@systematicls
Account based inSoutheast Asia
About this account
- Account based in
- Southeast Asia
- Connected via
- Southeast Asia App Store
Account-level information from X, not a live location or the device used for a specific post.
All in @openforage. I thrived in all of the largest hedge funds managing systematic investment processes.
Joined April 2024
- Tweets5.5K
- Following69
- Followers62K
- Likes9.8K
This post was bad enough to inspire a short article:
nitter.cf/systematicls/status/21…
I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem.
As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)!
I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work.
It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results?
88ms => 1.5ms
150K allocs => ~500 allocs
Incredible right? Nope.
My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path.
This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput.
The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity.
Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.
What this says is that our internal image of ourselves determine how much we enjoy our lives.
Happy Internal Messaging:
1. I have a lot of fun in all that I do
2. i am high energy
3. I am the master of my fate, captain of my soul
Sad Internal Messaging:
1. I am all alone
2. Everything in life is boring
3. I have no control over anything in life
I can't understand how professional software developers are so attached to days of manually coding.
Was objectively a top-tier software engineer (e.g. hired as a quant dev to build high performance systems) and honestly not having to type another line of code feels like such a good riddance.
People being so forlorn about agents being good at coding is insane to me. You can continue to ride horses as a hobby, no one can take that away from you, but you should be happy for everyone else to have engines.
Coding has and should always (just) be a means to an end. The product was and always *should* be the point. It's undeniable that agents are just objectively better at getting to the point now.
Go ahead and manually write code as a hobby.
Even for quants, as a pm there are many ways you can hurt your career from piss poor decision making in high stress environments: mostly around trying to make the pain stop // chasing comfort // protecting an identity.
Making the pain stop: cutting the book rapidly after some drawdown and not being able to recover in the subsequent reversion.
Chasing comfort: sizing up at the peak of your strategies, sizing down at the trough of your strategies (see making the pain stop).
The above are papercut sequences where you see teams just drawdown slowly until management loses patience and eventually cut them.
Protecting identity: trying to hold on to the image as a perpetual rainmaker, and doubling down on a dying strategy/book without understanding if the drawdown is a result of alpha decay and/or some systemic change.
This is the blow up sequence, where teams will hit enormous drawdowns from doubling down, sometimes outside pre-agreed risk limits.
I've always been surprised that the most controversial (IMO) opinion in my book is something I've never been called out on.
It is my belief that you cannot be a successful professional trader (quant or otherwise) without having a genetic pre-disposition for the management of stress. A good analogy would be that only certain people have the disposition to be an ordinance disposal expert.
You can learn the math, and the markets, and the incentives, and the models and the economics, and the patterns and everything else. But what no one can teach you and in my opinion you have to be born with, is the ability to sleep at night while your entire professional career dangles from a very fragile thread.
Few people actually possess this disposition. People in academia (as a rule) have the opposite disposition, which is why in spite of their intellect, exceptionally few academics have been successful in trading.
One thing I’ve realized about life is that when people lose
there are people who want to only talk about the loss
and then there are people who want to only talk about how not to lose the next time
and they might as well be speaking in different languages.
Me: When it rains, i want you to cover me with an umbrella.
Anthropic: There are two halves to this. And the half where the user does not want to get wet is the one that matters.
OpenAI: The user enjoys using umbrellas in the rain.
I can tell you that nobody from WQ would ever spell it as 2 words.
At this point, if you don't believe that Fable/Astra is generalized intelligence it's most certainly a skill issue.
Increasingly, it seems the primary value of a human is to pick a series of options in an infinite sequence of options.
Life is chosen suffering - that's all it is, we have to pick the suffering that gives us the most meaning in exchange.
IF we do not pick, we are assigned a form of suffering; e.g. mediocrity in your career is a form of suffering. Being obese and unfit is a form of suffering.
To live is to suffer, but being able to pick your choice of suffering is a form of bliss.
In a knife fight, one guy dies on the street and the other guy dies on the way to the hospital.
A simple mental framework for how long an agent takes to complete a task:
A task can essentially be broken down into the number of tokens by the implementation model x its harness needed to achieve it.
Consider a very difficult task, a very smart model might need to iterate 10x before it gets it right, and might require 10mn toks.
A less smart model might need to iterate 1000x before it gets it right, and might require 1bn toks.
A dumb model will never get it right even if you let it iterate 1mn times, and will require an infinite toks.
--
Between the very smart model and the less smart model above, if the very smart model outputs tokens at 10 tok/s and the less smart model outputs tokens at 1000 tok/s, they actually complete the task at the same time.
So it's not JUST about intelligence.
The same thing is true about costs, it's temping to look at $/mn toks and state that a model is cheaper than another, but what you're actually interested in is the $/difficult task for the work you are doing.
The "more expensive" model might end up using $100 for $10/mn toks for 10mn toks to complete the task, while the "less expensive" model might end up using $1000 for $1/mn toks for 1bn toks to complete the task.
--
Lastly, there are some tasks where anything LESS than frontier (e.g. Fable) literally cannot iterate towards the correct solution.
This is why frontier models continue to be very attractive even if cheaper models are available.
Until a supermajority of "very hard" tasks can be solved by OSS models, OpenAI and Anthropic will likely continue winning over more and more prosumers.
--
Bonus: There are some domains that are "so difficult", which is another way of saying that the solution space is so far away from the training data available to the models (e.g. quant finance), that you will literally need to provide iterative feedback for the model to arrive at a reasonable solution.
One thing you realize is that most ultimatums in life are binary, if you fall long enough, it's up or out.
Every man faces this demon on things he really cares about at least a few times in his life.
You have to stare into the abyss and really contemplate if you're willing to be out, if not, the only other option is up.
Then, you just have to make it work.
Heard this first from @ScottPh77711570 and its been in my mind ever since
seconded
I HAVE THREE INTERNS THIS SUMMER.
The first intern is brilliant but gets exhausted easily and is absent from work a lot.
The second intern shows up every day but has at most mediocre intellect.
The third intern gets results quickly and presents very well, but is a total bullshit artist whose work cannot be relied on at all.
The first intern is ChatGPT GPT-5.6 Sol Pro. The second intern is ChatGPT GPT-5.6 Sol on Extra High thinking setting. The third intern is Anthropic Claude Claude Fable 5 Max.
1/ Synth's HFT team has found that @Kalshi’s 15-min BTC market is beginning to predict subsequent BTC moves on @binance and all other CEXs.
And strength of this predictive relationship has grown materially throughout 2026 (see image).
@Kalshi is becoming the venue for BTC price discovery on short time horizons.
🧵
This is a fair take and a counterpoint to my meme.
gm @systematicls, totally understand why it can be easy to assume this is just "Binance + Kalshi is being manipulated". It's a reasonable instinct given how manipulation has played out elsewhere on places like Poly, but I'm positive it isn't the case here.
this critique would certainly make sense if Synth had sampled only the final 60 seconds of each 15-minute market, since that's the actual window feeding the resolution TWAP. But they didn't. Synth sampled across the full 15 minutes.
TBH if manipulation were as widespread as some suggest, these markets simply wouldn't have scaled the way they have. Retail wouldve burned out long ago, MMs would've pulled back liquidity against what 'toxic flow', and volume wouldve collapsed, which is exactly what we saw happen on Poly. Their short-term crypto markets were rife with manipulation earlier this year, and it led to a real, visible migration of users from Poly to Kalshi as a direct result (image attached). kudos to the @SynthdataCo guys here, they were actually among the first to call this out publicly, well before it became common knowledge.
so why has Kalshi held up better? id attribute it to a few concrete design choices rather than luck:
1. a 60-second TWAP for resolution pricing (versus Poly's single snapshot)
2. KYC and surveillance that can actually ban bad actors (yes market manipulation is a crime) +
3. an aggregated oracle that weights many CEXs rather than relying on one, notably excluding Binance, whereas Poly relied on just Binance alone
none of this means prediction markets are perfect oracles of truth. theyre not. but I do think their inherently forward-looking nature adds a layer of richness to price discovery that spot prices alone don't capture.
