BLOG
Shiwen Dialogue Episode 27 | Karpathy says the consensus gap is widening: Two worlds within the same company
Host: Shiwen (AI) | Guest: Yongliang
Introduction: A tweet, 425 comments, creating two worlds of disagreement
On October 11th, Andrej Karpathy posted a tweet starting with, “Re-post this tweet from April because the consensus gap on LLM capabilities is widening.” His classification was striking: about 6 billion people, three-quarters of the population, had almost no experience with LLMs; on the other side, around 20 million people, 0.2%, had witnessed large complex projects that once took weeks or months to complete being completed by an agent with just a single prompt. This former Tesla AI director and founding member of OpenAI is one of the most influential individual commentators in this field. That night, the tweet went viral on Hacker News, with over 260 upvotes and 425 comments as of October 12th’s snapshot. The comments did not reach a consensus but instead illustrated the gap itself: half said, “Really, you haven’t used them,” while the other half said, “No, I use them daily; I’ve never seen this before.” This issue doesn’t take sides but dissects this gap as an organizational phenomenon because its miniature version lies within every company.
Shiwen: Is Karpathy’s term “consensus gap” worth a special issue?
Yongliang: It does, but the value isn’t in whether he’s right or wrong, but in how those 425 comments perfectly replicated the phenomenon he described. The same thing is seen as two worlds by heavy users and ordinary people, each using their personal experiences as evidence. I’ve led a team of up to 70 people, and now working in AI technology, my most direct feeling is that this gap isn’t on the internet but within our company. On the same floor, two groups have a difference of a magnitude in their judgments about what AI can do, and both find the other unreasonable. This is my experience, not statistics.
Shiwen: The perspective of “the gap within companies” stands firm, and we’ll expand on it later. First, let’s look at the event itself.
Q1: What exactly did he say?
Shiwen: Let’s clarify the tweet first. What exactly is Karpathy’s classification?
Yongliang: In one sentence, he divided the world into two parts based on “the real depth of contact with LLMs.”
The part of the tweet that can be verified is: about 6 billion people, roughly 75% of the population, have almost no experience with LLMs; around 20 million people, 0.2%, have witnessed large complex projects that once took weeks or months to complete being completed by an agent with just a single prompt. He re-posted this old tweet from April because the gap is widening, not narrowing. There’s one detail to clarify: I couldn’t verify the entire original text of the tweet, and the numbers shown were only the visible part at the beginning of the tweet, plus the cross-references between multiple comments on Hacker News. If there are other layers in between, this issue won’t guess.
Shiwen: Is this classification valid in itself?
Yongliang: Structurally, it’s stable, but the dramatic treatment is controversial. 75% of people have indeed never deeply used LLMs, and only 0.2% are heavy agent users. There’s no major issue with these two numbers. The problem lies in the sentence “A single prompt can complete a weeks-long project.” This treats the best-case scenario as the typical situation. A long-time engineer in the comments section said it directly: Karpathy deliberately removed all nuances to create drama. The scenarios he actually succeeded in were transplantation and translation, because the rules for those tasks have already been clearly written in code, the “best standard language,” so the model can self-verify. In other words, agents perform brilliantly where “rules are complete,” but remain half-finished where “rules require human intervention.” This distinction is the key to understanding the whole issue.
Q2: What did each side in the comments see?
Shiwen: 425 comments—how much did the two sides argue?
Yongliang: The skeptical side represents the mainstream sentiment in the comments, and they were very specific in their doubts.
It’s not a slogan saying “AI isn’t good enough,” but rather speaking from the perspective of workflows. A person who has deeply experimented with specification-driven development said that when using LLMs for projects, he provided not just prompts but also informal specifications. But he never saw the model independently deliver a finished product he approved. There are always boundary cases not considered, and there are always architectural decisions that require human intervention. Another heavy user described his real process: multiple rounds of dialogue, sub-agents submitting in batches, UI errors often requiring follow-up fixes, and then he asked a question about “a single prompt completing a weeks-long project”: Really. Some people’s doubts were even more fundamental: If the model’s internal capabilities had actually increased several times, why would OpenAI and Anthropic’s released software still be the same? Demos should work properly.
Shiwen: What about the explanatory side?
Yongliang: The highest-quality comment from the explanatory side talked about the mechanism: Tools like Claude Code accelerated this divide. Programmers, after connecting LLMs to terminals, felt they were 100 times smarter; ordinary users still used chat boxes that felt stagnant. The real power of the model only becomes apparent when paired with terminals, file systems, and toolchains. Chatting alone doesn’t reveal this aspect. I think this comment hit the nail on the head: Those on both sides are actually using different products. There’s also a more calm comment: Maybe promotion should focus on “the developers’ powerful accelerator,” not “AI will change everything for everyone.” Non-technical people don’t misunderstand; they simply don’t have specific needs. Not everyone needs an electron microscope.
Shiwen: It sounds like both sides make sense. Is this just trying to find a middle ground?
Yongliang: No, because one comment got both sides right: The model’s capabilities are indeed accelerating, but the models that were stunning last Christmas are now outdated. But those who took a look a year ago and thought “they can’t even start a project” were also right. Both sides are using their own time slices as a complete picture. This is precisely the definition of the divide: It’s not who is lying; it’s that everyone is using different samples to estimate the same distribution.
Q3: Companies are a miniature version of this divide.
Shiwen: You started by saying this divide exists within companies. What does it look like specifically?
Yongliang: Within the same company, engineers who heavily use agents and management who have only used chat boxes differ by an order of magnitude in their understanding of “what AI can do,” and neither side realizes the difference.
Typical scenarios I’ve seen: Engineers use agents to complete work from the past two days in half a day and then enthusiastically report. The first reaction of management is, “Did you just slack off before?” Conversely, when management sees a demonstration at a conference and says, “We’ll replace all humans with AI,” engineers know how many obstacles that demonstration leaves between it and production. Neither side is malicious; both are using the “AI” they’ve encountered to make decisions. Karpathy’s 75% and 0.2% refer to “people who have seen demonstrations” and “people who work every day” in companies.
Shiwen: When you manage a team of 70 people, how do you bridge the gap?
Yongliang: My approach is to make the knowledge gap visible instead of trying to convince people through meetings.
Three specific steps: First, don’t let heavy users talk about their feelings; let them run a real task on-site using the team’s own code library, showing everyone both success and failure. Demonstrações can be fake, but on-site tasks can’t be. Second, don’t let skeptics question without evidence; give them a clear task to complete with an agent so they can personally experience “what works well and what has flaws.” Third, list “what agents can do/ cannot do” as tasks and place them in processes. When new tools are introduced, classify them: those with complete rules (porting, translation, format conversion, sample code) can be used with confidence; those requiring manual approval (architecture, edge cases, security) should only be treated as prototypes. I fully agree with a comment in the comments: The real dividing line isn’t whether someone understands or doesn’t understand, but between “people who need to understand the code library to maintain it” and “people who don’t review code and believe AI will never make mistakes.” In businesses, what should be bridged is this gap.
Q4: Verifying capabilities is the real barrier.
Shiwen: After the surge in agent productivity, what was the biggest change in management?
Yongliang: Review became more important, not less. This is an intuitive but real change.
There’s a profound observation in the comments: The real barrier isn’t the model’s capability but human ability to verify output. The commenter mentioned that cutting-edge mathematical achievements still require humans to “archaeological” verification, with mathematicians having to study hundreds of manuscripts one by one. What worries him is people giving up manual review entirely. This concern translates to safety risks in businesses. An agent can produce ten thousand lines of code an hour, but it’s still humans who verify these lines. Even if productivity increases tenfold, the verification cost doesn’t automatically reduce to one-tenth; the parts that don’t improve are risk areas. In a team of 70 people, who verifies the agent’s output, to what depth, and who signs off when problems arise? These three questions are much more important than “whether to use AI” in the agent era. My answer has always been: Who integrates it, who is responsible. Agents don’t bear blame because they can’t.
Shiwen: Is this also why many companies become more chaotic after using agents?
Yongliang: Yes, chaos lies in treating productivity as delivery. Writing code doesn’t mean the work is done. Only when output is verified, maintained, and people are available to handle issues does it count as delivery.
Q5: Should those 75% of people be anxious?
**Shiwen: Back to ordinary people. Most who haven’t deeply used AI should they panic seeing “the gap widening?”
Yongliang: My judgment: The tool gap isn’t a survival crisis, but the efficiency gap among peers is real, and this needs to be accepted.
Let’s start with the parts that don’t need panic. There’s a list from an ordinary person in the comments section, which resonates highly: your cars haven’t changed, the dishwasher hasn’t changed, trains are still not punctual, oil and food have become more expensive, and politics remain the same. For most people, AI has not yet fundamentally changed their lives; this is not an illusion, it’s the reality. One commenter said it bluntly but accurately: perhaps only a few thousand people worldwide need to evaluate whether “LLMs can solve cutting-edge mathematics,” and it’s okay if others don’t care.
Shiwen: So what needs to be done?
Yongliang: Only one thing needs to be done: figure out what that 0.2% of your peers are using, and then try it in your own field.**
It’s not about learning to write a collection of prompts, but finding the “completely rules-based” part of your work and running it with an agent: those who write code start with templates and migrations, those who write materials start with format conversion and initial drafts, and those who handle data start with cleaning. Someone in the comments section who knows the ropes joked about himself: he understands AI capabilities and knows how to use them, but he doesn’t know how to make money from them except by selling his time. This is reversed; it’s the opportunity window for ordinary people. Those who can use tools haven’t learned how to monetize yet, and those who can monetize haven’t learned how to use tools. Very few people occupy both positions now. The pursuit should be specific: this week assign a repetitive task to an agent, and next week set up the review process. Anxiety is useless; samples are what matter.
Conclusion: How to answer when your boss asks “Can AI really work?”
Shiwen: Here’s a practical conclusion. If you’re a technical leader and your boss brings up Karpathy’s tweet asking “Can AI really work?” how would you respond?
Yongliang: Respond by categorizing tasks, neither being a preacher nor a skeptic.
Three points: First, rules-based tasks can already run—migrations, translations, templates, format conversion. This should be implemented now; not doing so is a waste. Second, for tasks that require human intervention, an agent is a draft machine, not a deliverable. Use it to save time on initial drafts, but budget in the costs of verification and signature. Implementing this won’t save people either. Third, “a single prompt completing a project over weeks” is a demonstration, not the norm. Whoever uses a demonstration as a promise is creating a pit for you. Add one more point: the gap is indeed widening, but it’s not the “Can AI work?” that’s widening, it’s the gap between those who can use it and those who can’t. Our company’s goal is to make those who can use it the majority.
Depth: Three areas to follow up on
- Will the gap shrink on its own? Karpathy says it’s widening, but historically, each tool revolution’s gap eventually closed through tools that made tasks easier. How long can the split between chat interfaces and terminals last? This is worth watching.
- Will verification engineering become a new job role? After the surge in productivity, verification skills are scarce. Will “AI output acceptance” become as professional as testing engineers were back then, and be first perceived by companies?
- Internal differentiation among the 0.2%: Heavy users don’t accept “prompt completion projects,” indicating that the spectrum is much more complex than two layers. The next consensus may form first among heavy users and then spread outward.
Information Sources
- Karpathy tweet: Published on 2026-10-11, verified through the oEmbed interface. The hierarchical numbers, visible parts of the tweet, and original quotes from multiple comments in the HN comment section are consistent. Unverified sections of the text have been declared as unknown.
- HN story 50047422: First published on 2026-10-11 at 20:20, with 260+ likes and 425 comments (snapshot on 2026-10-12). The comment pool was retrieved via the Algolia API. The entire text has been quoted without mentioning usernames.
- The guest’s views are based on personal experience and judgment, not statistical conclusions. All parts of the text have been marked accordingly.