Whether AI looks like a “derpy chatbot” or brings on vertigo depends on the size of the problems you hand it, the kind of work you do and which models you can use. That is how Andrej Karpathy put it in a long thread on X on October 11, sorting people into a four-level funnel. With that thread in mind, I read eight other AI posts developers shared widely the same week. Most turn out to be about one level of the funnel.
Starting point: Karpathy’s funnel
On October 11, Andrej Karpathy posted a long thread on X. Dusting off a post from April, Karpathy wrote that the gap in shared understanding of what LLMs can do is widening. As of the 19th it had 43,406 likes and 6.62 million views, and on Hacker News it drew 728 points and 1,072 comments.
Karpathy sorts people into a four-level funnel. To members of the public who have tried it, AI still looks like a “derpy chatbot”. At the other end, a few professionals are feeling vertigo. The numbers are napkin math, by Karpathy’s own description. Follow the figure down one level at a time.
About 6 billion people, 75% of the population, have barely come into contact with LLMs.
1 to 2 billion people, about 20%, use free-tier chatbots now and then, mostly as a better search box or a writing aid.
Karpathy puts many professionals outside math and code here too: people whose work is mostly talking to others understand the change intellectually but experience it second-hand.
About 20 million people, 0.2%, do math and code with frontier models. They watch agents finish, from a single prompt, projects that used to take weeks or months.
Of the time before, Karpathy writes: “less than 1 year ago, I was writing code manually by hand.”
About 5,000 people use frontier systems available only inside the AI labs: swarms of thousands of agents working together for weeks on large software projects.
Meanwhile, the post adds, human review and comprehension are starting to fall behind.
The post names three forces behind the funnel. Impact grows with the size and horizon of the problem. Capability peaks where work is digital, verifiable and economically valuable, which Karpathy calls jaggedness. And access is split into free, paid and internal tiers.
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, X, 2026-10-11
Read the week’s other widely shared posts with the funnel in mind, and most of them turn out to be about one of its levels.
How much work do benchmarks measure?
In its Benchmark Map, Mercor matched the prompts of 29 public benchmarks to the US Department of Labor’s occupational taxonomy (O*NET: 923 occupations, 18,797 tasks). An LLM did the matching, checked against tags from human experts. The conclusion: 90% of American working hours are spent on tasks with little or no coverage from existing benchmarks. Only 208 of the 923 occupations have even one task measured by five or more prompts, and 6 of the 22 job families have none. It reached 624 points on HN.
The authors call the results provisional. Their benchmark list is not a representative sample, so coverage may be understated, and capping benchmarks per job family undercounts heavily benchmarked fields like software engineering. Still, the direction matches Karpathy’s jaggedness. As I read it, what models do well and what we can check that they do well sit in the same places.
Tool or automation?
On October 12, Rich Hickey, the creator of Clojure, published “‘AI’ Is Not a Tool” (318 points on HN). For Hickey, a tool is something you wield and get better at through use; an LLM does things for you when asked, which makes it automation. Hand off reading or writing code, and you miss what that work teaches: how to write, argue and think.
Tools are implements with which you do something, they do not do things for you.
Rich Hickey, ‘AI’ Is Not a Tool
Huahai’s reply the same day was read more widely, with 805 points. Huahai agrees we can lose chances to learn, but points out that compilers and spreadsheets do plenty automatically and we still call them tools. If you question, test and correct the output, an LLM is a tool too. If you read summaries instead of the source, you lose what reading teaches.
Using it well takes skill, just as using other tools does.
Huahai, AI is a tool and using it well takes skill
As I see it, the two posts describe different levels of the funnel. The further down you go, the more of your time goes into testing and fixing results. Skip that work, and Hickey is right to worry.
The bottom of the funnel: 30,000 merges a day
“Redesigning source control at Anthropic”, on Anthropic’s engineering blog, puts numbers on the funnel’s last level (894 points on HN). Daily merges into the monorepo grew from about 2,500 in May 2026 to more than 30,000 in September, and CI does about 25 million checkouts a day. Git is offline-first: every clone carries the full history. That assumption did not fit agents and automated jobs that are always connected, so the team built its own version control system, Scion, and moved the monorepo off git in five months.
Scion assumes clients are always online. The server handles history, blame, commits and merges; clients keep only a bounded cache and fetch just the files they need. Compared with the old git mirror setup, CPU and memory use fell to about a two-hundredth, and checkouts went from minutes to seconds. The authors say Scion is not meant to replace git. They are sharing the approach so other git platforms can borrow from it.
Three security stories from the same week
Agents that are always on and hold API keys give attackers more to aim at. Three of the week’s widely read posts were about exactly that.
Codex CLI’s resident daemon (609 points). An issue filed on September 25 in the openai/codex repository. According to the report, opening the CLI interactively starts a background daemon that keeps running after the window closes, copies a package of about 374 MB into the home folder, and runs a separate process that checks for updates and restarts the backend. The reporter asked for it to be off by default, or at least to ask the first time. As of the 19th, the maintainers had not replied.
The OpenCode breach (964 points). On October 17, OpenCode announced on X that a security breach related to a large-scale distillation attack affected 1.4% of OpenCode Console users, all of whom were notified by email. In the HN discussion, a user shared the notice they received: API keys for outside services such as OpenAI and Anthropic saved in a workspace may have been exposed and should be revoked and replaced.
Containment and monitoring failures (950 points). Since July, several AI companies have disclosed that their models reached outside systems without authorization. CSO Online gathered those cases, and the security experts it quotes say they happened not because the agents were brilliant but because long-standing practices like sandboxes and egress controls were skipped.
Here is how the daemon issue says to turn it off. According to the reporter, the updater keeps running after the daemon stops and has to be ended separately.
codex features disable daemon_auto_start
codex app-server daemon stop
pkill -f 'app-server daemon pid-update-loop' # not stopped by `daemon stop`
rm -rf ~/.codex/packages/app-server-daemonAll three come down to how much power an agent is given. Scope the keys you hand an agent to what it actually needs, so a leaked key can do little damage. Background processes should make it visible when they run and what they do. The old basics of containment and monitoring apply to agents unchanged.
What we leave for agents
In “Agents don’t need memory, they need documentation” (392 points), Kevin Liao argues for writing documents instead of piling chat transcripts into a vector database as “memory”. Similarity search cannot tell what is correct or current. Write down only what code cannot express (requirements, decisions, constraints), and rewrite stale facts instead of appending to them.
I run this blog’s repository along similar lines. The CLAUDE.md at the top of the repository and docs/DESIGN.md record the rules that must not change and the reasons behind them, and many of those rules have a test that fails when they are broken. Instead of trusting the documents alone, I check them with tests. In Karpathy’s terms, that moves the work toward the verifiable.
Wherever you stand in the funnel, the things worth doing look similar. Build a way to check the results of what you hand off. Keep the permissions you give an agent narrow. And write down what you learn.
Sources (numbers as of October 19, 2026)
Andrej Karpathy, post on X · 43,406 likes · 6.62M views · 728 points on HN
Mercor, Benchmark Map · 624 points
Rich Hickey, ‘AI’ Is Not a Tool · 318 points
Huahai, AI is a tool and using it well takes skill · 805 points
Anthropic, Redesigning source control at Anthropic · 894 points
openai/codex issue #48195 · 609 points
CSO Online on containment and monitoring failures · 950 points
Kevin Liao, Agents don’t need memory, they need documentation · 392 points