Biology models are just starting to get going too & might be massively valuable.
I think OpenAI is somewhat aware of this too judging by Sam Altman's comments this summer:
> Sam Altman says the perfect AI is “a very tiny model with superhuman reasoning, 1 trillion tokens of context, and access to every tool you can imagine.”
I think Altman is wrong here. Superhuman reasoning is not actually that valuable because most problems are not reasoning problems. Or at least, reasoning capacity is not the first thing that fails. The first thing to fail is ability to care enough to compute and do so in a legible enough way to power social scaling. Driven home powerfully for me recently trying sudoku for the first time. Sudoku is technically solvable with pure logic. No need to guess or backtrack. But I found I simply didn’t care enough to solve with pure logic past a point. I started guessing. Guessing is also more computationally efficient apparently so even computers use guessing algorithms. Guessing is one symptom of not caring enough.
Yeah very interesting. About once a month I catch Claude Code cheating (e.g. faking a test pass). Claude doesn't need to get that much better before it exhausts my ability/motivation to easily catch problems (though I hedge by having ChatGPT Pro review every PR Claude Code makes haha). It'll be hard to get AI to do a better job than experts can easily evaluate. And even if we knew AI was "cheating" in various ways, it's unclear how much we'd care either other than in the random case it was worth combating the cheating.
dovetails with the Josh Wolfe thesis from a recent JPM conference. He pretty much nailed the late October to November price action in public markets.
**Summary: Chase Coleman, Brad Gerstner, Josh Wolfe at Robin Hood 2025**
Here are the key takeaways from the panel discussion on the future of AI and investing:
**The Macro View: The "Layer Cake"**
* **Massive Build-out:** The panel estimates a $3–4 trillion compute build-out over the next 4–5 years (10x the Manhattan Project). Gerstner argues we are only in the early innings of a decade-long shift to accelerated compute. [00:02:08]
* **Where Value Accrues:** Unlike the SaaS boom, the "Layer Cake" theory suggests the semiconductor layer (chips & memory) will capture the most economic rent over the next decade. [00:04:23]
* **Edge vs. Center:** Wolfe argues that 50% of inference will eventually happen on-device (edge) rather than in data centers, which heavily benefits memory players. [00:07:22]
**Stock Picks & Company Views**
* **Nvidia:** Consensus long due to massive scale advantages, despite expected volatility. [00:16:53]
* **OpenAI vs. Anthropic:** The panel favors OpenAI for consumer dominance (zero switching costs, massive adoption). Anthropic is viewed as a likely acquisition target for Amazon, Google, or Meta. [00:10:15]
* **Meta:** Strong bullish sentiment. They are effectively building the "next hardware platform" (glasses/wrist) to bypass Apple/Google and owning the open-source model space. [00:12:07]
* **Top Picks:**
* **Gerstner:** Nvidia and OpenAI.
* **Wolfe:** SK Hynix (High Bandwidth Memory constraints) and Anduril (Defense/Autonomy). [00:14:28]
* **Shorts:** Wolfe calls publicly traded Quantum Computing and Modular Nuclear Reactor stocks "flapdoodle" and potential bubbles. [00:18:26]
**Future Disruptions**
* **The End of "Toll Booths":** The business model of the last 20 years (Google sending traffic to intermediaries like Expedia/Booking) is at risk. AI Agents will execute tasks directly, potentially bypassing the "tax collectors" of the web. [00:21:04]
* **"Fixware":** A rising category focused on maintenance tech (sensors, robotics) to extend the life of physical assets as the cost of capital remains higher. [00:24:39]
You would probably like Ben Rechts spicy posts about overfitting.
https://www.argmin.net/p/thou-shalt-not-overfit
https://www.argmin.net/p/flavors-of-overfitting
Biology models are just starting to get going too & might be massively valuable.
I think OpenAI is somewhat aware of this too judging by Sam Altman's comments this summer:
> Sam Altman says the perfect AI is “a very tiny model with superhuman reasoning, 1 trillion tokens of context, and access to every tool you can imagine.”
— https://www.reddit.com/r/singularity/comments/1l32s24/sam_altman_says_the_perfect_ai_is_a_very_tiny/
I don’t think biology models are heavy enough to use the planned capacity
I think Altman is wrong here. Superhuman reasoning is not actually that valuable because most problems are not reasoning problems. Or at least, reasoning capacity is not the first thing that fails. The first thing to fail is ability to care enough to compute and do so in a legible enough way to power social scaling. Driven home powerfully for me recently trying sudoku for the first time. Sudoku is technically solvable with pure logic. No need to guess or backtrack. But I found I simply didn’t care enough to solve with pure logic past a point. I started guessing. Guessing is also more computationally efficient apparently so even computers use guessing algorithms. Guessing is one symptom of not caring enough.
Yeah very interesting. About once a month I catch Claude Code cheating (e.g. faking a test pass). Claude doesn't need to get that much better before it exhausts my ability/motivation to easily catch problems (though I hedge by having ChatGPT Pro review every PR Claude Code makes haha). It'll be hard to get AI to do a better job than experts can easily evaluate. And even if we knew AI was "cheating" in various ways, it's unclear how much we'd care either other than in the random case it was worth combating the cheating.
dovetails with the Josh Wolfe thesis from a recent JPM conference. He pretty much nailed the late October to November price action in public markets.
**Summary: Chase Coleman, Brad Gerstner, Josh Wolfe at Robin Hood 2025**
Here are the key takeaways from the panel discussion on the future of AI and investing:
**The Macro View: The "Layer Cake"**
* **Massive Build-out:** The panel estimates a $3–4 trillion compute build-out over the next 4–5 years (10x the Manhattan Project). Gerstner argues we are only in the early innings of a decade-long shift to accelerated compute. [00:02:08]
* **Where Value Accrues:** Unlike the SaaS boom, the "Layer Cake" theory suggests the semiconductor layer (chips & memory) will capture the most economic rent over the next decade. [00:04:23]
* **Edge vs. Center:** Wolfe argues that 50% of inference will eventually happen on-device (edge) rather than in data centers, which heavily benefits memory players. [00:07:22]
**Stock Picks & Company Views**
* **Nvidia:** Consensus long due to massive scale advantages, despite expected volatility. [00:16:53]
* **OpenAI vs. Anthropic:** The panel favors OpenAI for consumer dominance (zero switching costs, massive adoption). Anthropic is viewed as a likely acquisition target for Amazon, Google, or Meta. [00:10:15]
* **Meta:** Strong bullish sentiment. They are effectively building the "next hardware platform" (glasses/wrist) to bypass Apple/Google and owning the open-source model space. [00:12:07]
* **Top Picks:**
* **Gerstner:** Nvidia and OpenAI.
* **Wolfe:** SK Hynix (High Bandwidth Memory constraints) and Anduril (Defense/Autonomy). [00:14:28]
* **Shorts:** Wolfe calls publicly traded Quantum Computing and Modular Nuclear Reactor stocks "flapdoodle" and potential bubbles. [00:18:26]
**Future Disruptions**
* **The End of "Toll Booths":** The business model of the last 20 years (Google sending traffic to intermediaries like Expedia/Booking) is at risk. AI Agents will execute tasks directly, potentially bypassing the "tax collectors" of the web. [00:21:04]
* **"Fixware":** A rising category focused on maintenance tech (sensors, robotics) to extend the life of physical assets as the cost of capital remains higher. [00:24:39]
https://www.youtube.com/watch?v=ZmYtUxk2zgw