That's being pretty well solved by the right agent orchestration nowadays imo. My point is that it's just next to ridiculous to call a 200K model suitable for long-running tasks. It can't even sketch a task/agent hierarchy without compacting (=losing thought). Grok Build is ok with that though.
With all due respect, the video is slightly misleading since it doesn't show how intelligent or long-running it is. What I see is a simple TS/UI prompt followed by 4 reads, which took 30% of the context window. Maybe the model is great, but the context window is damn small, even for orchestrating stuff well.
It compacts very well, I have to admit, better than most of the other models, but it's just not suitable for complex problems.
It might work for short-running, simple tasks that benefit from speed of delivery, which is the exact opposite of what's being marketed.
Although I'm very excited about Grok Build, I love the approach of shipping stuff early and letting it mature. I am happy I'm that early.
@daptonai@xai No need to test to understand that the 200K context window coming with Composer is 5x less than what people are used to nowadays, which is still not enough for long-running tasks ngl
The thing is amazing: fast and deep, but the pricing is unrealistic. If it were x3 more credits/lower price, it'd be competitive. I'm a max subscriber of all of them. I've used up the monthly Grok Build credits in 3 days (which is $300 ngl) - on max effort ofc, pipelined agents, extensive coding of complex stuff with low-latency engineering and lots of math; with more rational use, I reckon it'd be in around a week. Let's say 1w/4w it's 25% efficient. Codex worth $100 is okay; I'm usually taken weekly limits with 1-2 days left to go, the same for the $200 Claude Code (6/7 days, 85% efficiency). Given Claude's price, it doesn't make Grok look very good. I'd prefer x3-x5 Grok credits at this price, which makes a lot of sense to me given its parallel work. Until then, Claude Code is still the best RR on the market.
Curious about how market makers actually supply liquidity? The Grossman-Miller model helps get that! The math might seem a bit daunting, though.
Check out the free article and the app that lets you see how the economics unfold in real time
https://t.co/sxI5vlshQ0
Just posted a very quick and simple derivation of equilibrium spreads for GlostenโMilgrom Market Maker before diving into the next level stuff.
Grab a refresher of market maker intuition here https://t.co/81SfqvufsZ
@LilQwantXBT FalconX does that, maybe there are more. it'd be cool to have a number of low-tier cex-es/new perp dex-es in a thing like that. Having one unified liquidity account is indeed cool idea
@lopezdeprado called these the deflated Sharpe. People know the "Advances in Financial ML" book, but there are more great reads like this one - short, easy to digest and practical, and it has better charts in it to illustrate the point (the one above is just pure lack of edge with zero mean SR)
https://t.co/uRdMINBoSY
@HF_Trader Occam's razor. Never fails.
That's the very reason why I've started organising and sharing mm-related content. Simple microstructure models do work, and this knowledge is essentially practical and hugely underestimated.
@zkcno considering the decreasing/low realized vol (30% now), that high corr really says we're now into broad accumulation and early trend strength, pure breadth thrust
lots of cool papers out there, but this one is not only a waste of time but just harmful. 2019 data - an eternity ago, they celebrate 90% "accuracy" and everybody knows that's nothing if the rest 10% of losses can be total meltdown, classic backtesting fantasy: survivorship bias, look-ahead bias, no tx costs - complete "academia package" is there.
I opened it to find failed models, but only "successes" are here. Damn if it's not selection bias, then what?
Don't waste time and learn Market Microstructure; it will give you a fundamental understanding of how things are.