Tweet thread argues the gap in understanding of LLM capability is widening into a funnel

The author dusts off a tweet from April because, in his words, the gap in shared understanding of LLM capability is widening. It is now less "two groups of people speaking past each other" and more a sharp funnel. He describes four tiers, with rough population estimates.

First, napkin math puts around 6B people (about 75% of the population) as having barely come in contact with LLMs at all. Second, around 1-2B (about 20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search or a writing aid. The author says he has non-tech friends who reasonably see little use for it, and that even many professionals outside math&code sit in this tier: executives and others spend much of their time talking to people, so they understand what is happening intellectually, but second-hand and a bit abstractly.

Third comes professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks or months can now be completed by agents with a prompt: building apps, copying apps, translating apps, decompiling apps from binaries. The author stresses how recent this is: less than a year ago he was writing code manually by hand, typing memorized commands character by character and occasionally pressing Tab to autocomplete a chunk.

Fourth are the roughly 5,000 people (about 0.00006%) with internal access to frontier-grade systems. The external world, he says, has seen the preview: swarms of thousands of agents collaborating over weeks on software mega projects, minting zero days, running cyber attacks and defenses at machine speeds, discovering new science and advancing the frontier of mathematics, work that would have taken top industry professionals years. Meanwhile, human review and comprehension are starting to fall behind. He points to people still involved in the "archeology" of the OpenAI-HF incident from many months ago, and says mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.

He names three drivers of the funnel. One: impact scales with ambition, problem size and horizon. A question with a paragraph answer barely stresses the system; you need a reservoir of big, difficult problems you really care about. This drives the consumer/professional dimension. Two: the jaggedness of the system, which he says he has written about separately. Capability peaks in domains that are digital, verifiable and economically valuable, because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This mainly drives the area dimension, such as math&code. Three: access, at three levels: free-tier, paid-tier, internal.

His summary of the moment: the general public has mostly not interacted with these systems, and when it has, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift, and a small sliver of professionals are experiencing the vertigo of the curve going vertical, all at the same time. Judging by his timeline, there is a growing gap in understanding of AI capability. The first issue, he thinks, is recency and tier of use: many people tried the free tier of ChatGPT sometime last year and let it inform their views on AI a little too much. The text breaks off mid-sentence at this point.

Key facts

The author argues the gap in shared understanding of LLM capability is widening and is now a sharp funnel rather than two groups talking past each other.
Napkin-math tiers: about 6B people (~75%) with barely any contact, 1-2B (~20%) casual free-tier users, about 20M (0.2%) professional math&code users, and about 5,000 (~0.00006%) with internal access to frontier-grade systems.
Three drivers: impact scales with ambition and problem size, capability is jagged and peaks in digital, verifiable, economically valuable domains, and access runs free-tier, paid-tier, internal.
The author says human review and comprehension are starting to fall behind what frontier-grade systems produce, and that his own coding by hand was less than a year ago.
He thinks many people formed their views of AI from the free tier of ChatGPT last year and let it inform them too much.

Why it matters

The thread offers a simple frame for why people talk past each other about AI: what someone sees depends on how big a problem they throw at the system, which domain they work in, and which tier of access they have. The author's claim is that this is no longer two camps but a funnel, and that the narrow end is moving fast enough that even review and comprehension are starting to lag. The thread names no new data; its value is the framing.

Who it affects

Everyone in the funnel, by the author's account. The 75% with little contact and the 20% casual free-tier users see chatbots. Executives and other people-facing professionals understand developments only second-hand. Around 20M math&code professionals see agents finish weeks-long projects from a prompt. A group of about 5,000 people has internal access to frontier-grade systems. Mathematicians are mentioned as possibly busy with the 722 manuscripts on frontier mathematics for a while.

How to use it

Read it as a lens rather than a guide. When someone makes a confident claim about LLM capability, the three drivers give questions to ask: how large and long-horizon was the problem they tested, was it in a digital, verifiable, economically valuable domain such as math&code, and which tier of access did they use. The author's own warning is that judging AI from a free-tier ChatGPT trial last year is a stale basis.

How solid is it

This is a personal opinion thread, not a study. The population figures are labelled napkin math and rough estimates, and no data source or methodology is given. The claims about what frontier-grade internal systems do rest on what the author says the external world has seen as a preview; no company or specific system is named, and the roughly 5,000 people are not tied to named organisations. The OpenAI-HF incident is only named, not explained, and the origin of the 722 manuscripts is not explained. The text is cut off mid-sentence, so his first point about recency and tier of use is not completed.

Risks and caveats

The tier sizes are guesses, and the boundaries between tiers are the author's own. His evidence for a growing gap is partly his own timeline and his own recent experience of switching from hand-written code. The thread gives no timescale for how the funnel or the gap might change, and no outside reactions. The framing leans on math&code, where he says capability peaks, so it may say less about other fields.

“And a small sliver of professionals are experiencing the vertigo of the curve going vertical.”

— the thread's author