ByteSmith 𓂆 @dbytesmith

This is really the problem of the golden hammer. Unfortunately AI is one of the greatest golden hammers we have ever created, and because a lot of the pain can just be smoothed over by "no mistakes" we find ourselves with terrible interfaces

dax @thdxr

15 hours ago

mcp is so tied to the idea of 1 process = 1 session there's some things in the spec that force you to spawn a new MCP server for every active session you have ridiculous overhead - this is why i kept saying it's too early to standardize

102 28 935 115K 145

@Maverick_Quant @octopus_x_trade

@jace48 Yeah, that would be something to look forward to 😅

⚠️🚨⚠️ Before “AI/LLM”, I was apparently a caveman. After “AI/LLM”, I am apparently an enlightened digital lifeform 😄 Yet I still need food, sleep, health, peace of mind. I still get tired. I still make mistakes. I still have gud n bad days. The funniest part is watching folks talk as if access to an “AI/LLM” made them the culmination of human evolution. Bro, chill. Your ancestors crossed oceans, survived famines, wars, plagues, and built civilizations with tools we'd laugh at today. They're probably looking at us arguing over prompts and having a good laugh. Humanity remains as fragile, replaceable, and temporary as ever. We just have better big boys toys now.

It's always a new learning experience when you read about product optimization blogs - and this won't disappoint you! zed.dev/blog/zed-decod…

Vivek Galatage @vivekgalatage

20 hours ago

Continuing the deep dives into the data structures that power the editors, here is an excellent article and a companion video to @zeddotdev's rope implementation! zed.dev/blog/zed-decod…

1 9 65 9K 40

@Nick_khandelwal @Souvik131 @dhruvbewtra

@indaiportfolio @Souvik131 @shreyanshdaga 🤝🥂

I thnk mayb d ans lies in d "deep research and projections" part. every1 hv access 2 d same data as public data has been commoditized for a very long time now. so i think d edge ws rarely about access 2 data. n tbh, internal affairs, privileged access n insider info have always remained d elephant in d room staring all of us in the face. d interesting part is how 1 reasons abt d info readily avlbl 2 dem, wht dimensions dey consider, wht assumptions dey make n ultimately how dey arrive at an inference. individuals/ai/llm's can hv a look at d same company n walk away wid completely opposite/contra/diffrnt conclusions/infrence. n i thnk dats d obscure corner whr d edge likes to hide nd whn exposed it usually stop being 1 lik it nvr existed in d 1st place Bt given d apce at which entities n enterprises r onboarding AI/LLMs, I do look forward 2 d day wen info asymmetry shrinks 2 levels we've nvr seen b4. and hoping dat d future models continuously do training or adapting on live streams of entities n enterprises n public info and simultaneously making trading decisions on top of it, and d time gap btwn info becoming public n being reflected in price may eventually approach 0, those are going 2 be sm crazy days 2 witness. 😅

@Souvik131 @shreyanshdaga Hi there 👋. Cool stuff 🚀, looking fwd 2 seeing how dis evolves.

@Souvik131 Interesting stuff 🚀, will be following this 1 as well as i usually do wid your works.

Glad this has finally been picked up. "Just add io-uring support" is 1 of those things dat sounds simple until we actually strt working on it. 😅 Looking fwd 2 contributing as well, quite a few intrsting prblms 2 solve here.

tokio's copy_bidirectional() is still just a read/write relay loop under the hood. kernel → userspace → kernel, for EVERY byte. that's 2 copies + 2 ctx switches per direction and the app isn't even touchin the payload, no parsing no decrypt nothin linux has had splice() since forever, and io_uring is right there too. And yes i'm aware of tokio-splice2 existence [github.com/hanyu-dev/toki…] and i know fasterthanli did the [fasterthanli.me/articles/remot…] writeup touchin on this too. is it portability? complexity tax? or is memcpy just stupid fast nowadays and nobody cares anymore for real. tbh curious if anyone actually benchmarked this on real traffic (small bursty msgs, not some iperf3 bulk transfer flex) not just vibes or just assumin its free perf 🦀💀

Useterminal.in We have revamped the website, keeping it free. Any feedback is welcomed 🙏 We are still perfecting the code and the goal is to keep the time minimum between dissemination of information on exchange and delivery to final user. Cc @ajit013

@doconclock 🫂

I personally feel that,⚠️backtesting derivatives is walking backwards trying to reach forward. the path that happened is one path, market doesn't care. why not build something different: precompute every plausible path a derivative can take, offline, across the full variable space. IV surface, skew, term structure, higher order greeks, vol models (SABR/local/stochastic-local), rates, carry, dividend adjustments, OI concentration, GEX, flow, futures basis, RV/IV regime, vol of vol, macro, geopolitical sentiment. all of it. combination space is Factorial Scale. billion/trillion/septillion/googol/googolplex of paths. GPU/CPU cluster job. then in live market you stop predicting. you look up. current signature hits the table. here are all paths matching this signature. here is their terminal state distribution at DTE0. And as market shifts you don't hold the old path. you hop. new signature, new cluster, new probability map. continuous, real time. Backtest is archaeology. this is a live probability map that updates as the market breathes. storage and compute are solved problems if we have the budget. the real unsolved problem is matching a live market signature against a billion/trillion/septillion/googol/googolplex-path table in low single digit μs to low double digit ns. and that's not a database problem. that's more like FPGA, RDMA, HBM, HPC, custom kernel with bypass or OpenOnload, NVIDIA BlueField/ConnectX DPU, 400GbE NIC/Fabric Interconnect etc territory. anyone working on this?

yeah dats actually a fair point, and a good food for thought. thanks man. wat I am struggling wid z whether d projected realized path contains all d info, or merely d final projection of a much higher dimensional state. 2 vry different paths can or may look vry similar while being driven by vry different positioning, liquidity, volatility n flow regimes. mayb I m overthinking it, bt dats exactly the rabbit hole I m struggling wid. Like How much info does d realized path actually contain n in wat fidelity? Is price/iv/ir/basis d state itself or merely a projection of a much larger state space? the crude raw intuition of mine suggests dat every new trade, quote, flow, event, or volatility change should continuously update our confidence in d assumptions behind d path. or mayb dat process collapses back to d realized path anyway. or maybe it completely breaks d assumptions n forces a different view of d state space. guess have to still try to figure that part out...

> You could add some microstructure data, but then we enter the realm of microstructure alphas that's not really the domain for retail.. > But I feel this is the right way to backtest - Yeah, True 🎯🤝 from your quoted thread (my bookmarked 😅) > This is a simulated option chain evolving through time that naturally aligns with carry & basis - dats actually where my mind was goin, but time alone feels like only 1 dim of the chain. I keep wondering & thinking whether things like order flow, T&S buckets, IV hotspots, OI migration across expiries, cross expiry (current/forward/leap/yearly etc) positioning, news, earnings, macro and geopolitical events should also be part of the state representation. maybe dats exactly where d combinatorial explosion begins or maybe d reason d search space explodes is because we are trying to compress all of that into a much smaller representation, don't know exactly or partially yet, but want to find out or let anybody familiar to this share their experiences.

> very depressing results on backtests 😅 Everything is negative EV do you think higher fidelity data with more dimensions would change the outcome at all? as most backtests operate on a quite low dimensional representation (compressed view of reality) of the market compared to reality.

@jace48 Sir 🙇, consider me put in my place🥹

@SachaSucha 🤝🫂

Bhai, I never take ur words lightly 🫡🙌. dey usually carry wisdom & experience I don't have yet 🥹 And tbh thank you for sharing this (genuinely). i am probably behaving like just a stubborn engineer (childish inquistive behaviour) repeatedly banging his head against a wall 2 see whether d wall breaks or my head breaks or something interesting emerges in between 😅