> The tricky part about any machine learning model is that, when you are iterating through trillions of parameters, even a 99.999% accurate signal will amplify a ton of noise.
Reminds me of Friedrich Hayek's idea about information in The Use of Knowledge in Society. Once you have the full data, it's just a matter of applying logic to find the best strategy. Gathering the data is the hard part.
>In a complex system, profit and loss is about as good as you can get.
I think this is one reason prediction markets are particularly interesting. Check out some of the companies building prediction bots using LLMs like Mantic. Prediction markets need to resolve, so they provide a fairly clean signal, unlike price movements of stocks.
>The philosophy of raising as much money as possible for compute to train as large a model as possible and justify it through “well, the future applications will justify the spend” is not possible if the EWM is constructed properly — because the entire point is that profit is the signal.
If that's the case, how would wacky, out-of-distribution ideas be funded, like the internet or EVs? I'm sure the EWM's search space would've eventually led to the discovery of some signal showing that the internet could be useful, but I would think this would've taken a much longer time to uncover vs. someone saying fuck it and building it out.
1) Yep, it's kind of silly that everything boils down to unique, clean training data at some point. We're finally building new structures, but that's one of the funny parts of the original machine learning craze from the mid 2010s — the math has been known for a long time, but it was thought to be too impractical to use. Really, we're finally at the point of taking advantage of the exponential increase in hardware capacity.
2) My main issue with prediction markets is that they are turning complex questions into binary (ignoring the legal loophole issues for the time being.) The most crude, extreme example I can think of is Bitcoin binary up/down 5 minute markets — sure, you might build something that can make money on this coin flip over time, but what exactly is the broader economic purpose of building such a bot? You might get a good trading strategy out of it, but the nice part about a stock price is that it's a reflection of the price that actors with wildly different intentions are willing to trade at. A simple example of the utility of predicting stock prices over prediction markets for a company would be determining when to issue stock to maximize capital raised in offerings. (A company like ASTS, for example, that dilutes their stock into pumps can raise a lot more cash if they had an internal price prediction algo — after all, that is the original utility of the stock market. The money is supposed to flow in and out of the company, and even the utility of making efficient buyback decisions would be much greater than interacting with the synthetic market of binaries that aren't stock options.)
3) I think fundamentally an EWM would depend on data being present. I don't think that white space startups like SpaceX, for example, would fall under predictive capacity (or a different method of applying the model would be needed.) The main benefit, in my eyes, is that I tend to believe that the classic model of VC is dying out. More on this later, but given the sheer size of IPOs, and how much of the return has been privatized, the raise/dilute/scale/square away finances model doesn't seem to work. I think the BTC treasury company model, on the other hand, is a silly version of a much better idea — create a company mechanism that doesn't require raising money and dealing with varying viewpoint and diluting ownership, while shoring up the finances as quickly as possible. (More on this in my next post.) I define intelligence as extremely accurate pattern recognition based on priors. I don't think innovation is going to be replaced by artificial intelligence, it's not possible to mimic human intuition if we can't even explain it.
This post is definitely a sort of precursor to what I've been thinking about on the golf course for months, and the "bet on networks" investment model I've come up with. Rather than the VC all-or-nothing model, the "efficient frontier" of capital held should follow more of a barbell model.
2) I totally agree with your point. But what I was trying to get at was that companies or individuals building prediction engines (or as I call it simulators and super forecasting machines) for more qualitative questions like when the war in Iran be resolved can find interesting signal for what seems like a complex problem, finding a new level of abstraction that reveals a solution that may not have been discoverable without deep neural networks. My take is that there will be some interesting breakthroughs in this area in the near future.
3) I'm looking forward to the next piece. Curious what angle you will take on the BTC treasury topic and what your take is on VC asset class (especially as a VC myself).
I actually think innovation and creativity will get arbitraged away by AI. Both largely results from your consumption habit - this has been echoed by Pierre-Alexis Dumas and Rick Rubin when they refer to the subconscious. Is there real creativity in music or is it the different pacing of the same set of chords? I think AI will create more of the Kepler and Newton types of intelligence first, where pattern recognition plays a big part, before producing the Einstein level of intelligence where creativity plays a much larger role.
Oh, yes, I very much think the financial analyst role has been replaced by LLMs. My methodology for that mode of thinking is not testing predictions on markets, but just using them as infinitely read, turbocharged information processors to spit out takes where I use my own discernment as to whether they're useful or not. When one has sufficient expertise, one can extract insane value out of even base LLMs. For example, with club fitting, I talked extensively to Gemini to map out how to build my bag. The same goes for law, or any topic I'd normally deep dive a niche subreddit or forum or topical wiki for. LLMs are insanely good tutors when you understand a field well enough such that you just need a rapid fire perspective or a concept reframed in another way. When I was trading all of 2025, I had separate instances of Grok (to test out my twitter attention trade thesis) and Deepseek (to get around manual PC gates on ChatGPT) to pingpong ideas off of. Hilariously enough, Nick Denton, the old Gawker guy, was doing something similar to trade stocks, which he highlights in this interview.
My real worry is that this kills the entry level "intern to junior analyst" pipeline, bc why do I need a kid building excel models about industries they barely understand? Something needs to be put into play such that new grads are thrown a bone.
As to VC, I generally believe the era of huge fundraises at insane valuations to exit into markets or a large buyer is mostly over. When you have trillion dollar IPOs and series H's, inherently, it's an acknowledgement that the valuation isn't particularly "real" and control over the business matters the most. Smart founders (and something I am actively positioning myself to play a role in) will prioritize breaking even on revenue as quickly as possible, scaling through AI rather than labor (which is why fundraising required a ton of capital in the 2010s in the first place — hiring away talent from a big tech company was ludicrously expensive), and retaining control through profitably managing the treasury to avoid dilution will be the "new" model I see for venture capital. Certainly some of the big guys will always exist, but I think those are more pure logistics/hardware/compute plays. It's not going to be possible for trend chasers coasting off of crowding into rounds, like the current seed environment, to generate returns. Which is a good thing, the private market must get more efficient now that the public market is full of junk.
I think I'll write a separate post and think more about the AI consumption phenomenon, but I think I am in the ballpark of framing it around literature and "detecting" AI writing as I wrote here: https://x.com/edgefills/status/2059244492797137059
> The tricky part about any machine learning model is that, when you are iterating through trillions of parameters, even a 99.999% accurate signal will amplify a ton of noise.
Reminds me of Friedrich Hayek's idea about information in The Use of Knowledge in Society. Once you have the full data, it's just a matter of applying logic to find the best strategy. Gathering the data is the hard part.
>In a complex system, profit and loss is about as good as you can get.
I think this is one reason prediction markets are particularly interesting. Check out some of the companies building prediction bots using LLMs like Mantic. Prediction markets need to resolve, so they provide a fairly clean signal, unlike price movements of stocks.
>The philosophy of raising as much money as possible for compute to train as large a model as possible and justify it through “well, the future applications will justify the spend” is not possible if the EWM is constructed properly — because the entire point is that profit is the signal.
If that's the case, how would wacky, out-of-distribution ideas be funded, like the internet or EVs? I'm sure the EWM's search space would've eventually led to the discovery of some signal showing that the internet could be useful, but I would think this would've taken a much longer time to uncover vs. someone saying fuck it and building it out.
1) Yep, it's kind of silly that everything boils down to unique, clean training data at some point. We're finally building new structures, but that's one of the funny parts of the original machine learning craze from the mid 2010s — the math has been known for a long time, but it was thought to be too impractical to use. Really, we're finally at the point of taking advantage of the exponential increase in hardware capacity.
2) My main issue with prediction markets is that they are turning complex questions into binary (ignoring the legal loophole issues for the time being.) The most crude, extreme example I can think of is Bitcoin binary up/down 5 minute markets — sure, you might build something that can make money on this coin flip over time, but what exactly is the broader economic purpose of building such a bot? You might get a good trading strategy out of it, but the nice part about a stock price is that it's a reflection of the price that actors with wildly different intentions are willing to trade at. A simple example of the utility of predicting stock prices over prediction markets for a company would be determining when to issue stock to maximize capital raised in offerings. (A company like ASTS, for example, that dilutes their stock into pumps can raise a lot more cash if they had an internal price prediction algo — after all, that is the original utility of the stock market. The money is supposed to flow in and out of the company, and even the utility of making efficient buyback decisions would be much greater than interacting with the synthetic market of binaries that aren't stock options.)
3) I think fundamentally an EWM would depend on data being present. I don't think that white space startups like SpaceX, for example, would fall under predictive capacity (or a different method of applying the model would be needed.) The main benefit, in my eyes, is that I tend to believe that the classic model of VC is dying out. More on this later, but given the sheer size of IPOs, and how much of the return has been privatized, the raise/dilute/scale/square away finances model doesn't seem to work. I think the BTC treasury company model, on the other hand, is a silly version of a much better idea — create a company mechanism that doesn't require raising money and dealing with varying viewpoint and diluting ownership, while shoring up the finances as quickly as possible. (More on this in my next post.) I define intelligence as extremely accurate pattern recognition based on priors. I don't think innovation is going to be replaced by artificial intelligence, it's not possible to mimic human intuition if we can't even explain it.
This post is definitely a sort of precursor to what I've been thinking about on the golf course for months, and the "bet on networks" investment model I've come up with. Rather than the VC all-or-nothing model, the "efficient frontier" of capital held should follow more of a barbell model.
2) I totally agree with your point. But what I was trying to get at was that companies or individuals building prediction engines (or as I call it simulators and super forecasting machines) for more qualitative questions like when the war in Iran be resolved can find interesting signal for what seems like a complex problem, finding a new level of abstraction that reveals a solution that may not have been discoverable without deep neural networks. My take is that there will be some interesting breakthroughs in this area in the near future.
3) I'm looking forward to the next piece. Curious what angle you will take on the BTC treasury topic and what your take is on VC asset class (especially as a VC myself).
I actually think innovation and creativity will get arbitraged away by AI. Both largely results from your consumption habit - this has been echoed by Pierre-Alexis Dumas and Rick Rubin when they refer to the subconscious. Is there real creativity in music or is it the different pacing of the same set of chords? I think AI will create more of the Kepler and Newton types of intelligence first, where pattern recognition plays a big part, before producing the Einstein level of intelligence where creativity plays a much larger role.
Oh, yes, I very much think the financial analyst role has been replaced by LLMs. My methodology for that mode of thinking is not testing predictions on markets, but just using them as infinitely read, turbocharged information processors to spit out takes where I use my own discernment as to whether they're useful or not. When one has sufficient expertise, one can extract insane value out of even base LLMs. For example, with club fitting, I talked extensively to Gemini to map out how to build my bag. The same goes for law, or any topic I'd normally deep dive a niche subreddit or forum or topical wiki for. LLMs are insanely good tutors when you understand a field well enough such that you just need a rapid fire perspective or a concept reframed in another way. When I was trading all of 2025, I had separate instances of Grok (to test out my twitter attention trade thesis) and Deepseek (to get around manual PC gates on ChatGPT) to pingpong ideas off of. Hilariously enough, Nick Denton, the old Gawker guy, was doing something similar to trade stocks, which he highlights in this interview.
https://www.vanityfair.com/news/story/nick-denton-interview-thiel-musk
My real worry is that this kills the entry level "intern to junior analyst" pipeline, bc why do I need a kid building excel models about industries they barely understand? Something needs to be put into play such that new grads are thrown a bone.
As to VC, I generally believe the era of huge fundraises at insane valuations to exit into markets or a large buyer is mostly over. When you have trillion dollar IPOs and series H's, inherently, it's an acknowledgement that the valuation isn't particularly "real" and control over the business matters the most. Smart founders (and something I am actively positioning myself to play a role in) will prioritize breaking even on revenue as quickly as possible, scaling through AI rather than labor (which is why fundraising required a ton of capital in the 2010s in the first place — hiring away talent from a big tech company was ludicrously expensive), and retaining control through profitably managing the treasury to avoid dilution will be the "new" model I see for venture capital. Certainly some of the big guys will always exist, but I think those are more pure logistics/hardware/compute plays. It's not going to be possible for trend chasers coasting off of crowding into rounds, like the current seed environment, to generate returns. Which is a good thing, the private market must get more efficient now that the public market is full of junk.
I think I'll write a separate post and think more about the AI consumption phenomenon, but I think I am in the ballpark of framing it around literature and "detecting" AI writing as I wrote here: https://x.com/edgefills/status/2059244492797137059