---
type: idea
title:
created: 2026-07-12
status: seed
origin_sources: []
related_auto: []
used_in: []
---

## Claim
- Models are getting better on benchmarks and raw intelligence
- Intelligence/Token is an increasingly important metric
- Lets call it raw intelligence
- Then there is learnt intelligence - solving a math problem by bringing solutions from adjacent families
- Cost goes up dramatically(?) Don't know but definitely higher -- do we know a way to measure this? Ask models to solve something on its baseline - then on post-training (customer still bears the cost as baseline increase) - and then there is learnt intelligence (customer bears the cost)
## Why it matters
- Enterprises will only pay on ROI eventually once we are past this experimentation phase
- S&P 493 EPS growth is 9% over the last two years? Most of it driven by pricing power, etc. 
- BG argued that a lot of revenue upside that is on the cusp - breakthroughs, discoveries, enabled by these agents have not even be commercialized
- I think this is an important measurement problem 

## Evidence / examples
![[Screenshot 2026-07-12 at 5.51.31 PM.png]]

![[Screenshot 2026-07-12 at 5.51.55 PM.png]]
![[Screenshot 2026-07-12 at 5.53.15 PM.png]]![[Screenshot 2026-07-12 at 5.54.02 PM.png]]![[Screenshot 2026-07-12 at 5.54.39 PM.png]]## Counterpoints

## Related

## Draft hooks
- Possible headline:
- Possible angle:
- What this is really about:
  

[https://x.com/yunta_tsai/status/2076323348217565485?s=20](https://x.com/yunta_tsai/status/2076323348217565485?s=20) - interesting discussion. for context, there seems to be a category of investors (those on captabls of Anthropic and OpenAI) who think real ROI of tokens is on the horizon, as discoveries, breakthroughs, etc. start becoming commercialized. This discussion is not conclusive - I am sharing it here for y’all to just get a sense of the way people are looking at models.I think this discussion feeds into or eats away from the investors’ narrative. If it is proven that that models can’t make jumps (refer the quoted tweet), then I’d suspect a steep valuation drop for the labs, since all of the sudden they would be operating in a constrained TAM (defined by enterprise budgets and cost efficiencies, and the ability of enterprises to absorb the technology to drive productive gains - today only the tech-native players like Uber, Doordash, Databricks, Amazon have demonstrated the ability to deliver on that). If novel discoveries become achievable because of how good the next Anthropic or OpenAI model is, their TAM is basically unconstrained. Regardless, a big chunk of the medium-term value is dependent on the first layer - that is the visible slice up for grabs.