+1 - the biggest issue is not being able to fine tune the llm to learn the specifics of how to make a tool call better over time, which an approach like this can bring to the table.
I was using DeepSeek Coder and Claude for verilog quite successfully back in June - saved me a lot of time creating testbenches. Probably better now. I need to try the latest Gemini 1.5 pro to see how it does. I used DeepSeek Coder to create a testbench for some legacy verilog code that was poorly documented - it did quite well, better than I was expecting. It even collected repeated code into tasks which kind of surprised me.
I hadn't even considered trying to use ChatGPT with Verilog. I strongly suspect it will get in the way. My design is pretty darned small, just a grid of uniform cells, with not much else in the way of I/O, etc.
I skimmed through the linked github and reddit post, and maybe I missed it, but this looks like the kind of strategy that gets blown up when a negative black swan occurs. They're "harvesting premium" by selling options, based on the fact that "implied volatility is typically higher than realized volatility on average." Which is true, until it isn't. Like, yikes dude.
For example, the author recommends spreading selling options across multiple sectors to avoid getting blown up by any one sector moving against you, implicitly assuming no correlation across sectors. But how many times do we have to relearn that lesson that correlations do happen, and more frequently than we expect.
One of the commenters in the reddit thread even says that - they retired just before 2020 b/c all their Monte Carlo simulations showed they would have plenty to retire on, unless a highly unlikely correlated whole-market crash occurred. Which happened in the same year after they retired, blowing up their retirement planning.
Absolutely. I call it the options fallacy. It gives the illusion of risk free money but in reality it’s just capitalizing on a particular market pattern. There’s no free lunch.
Yeah selling options makes you an insurance firm, so you should have a better handle on valuing risk than your counterparty or you are probably the fool.
It trains ML models based on historic data and custom features and then uses them to generate a kind of intelligent indicator between -1 and +1. This intelligent indicator is then used to make trade decisions. Frequency is a parameter and can vary from 1 minute for crypto trading to 1 day for normal exchanges.
there are many more examples of slower/shorter term strategies than merely fundamental, there always have been...especially in the institutional space where the race to zero is functionally over.
I guess that applies when you're not able to fine-tune the LLM you're using. Presumably Anthropic has a lot of data too.