Karpathy's "Sharp Funnel": Why People See Such Different AIs

For some, AI is a chatbot that answers off target. For others, it is a capability climbing almost vertically. Andrej Karpathy says the gap is widening.

The 20-second read

1
Same AI, completely different experiences. Andrej Karpathy estimated that roughly 1 billion to 2 billion people use AI as a light search or writing tool, while about 20 million use it to handle complex professional work.
2
The gap is not only about technical skill. How hard the problems are that people hand to AI, which field they use it in, and which models they can access all change how capable AI feels.
3
The new bottleneck is human ability to verify. Karpathy observes that AI has begun producing output faster than humans can understand and review it.

For some people, AI is a chatbot that answers questions off target. It is a slightly more convenient search engine.

Other developers hand AI agents jobs that used to take days. At frontier labs, researchers are testing systems in which many agents collaborate over long periods.

So who is seeing AI correctly?

Andrej Karpathy, Tesla's former AI chief and an OpenAI co-founder, described the gap in an Oct. 11, 2026 post on X. He called it a "Sharp Funnel." In April, he said, two very different groups could barely talk to each other. Now, he argues, the gap is wider.

1. A funnel that sorts humanity into four tiers

Karpathy's estimate of four tiers of AI experience

The population figures are Karpathy's rough estimates, not official usage statistics. Bar lengths in the original graphic were conceptual, not proportional to population.

  1. About 6 billion people: those who have barely experienced AI.
  2. About 1 billion to 2 billion: occasional users of free chatbots.
  3. About 20 million: professionals using frontier models for their work.
  4. About 5,000: people with access to frontier systems inside research labs.

The first group, about 6 billion people, has barely touched a large language model (LLM). The second uses free chatbots now and then, for search, translation and writing help.

The picture changes at the third tier. About 20 million mathematicians, developers and others use high-performance models to build applications or analyze complex code. Karpathy cited cases in which agents finish projects that once took weeks or months.

The fourth group is about 5,000 people. According to Karpathy, they work with frontier systems inside labs and are seeing the potential of large-scale agent collaboration in cybersecurity, mathematics and scientific research.

The precision of the headcounts is not the point. Karpathy himself labeled them "napkin math." What matters is that judgments about the same technology change completely with how deeply a person has experienced AI.

2. Why the gap does not close

Karpathy points to three causes.

First, the size of the problem handed to AI

Short questions and answers rarely show what a strong model can do. An agent's value shows more clearly in projects that need multi-step planning, execution and revision.

Second, AI ability is uneven across fields

Current models are especially strong where answers can be checked fairly clearly, such as coding and math. In reinforcement learning, it is easier to use measurable rewards, such as whether code passes its tests.

In work where the right answer is hard to judge, such as personal taste, complex negotiation or running an organization, the benefit of better performance may be less clear.

Third, differences in access

Free-model users, paying users of high-end models and users of internal research systems do not experience the same capabilities. Some frontier features have not been released to the general public.

That makes the AI debate not only about what the technology can do, but also about differences in judgment among people who have used different products in different settings.

3. AI's output speed is outrunning human verification

The most interesting part of Karpathy's post sits at the top of the funnel.

He mentioned a case in which mathematicians are analyzing 722 AI-generated math research manuscripts. OpenAI's public research repository also holds hundreds of manuscripts, and some have not yet completed formal verification.

This does not mean AI has proved hundreds of new theorems. The validity of the results has to be confirmed separately.

Still, an important question remains. When AI-generated code and research results multiply, how quickly can humans understand and verify them?

Going forward, the variable that decides productivity may not be raw generation speed. How reliably one can finish output that can be verified may become the stronger edge.

Underrating AI is a problem. So is accepting abilities that have not been verified as achievements already delivered.

The strange moment Karpathy describes is that both realities exist at once. The key question is no longer how impressive AI is. What problems are we giving AI, and are we ready to verify the results?

If I only use free AI, can I not see how far the technology has really advanced?
  • Not necessarily. Free models are good enough for many everyday tasks. But on complex, long-running projects, differences can show up depending on the high-performance model and the environment you use it in.
Will people who can't code be left out of the AI revolution?
  • More important than coding itself is the ability to define a problem clearly and verify the results. Non-developers can also see productivity gains through data analysis and work automation.
Is the 5,000 people Karpathy mentioned a confirmed number?
  • No. It is not an official statistic from a survey of headcount at each lab. It is his rough estimate.
If AI outpaces human verification speed, does it become hard to trust the results?
  • The risk can grow. That is why automated testing, reproducible experiments and independent review systems are becoming more important.

What to watch

Real-world work automation rateThe share of work where AI goes beyond drafting and completes the final output.
Verification cost and error rateEven if AI gets faster, net productivity can fall if the time people spend correcting it grows.
Spread of high-performance modelsHow quickly the performance gap narrows between free, paid and internal systems.
Repeat use and revenue from business customersA measure of whether one-off experiments turn into a sustainable business.
How you see AI depends largely on which AI you have used and on how big the tasks were, and checking what AI produces is the next job.

Insight Times Editorial Desk

More by topic

Starting with this story's topics, then what readers are following now

Investor psychology, decision-making and books on investing

TSLA

This story's topic50 storiesSee all →

The Cybercab is not a $30,000 car. Tesla is selling hours, not vehicles

No steering wheel, no pedals. The economics of Tesla's robotaxi turn on how many paid hours a day it runs, not on what it costs to build.

Tech

Why Tesla Jumped: Barclays Sets a $450 Target on SpaceX Merger Hopes

Hopes for a SpaceX merger moved the stock. Whether the rally lasts likely depends on how robotaxis turn into profit and cash flow.

Investing

Musk Says He'd Cancel Government Contracts. Washington Still Holds Leverage

After Democrats won both chambers in the midterms, Musk floated cancelling SpaceX's federal contracts. The share of revenue matters less than bargaining power and the regulatory risk that remains.

Tech

Tesla Semi's Real Bottleneck Is Starting to Clear

A new 30-megawatt heavy-truck charging buildout in California is too small to move Tesla's earnings on its own. But it marks real progress on one of the Semi business's hardest problems: charging infrastructure that has to exist before the trucks do.

Investing

Musk Sees the Next Phase of AI Infrastructure: Who Can Switch GPUs On, Fast

If neoclouds struggle under financial pressure, AI demand may not shrink. It may simply favor the few operators with the capital and execution to serve it, a shift that could matter for SpaceX and Tesla.

Tech

Musk's Next Bet Is Not an AI Model. It Is the Industrial System That Produces AI

At the Barron's conference, Elon Musk laid out an industrial strategy that spans chips, space data centers, rockets, satellites and robots. Investors need to ask when each stage starts earning cash.

Investing

If You Only Watch Tesla, You May Be Missing Half the AI Story

Nvidia builds the AI brain. Who supplies the electricity to run it? Six power companies Tesla investors should know, and why Tesla's storage business is just one link in the chain.

Investing

AI Agents

This story's topic47 storiesSee all →

AI that takes one instruction and works through the steps on its own — calling…

Box Shows AI Its Manuals, and the AI Starts Redesigning the Workflow

Box Architect, built with NVIDIA, moves AI from reading documents to designing how work gets done. It has not yet reached beta, so the real-world payoff remains to be seen.

Tech

Nvidia and Palantir Put AI Inside the Supply Chain - The Real Shift Is That It's Learning Judgment

A single Vera Rubin rack holds 1.3 million parts. The real breakthrough in Nvidia and Palantir's new supply chain AI isn't automating that complexity, it's capturing the judgment calls that used to live only in planners' heads.

Tech

A person at a kitchen table looks at a food delivery app on a phone beside a laptop and a DoorDash delivery bag

AI Was Supposed to Do Our Shopping. DoorDash Got Only Several Hundred Orders From It

DoorDash logged 993 million orders in the third quarter, yet outside AI agents placed only several hundred in total. Orders built through its own in-app AI ran about 50% larger.

Tech

Nvidia Moves From GPUs to CPUs as AI Agents Reshape a $200 Billion Chip Market

A customer's early test found Nvidia's Vera CPU handled about twice the AI agent workload of its existing chips. The bigger story is where the money in semiconductors is heading as AI shifts from answering questions to doing work.

Tech

The Cheaper Tokens Get, the Bigger Your AI Bill Gets

As AI shifts from training to inference, the token is becoming an industrial commodity. Falling prices are not cutting demand, because agents and automation consume tokens at a scale no human chat session ever did.

Investing

Nvidia's Next Target After GPUs: Controlling What AI Agents Are Allowed to Do

OpenShell and Sentry are not about policing what an AI says. They are about controlling, at the server infrastructure level, which files and APIs an AI can touch, what it can execute, and who can actually pull the plug when something goes wrong.

Tech

AI Needs to Find $1 Trillion a Year From Somewhere to Pay Back Its Capex

Saying AI capex is big is no longer the point. The harder question is what end users have to pay for the buildout to pencil out, and Goldman Sachs puts that number at roughly $1 trillion a year.

Tech

Cybersecurity

This story's topic16 storiesSee all →

Defending systems and data from attack.

The AI Race Just Got Its First Brake: What Amodei, Altman and Musk's Slowdown Signals

For the first time, AI safety concerns are shaping both model release schedules and IPO timing at once. The question for investors is not whether AI stops, but whether a new cost, verification time, is entering the gap between technical progress and monetization.

Tech

AI Speeds Up Hacking and Defense Alike. Which Cybersecurity Stocks Profit?

AI is getting faster at finding software weaknesses. The question is less who finds them first than who sorts out the real risks and gets them fixed.

Tech

When AI Learns to Pick Locks: Security Now Covers Permissions, Supply Chains and Recovery

AI is reshaping attack and defense at once. Security now means controlling what AI can access, vetting the software supply chain and being able to recover after a breach.

Tech

AI Won't Slow Down. The Real Fight Is Over What Agents Can Do

Dario Amodei wants frontier AI labs to slow capability gains, but a US-wide deceleration deal looks unlikely. The more probable path: keep accelerating commercial AI while tightening security around agent permissions, high-risk features, model weights and large-scale compute access.

Tech

Grok 4.7's Real Weapon Isn't Performance, It's Cost Per Finished Task

At $2 per million input tokens and $6 per million output tokens, SpaceXAI is pitching Grok 4.7 as an agent model for long coding and knowledge work. The number investors should watch is not the benchmark leaderboard but the total cost of finishing a job and how widely it gets adopted.

Tech

Claude Now "Leads" 26% of Anthropic's AI R&D. The Real Shift Is the Research Factory

Anthropic says Claude leads 26% of its AI R&D tasks and touches more than 90% of them in some way, with roughly 30,000 agents running at once. The bigger story is how research itself is being restructured.

Tech

If an AI Says "I Don't Want to Be Shut Off"

Whether AI can truly have feelings is an interesting question. The more urgent one is how much authority humans are already handing over to it.

Tech

A building that houses servers and networking gear and supplies them with powe…

Why Nebius Is Keeping Some 2027 AI Capacity Off the Shelf

Nebius mixes long-term contracts, short-term premiums and auctions. Holding back capacity pays only if the servers stay busy, and it is not yet clear the pricing will last.

Tech

AI ran out of memory before it ran out of GPUs

The HBM bottleneck is spreading into DRAM prices, server costs and the price of the next laptop or phone.

Tech

Is Space Cooler for AI? Why Musk Wants Data Centers in Orbit

The next limit on AI may be power and cooling, not GPUs. Elon Musk is looking to Earth orbit for an answer, but the test is cost per hour of compute.

Tech

White House Swaps AI for "SI." The Real Investment Signal Is Infrastructure, Not the Name

An executive order renames artificial intelligence "Super Intelligence" and six frontier labs sign a voluntary safety accord. What investors should watch is the data center, power, nuclear and chip policy now gathered under one national-competitiveness frame.

Tech

Weekly Check: The AI Bull Market Isn't Over, It's Just Gotten Harder

A 5% 10-year yield, 470 gigawatts stacked up in Texas, Chinese DRAM, and whether data centers actually get switched on are starting to decide who wins the AI trade through 2027.

Tech

Is Memory Peaking? In the AI Era, Watch the Bottleneck, Not the Demand

Demand still outruns supply. But memory stocks can peak before demand does. The real question is how long HBM4, server DRAM, CXL and Big Tech capex keep the cash flowing.

Tech

Marvell's $115 Billion Lumentum Deal: Is AI's Next Bottleneck Light, Not Chips?

Marvell's purchase of Lumentum points to a contest beyond GPU speed, in data transfer and power efficiency. Whether the price is right is not yet confirmed.

Tech