Post Argues Shared Understanding of LLM Capability Is Splintering Into a 'Sharp Funnel'
Summary
- • Post argues the gap in understanding of LLM capability is widening into a 'sharp funnel'
- • Napkin math: ~75% of people barely touched LLMs; only ~5,000 have internal frontier access
- • Author claims insiders see agent swarms finding zero-days and advancing mathematics, with human review lagging
- • Author names three drivers: problem ambition, jagged capability from verifiable-reward RL, and access tiers
Details
Author argues the gap in shared understanding of LLM capability is widening into a 'sharp funnel'
People at different exposure levels see very different things and increasingly cannot agree on what the technology can do. This is the author's argument, not measured fact.
Author estimates ~6B people (~75%) have barely touched LLMs
This is the widest tier of the funnel. The figure is napkin math, not survey data.
Author estimates 1-2B people (~20%) are casual, mostly free-tier users
The author implies this group's impressions are shaped by weaker, freely available models rather than the frontier.
Author estimates ~20M people (~0.2%) are professionals using frontier LLMs in math and code
According to the author, this group sees agents complete projects that would take humans weeks or months from a single prompt.
Author estimates only ~5,000 people have internal access to frontier systems
The author claims the external preview looks like swarms of thousands of agents collaborating on software megaprojects, finding zero-days and advancing mathematics. These claims are unverified in the post.
Author says human review is falling behind AI output, citing an 'archeology' of the OpenAI-HF incident and 722 manuscripts on frontier mathematics
These are offered as examples of output volume outpacing people's ability to check it; the post gives little further detail.
Drivers: problem ambition, jagged capability from RL on verifiable rewards, and access tiers
Capability peaks in digital, verifiable, economically valuable domains like math and code, so users elsewhere see a weaker picture.
The post was originally a tweet from April 9 and resurfaced on October 11, 2026
The picture reflects April and is being re-read against later developments. The figures are rough estimates.
Stat = author's estimated figures, Insight = author's argument, Context = background on the post
What This Means
The post argues public debate about AI is hard because participants are effectively looking at different technologies: casual users, professional coders and a few thousand insiders each see something distinct. The claims about internal systems and the numbers are unverified estimates from one author and should be read as opinion.