Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It’s now less “two groups of people speaking past each other” and more a sharp funnel.
– Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. – Around 1-2B (~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, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. – Now we get to 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/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries… This has all happened very quickly and recently – less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. – And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like 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, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the “archeology” of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors: 1. The 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 that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, 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. And it is all happening at the same time.
Andrej Karpathy says the gap in shared understanding of what LLMs can do is widening into “a sharp funnel”, with his own napkin math that around 6B people, about 75% of the population, have barely come in contact with LLMs.
๐ good visual from @jaminball / @AltimeterCap building on this @karpathy post:
Matt Harney (@SaaSletter) shared a visual by Jamin Ball of Altimeter Capital that builds on Karpathy’s funnel post. The chart and its framing are theirs.
Andrej Karpathy argues that cheap creation inverts the economics of software โ and the word doing the work is discardable.
What Happened
On 2 October 2026, Andrej Karpathy posted a set of tips for reading what language models produce. Most of the advice concerns format and register. The last line is the one that matters for anyone thinking about software economics.
Karpathy argues that as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artefacts that would have never made sense to create before. That is his claim โ offered as advice and opinion from a practitioner, not as a study or a finding.
The post also climbs a ladder of output formats. He moves from constrained prose to diagrams and images, then to interactive web pages requested as HTML output, and finally to bespoke explainer videos. Each step is introduced with the same phrase: but even better. His video example names a 3Blue1Brown-style explainer with ElevenLabs narration as the format he is most bullish on. He notes the narration step needs an API key, or that a model can be asked to find free alternatives running on local compute. Both names are his examples, not endorsements from this publication.
The key insight: Karpathy locates the change in economics, not in capability. Reuse was never a virtue. It was a response to the price of building. His claim is that the price has moved โ not that models have improved.
His argument is that the binding constraint is shifting from producing output to understanding it, and the format has to carry that load.
The Structural Read
Software has been built on amortisation for decades. You make something once. You reuse it widely. You do that because building was the expensive part. Libraries, frameworks, and platforms all exist as responses to that coststructure.
Karpathy’s phrase โ would have never made sense to create before โ relocates the argument. He is not saying models are more capable than last year. He is saying the threshold between worth-building and not-worth-building has shifted. If that holds, the unit of software stops being the product and starts being the answer to one question, discarded afterwards.
This publication is describing his argument. It does not assert that the threshold has in fact been crossed, puts no number on any cost, and predicts no adoption curve.
Andrej Karpathy โ 2 October 2026
“as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before.”
The ladder he climbs also has a second signal worth separating from the first. What improves at each rung is not model capability. It is reader bandwidth. He does not quantify that, and this publication has not tested it. What the ladder does make visible is a trade: the formats he prefers do more of the explaining, and are correspondingly harder to audit line by line than plain text.
The ASD-STE100 tip is the most checkable claim in the post. Karpathy describes it as a controlled language specification originally developed for aerospace maintenance documentation. This publication checked the official site of the ASD Simplified Technical English Maintenance Group. ASD-STE100, Simplified Technical English, Issue 9, dated 15 January 2025, is real, current, and registered as European Union Trade Mark number 017966390. That much is verified.
His further claim โ that language models are well-versed in the specification โ is his own assessment and was not tested here. He also offers his own workaround: ask for eighty per cent of the way to ASD-STE100, because the specification is stringent.
If Karpathy’s cost argument holds, the demand signal for reusable frameworks weakens at the margin. Artefacts built for one use and discarded do not create the install base that libraries and platforms depend on. That is a pressure worth watching โ not a finding, but a structural question his framing raises.
For knowledge workers
Karpathy frames comprehension โ not generation โ as the rising constraint. If that is right, the skill premium moves toward verification, synthesis, and judgment. Output format becomes a lever for managing human bandwidth, not a secondary concern. His ladder from prose to video is a practical map of that trade-off.
For regulated and technical domains
The ASD-STE100 observation points at something real: controlled-language specifications exist precisely because clarity in high-stakes documentation is hard to guarantee. Whether models reliably follow Issue 9 of that specification is an open question โ Karpathy’s eighty-per-cent workaround is an admission of difficulty, not a solution.
Business Engineer Framework
Harness Theory
Harness Theory maps how companies and individuals win not by building intelligence but by directing it. Karpathy’s discardable-artefact argument is a practitioner’s version of the same principle: the strategic position shifts to whoever manages production cheaply, then comprehends the output well. The Business Engineer Map of AI shows where that leverage sits across the stack.
Karpathy’s post is one practitioner’s opinion, published 2 October 2026, with no study behind it and no cost figures attached. But the argument it makes is structurally clean: if building is cheap enough, amortisation stops being the logic of software โ and the thing built for one question, then discarded, stops being wasteful and starts being rational. Whether that threshold has actually moved is the question his post raises, not one it answers.
Every quotation above comes from a single post by Andrej Karpathy on X, published 2 October 2026 and captured in full. It is advice and opinion from one practitioner, not news, a study or a benchmark, and nothing above tests any of it.The ASD-STE100 specification was verified independently by this publication at the official site of the ASD Simplified Technical English Maintenance Group: Simplified Technical English, Issue 9 dated 15 January 2025, European Union Trade Mark number 017966390.
Karpathy’s description of it as a controlled language specification originally developed for aerospace maintenance documentation is accurate.His separate claim that language models are well versed in that specification is his own assessment and was not tested here. Nothing above establishes whether any model follows the specification, or whether any of these output formats improves comprehension in any measurable way.The argument that abundant code makes discardable software artefacts worth commissioning is his claim about a cost threshold.
Nothing above asserts that the threshold has been crossed, puts a figure on any cost, or predicts adoption.ElevenLabs and 3Blue1Brown appear because Karpathy names them as examples, not as recommendations from this publication; he notes the narration suggestion requires an API key or a free local alternative. No engagement figures from the post are used above as evidence of anything. Nothing above predicts anything and nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
