_NFT]

https://t.co/CjpmhF5hEt

[Morven @xmorvenn]

@xdurdenn real evolution, it's shocking

[Morven @xmorvenn]

@RitOnchain crazy sharing, thanks