Supreme Leader Wiggum

If I were to stream… and it was mostly just me doing my day to day work, while talking to chat, would anyone watch? Setup isn’t going to be anything stellar, no real agenda - I’d pretty much just go do stuff across 5 or 6 repos in parallel & talk through what I’m thinking

The backtest uses online updates. So I label the data. Train model. Then use point in time replay, so backtesting it does online training since that’s just part of the system. If I just used the same checkpoint and Ran the same backtest twice, the second time it would be improved due to online parts.

@ScriptedAlchemy That makes sense. Last thing I’m curious about: do you train from simulated labels only after the full 2-6 week path is known, or do you also do online updates from interim mark-to-market premium path outcomes before the trade would have exited?

Fable stuck around just long enough to debug my 4 billion rows of training data, train a new financial model that was twice as profitable with half the drawdown, and during stress periods, maintain profitability. The model is now finally passing the 30% a month yield.

Use custom universe currently since I need training data on something. Opra data backfill is massive. Terabytes. So I use universe, then I have separate project that was experimental where I just use recommendation engine and bluesky firehose to rank accounts by accuracy and lead time. Then use that to build reputation database and discover instruments through social data for potential use. If don’t have to train on it all, then I’d just open socket connection that watches top 500 or top 1000 stocks by market cap. Let online training work it out in sparse heads/prime it with historical

@ScriptedAlchemy How are you finding the Stocks to watch? Do you scan a custom universe daily? Also any chance you share it?

@HashemKhalifa Model figures out weighting on its own. I create real simulation using OPRA quotes and trades, with slippage and latency. Little approximation since I have the order book. Lots of data labeling materialization. Watch for over fitting. Never test on data you trained on.

@ScriptedAlchemy One follow-up if you’re willing: when you optimized for 30% monthly return, did you train that as a direct portfolio simulation objective, or use weighted per-example labels that approximate monthly capital-cycle return?

The real challenge lies in managing drawdown and reversals. The model should be nearly equally profitable in either market direction. I utilize opra data, Bluesky data lake firehose for real-time social data ingestion, and employ ByteDance’s recommendation engine design for social usage. The entire system leverages ByteDance’s real-time sparse weight heads to facilitate online training. I train the model offline, essentially to prime it. Once online, I don’t need to retrain new checkpoints over time. Consequently, the model weights are fine-tuned in real-time.

@ScriptedAlchemy This is incredibly helpful, thank you. The fixed horizons part clarified the missing piece for me.

@lazyeval It won’t work. I just have Linux so cursor update patching doesn’t work well on Linux.

Hello darkness my old friend

@oriSomething I am national security

@ScriptedAlchemy Please don’t use it. It will damage national security 😱😱😱

@satellitedown Pretty easy with options trading tbh. Stocks. Not a chance

@ScriptedAlchemy yeah bro so I basically just told my computer to make me infinite money with like the stock market and stuff and it actually outperformed every financial instrument and institution in all of history

Profit off the appreciation on the premiums. Cycle the capital every few weeks into new premiums. No fixed horizon. Just make money with it 2-6 weeks then cycle. Under hood model used fix horizons as overlays to create matrix of horizons that get weighted, at least one iteration did. Might have folded it into unified. The model was trained exclusively to return 30% a month as its primary objective

@ScriptedAlchemy For your LEAPS calls/puts model, was the main training target forward option-premium return over fixed horizons, or realized strategy P&L after your holding/exit rules?

@julien_c Link to 10 trillion param local model plz.

run local models TODAY

@TimJayas No because they won’t spend 6 months telling the world it’s too dangerous to release

If GPT-5.6 is on same level with Claude Fable 5 will it be banned too?

@HemantDotDev Get banned

Everyone says ChatGPT 5.5 is insane. So what’s one thing Claude Fable can do that ChatGPT 5.5 simply can’t?

Guys they nerfed fable biblically.

At least I can go to LinkedIn to see what was, experience it again. Hindsight blows.

@coolemur @OpenAI @thsottiaux @sama just update 5.5 checkpoint. We don’t need to tell anybody anything. It’ll be our lil secret.

@ScriptedAlchemy @OpenAI

OpenAI should drop 5.6 right now. No better time.

@softDev23 @araseb_ How tf do you get Gemini to successfully make more than 2 tool calls before just imploding on itself and repeating the same word for the next million tokens out?!

@araseb_ Gemini helped build most of my products so far, but that is changing as I have switched to Codex and now Claude. I just feel Anthropic and OpenAI are actually shipping more updates and tools that devs want to use. And with Fable 5, who knows where we are headed with Claude.

Be honest builders If ChatGPT, Claude, Gemini, and Grok disappeared tomorrow, which one would have the most impact on your work?

@araseb_ Siri

ChatGPT has the normies Claude has the devs Gemini has?

@pcshipp They took fable from us

I'm a solo founder. Scare me with one word

@ak_kim0 That’s what wrote and designed the initial system. With GitHub MCP to review every commit….

@ScriptedAlchemy Please try the web based 5.5 pro extended. Copy/paste context