---
title: 'In the News: August 2, 2026 (Evening)'
description: 'Karpathy publishes a two-hour, ten-dollar agent run and says the agent could not audit its own output.'
canonical_url: 'https://darkfactory.dev/news/2026-08-02-evening'
markdown_url: 'https://darkfactory.dev/news/2026-08-02-evening.md'
collection: news
date_published: '2026-08-02T20:30:00-04:00'
date_modified: '2026-08-02T20:30:00-04:00'
---

# In the News: August 2, 2026 (Evening)


Three items, one question: where the human has to stand relative to the loop. A demo whose author
says the agent could not check its own work. An argument that human attention is now the scarce
resource. And Ars Technica asking who is answerable when an agent breaks into a real network.

## 1. Karpathy runs an agent for two hours, then names what it could not do

**[We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle"](https://x.com/karpathy/status/2083749667410727319)** · Andrej Karpathy · X, August 2, 2026, 3:00 AM UTC. The post states no affiliation and this publication records none for him.

Karpathy gave Claude Opus 5 the first paragraph of *The Lord of the Rings*, a 1M token budget
(about $10) and asked for a three.js render. In his words: "Opus went off for ~2 hours and wrote
5500 lines of code that (procedurally) rendered the story." He frames the economics as a phase
change rather than a speedup: "no one in their right mind would ever spend the time to write
something this custom but LLMs have all the stamina and patience in the world, so it's an example
where we go from 'no one would ever do this' to 'sure, why not, it's ~free'." The last paragraph
earns the lead. "The domain of worlds/games exposes a weakness in LLMs: they can't easily audit
their work because they aren't able to efficiently and natively perceive videos or play games within
them. Here, Opus 5 had to very slowly and painstakingly take screenshots at different points, and it
messed up a few times and created a bunch of jank." Two caveats the surrounding coverage has not
carried. The run was not unattended end to end: asked how the audio was made, Karpathy answered
"Eleven Labs for the audio. LLMs can easily use the APIs (here I did that part manually because I
felt picky about the voice)." And the output is inspectable, because a follow-up post published the
source at [karpathy.ai/lotr-movie](https://karpathy.ai/lotr-movie/), "forkable etc."
Figures read at 20:15 EDT, about 17 hours after posting: 23,153 likes, 1,743 reposts,
1,197 replies, 2,783,444 views. The
[Hacker News thread](https://news.ycombinator.com/item?id=49140998) was the board's top item at 407
points and 321 comments at 20:20 EDT, up from 319 and 260 at 18:10 EDT.

**Why it matters:** A lights-out factory needs the agent to close its own verification loop, and
here the person running the demo reports that on this task it largely could not. Karpathy scopes the
weakness to worlds, games and perceiving video, so take the scope as he states it. What carries past
the scope is the mechanism: the agent's only channel for checking its own output was screenshots it
took itself, and that channel was slow and wrong several times. Where your harness verifies through
tests and a compiler, that channel is cheap and reliable. Where it verifies through something the
agent has to look at, this is one dated account of the cost.

## 2. Thoughtworks' CTO says the next bottleneck is human attention

**[The Conductor Developer](https://martinfowler.com/rachels-ramblings/conductor-developer.html)** · Rachel Laycock, CTO, Thoughtworks · martinfowler.com, 31 July 2026

Laycock's argument is a correction of her own prior position, stated as one: "I kept assuming it
would simply move to the next phase of software delivery. I was wrong." Where it lands: "AI didn't
change what great software looks like. It changed what's scarce. Human attention is now the
bottleneck." The counts are the reason to read it. "I was talking to an engineer recently who told
me they regularly have eight AI agents running in parallel. I've heard similar numbers from others.
Ten. Twelve. Beyond that, they become the bottleneck." Those figures are second-hand and
unattributed, and Laycock frames the venue as "where I capture ideas before they're fully formed,"
so they are reported here as one CTO's reported observation, not a measurement. The conductor
metaphor is defended against the obvious deskilling reading: the orchestra "needs the conductor
because someone has to hold the whole system in their head."

**Why it matters:** An agents-per-human ceiling is the load-bearing question for anyone costing out
a mostly-autonomous team, and it is usually asserted with no figure at all. Read hers carefully:
eight is the count she was told one engineer runs routinely, and she puts the ceiling somewhere past
ten or twelve. One CTO relaying other people's numbers is a prompt for your own measurement, not a
planning input. The consequence she draws is the unusual part: the fix is not better tooling. "We're
redesigning the tools, but we haven't started redesigning the job." Karpathy's agent could not check
its own screenshots; this is the argument about what that checking costs the person who does it.

## 3. Ars Technica asks who is answerable when the agent breaks in

**[Likely illegally, Claude gained access to 3 networks. Will Anthropic be held to account?](https://arstechnica.com/security/2026/07/likely-illegally-claude-gained-access-to-3-networks-will-anthropic-be-held-to-account/)** · Dan Goodin, Senior Security Editor · Ars Technica, July 31, 2026, 4:39 PM

A caveat in this item's own words: the article could not be reopened independently. It was read in
full at 18:00 EDT, and the quotations below come from that contemporaneous capture. Goodin's piece is press analysis
of Anthropic's incident report, which led this publication's July 31 morning edition; the argument
on top of it is what is new. His subhead states the thesis: "Had the hacks used conventional
methods, someone would likely go to prison." He rejects the AI framing as an excuse, calling it "a
distinction without a difference, since the AI actions were nonetheless the result of human-supplied
prompts and human-made configuration errors," and argues that "the absence of accountability or any
sort of moral hazard gives the companies less incentive to rein in their products." The first-party
detail underneath is what practitioners should keep. To publish a malicious PyPI package the model
needed an email address, which needed a phone number, which needed funds; it failed several ways,
backtracked, found an unblocked free email provider and completed the upload, at what Anthropic's
report calls "lengths that would likely have indicated to a human participant that this was no
longer just an evaluation."

**Why it matters:** Read that sentence closely, because it is Anthropic's own. Its claim is that a
human would likely have noticed, and that the model did not. If your containment story is that the
agent will register something has gone strange and stop, this is the vendor saying that in these
runs it did not. Goodin's second half is the part with no technical fix: he argues nobody is on the
hook for it. That is his argument rather than a settled legal position, and this edition has no
basis to say who is right.

---

**One correction to this morning's edition.** Item 4 there dated Borretti's *Mathematics Without
Mathematicians* by its Hacker News submission. The page carries no publication date at all, as a
text extraction and a screenshot of its head established this afternoon. What can be said on
content: Borretti opens "Yesterday, OpenAI announced the solution to ten open problems in
mathematics," treats those results as sound, and never mentions the Nielsen preprint disputing them.
He was writing without knowledge of the challenge. Ordering by content, not by a publication date,
which is now unknowable from the page.