Hello there - been playing around w your initial agent - very insightful. I know you prescribe Claude as the default LLM - correct? There does seem to be a lot of variation in the outputs dep on the LLM used, and also spot price always seems to come up wrong on first iteration. Lots of follow on questions and checks for accuracy needed, as you indicate. But these are good exercises to confirm and check one's own biases etc. Re the latest I think the Lassonde curve thing is very promising interested to see that. For me one of the biggest questions is how to ballpark the timing of exit for an exploration/development stage stock..i.e. at the top of the curve, but also coinciding w really bullish (high gold price) sentiment. So are there historical correlations or indicators that could be embedded in an agent to signal that? eg. combine (a) normal valuation metrics based on assets/share price, NAV, etc... (2) (bearish or bullish macro sentiment/momentum and (3) stage of assets along Lassonde Curve. It seems to me that if those three things could be blended in some way it would be very helpful to make selling decisions for companies at this stage. Thanks
I use reasoning every time. Claude sonnet 4, but reasoning/thinking, low temperature is the more important request. Variance goes down with specificity in the agent itself. If you notice the email is slightly different than this post... its because I accidently published the wrong output as they looked very similar. Convergence is test criteria.
So far the agents are somewhat price agnostic. I call the latest prices just so that if the document is saved, you'd know the context in terms of price. I think when to purchase is an issue of charts in the short term. For this project, I'm focused on gathering a high level opinion of risk of company action. This is highly opaque to the outsider.
Short answer is, yes. It's possible, but if you try to do too many untestable things with LLM, you'll get cooked by ambiguity. Eric Sprott could have the wrong thing for breakfast and agree to 5 deals by afternoon that are way lower than we all thought. We'll all think he was thinking one thing, but it was breakfast. A complete outlier. I'm warry of too much needless precision.
Lassonde curve works, it just doesn't look good yet. I need to do some R&D to make it look good. Ultimately its a visualization of my sum of parts sourcing from 43-101. I don't think the forecast needs to be done with the LLM, only the data sourcing and standardization. Companies report 43-101s 30 different ways, tonnes, oz, P&P, M&II, combine this that and the other. Once I have the data, its python to the forecast and share price. I don't want LLMs futzing around with valuation. Same with DCF. I want the LLM to develop and prescribe inputs to a static forecast DCF model. Handle the most subjective stuff.
Hello there - been playing around w your initial agent - very insightful. I know you prescribe Claude as the default LLM - correct? There does seem to be a lot of variation in the outputs dep on the LLM used, and also spot price always seems to come up wrong on first iteration. Lots of follow on questions and checks for accuracy needed, as you indicate. But these are good exercises to confirm and check one's own biases etc. Re the latest I think the Lassonde curve thing is very promising interested to see that. For me one of the biggest questions is how to ballpark the timing of exit for an exploration/development stage stock..i.e. at the top of the curve, but also coinciding w really bullish (high gold price) sentiment. So are there historical correlations or indicators that could be embedded in an agent to signal that? eg. combine (a) normal valuation metrics based on assets/share price, NAV, etc... (2) (bearish or bullish macro sentiment/momentum and (3) stage of assets along Lassonde Curve. It seems to me that if those three things could be blended in some way it would be very helpful to make selling decisions for companies at this stage. Thanks
I use reasoning every time. Claude sonnet 4, but reasoning/thinking, low temperature is the more important request. Variance goes down with specificity in the agent itself. If you notice the email is slightly different than this post... its because I accidently published the wrong output as they looked very similar. Convergence is test criteria.
So far the agents are somewhat price agnostic. I call the latest prices just so that if the document is saved, you'd know the context in terms of price. I think when to purchase is an issue of charts in the short term. For this project, I'm focused on gathering a high level opinion of risk of company action. This is highly opaque to the outsider.
Short answer is, yes. It's possible, but if you try to do too many untestable things with LLM, you'll get cooked by ambiguity. Eric Sprott could have the wrong thing for breakfast and agree to 5 deals by afternoon that are way lower than we all thought. We'll all think he was thinking one thing, but it was breakfast. A complete outlier. I'm warry of too much needless precision.
Lassonde curve works, it just doesn't look good yet. I need to do some R&D to make it look good. Ultimately its a visualization of my sum of parts sourcing from 43-101. I don't think the forecast needs to be done with the LLM, only the data sourcing and standardization. Companies report 43-101s 30 different ways, tonnes, oz, P&P, M&II, combine this that and the other. Once I have the data, its python to the forecast and share price. I don't want LLMs futzing around with valuation. Same with DCF. I want the LLM to develop and prescribe inputs to a static forecast DCF model. Handle the most subjective stuff.