Jobs, war, government power, and a 25% collapse bet, from the Circuit documentary.
Anthropic is worth nearly a trillion dollars. Its CEO puts civilization’s odds of collapse at 10 to 25 percent, and builds harder anyway.
That refusal to flinch is the whole story.
I watched the full Circuit documentary so you can skip it. Here are the 10 takeaways that matter.
One thread runs through them: once AI takes the execution, the edge moves to judgment and go-to-market craft. Fitting, given who powers this issue.
Together with Attio:
Dario’s point cuts to one question: what do you do that AI struggles to replicate?
is the operators answering it, with six new entries live:[GTM Atlas Volume II]
▫️ by Travis Bryant, Anthropic’s Head of US Mid-Market GTM
[What only a human can deliver](http://Now, to Dario, starting with the half of the automation curve almost everyone skips.)
▫️ by Cristina Cordova, COO of Linear (ex-Stripe, Notion) [Hire for slope, not for pedigree](http://Now, to Dario, starting with the half of the automation curve almost everyone skips.)
▫️ by
[Your product is the pitch](http://Now, to Dario, starting with the half of the automation curve almost everyone skips.)
Elena Verna, Head of Growth at Lovable Plus a perk stack from Wispr Flow, Granola, Notion, and Anthropic.
[Steal the systems](http://Now, to Dario, starting with the half of the automation curve almost everyone skips.) the best operators run right now:
Now, to Dario, starting with the half of the automation curve almost everyone skips.
1. The automation curve has a brutal second half
“You automate 90% of the job, great. People are 10 times more productive in the other 10%. But eventually it gets close to 100%. Now, the sequel to that is, well, then you have to find something else for them to do.”
Most analysis stops at 90 percent and calls it a productivity revolution. Dario keeps going. For some workers, AI already replaces rather than augments, and augmentation and replacement turn out to be the same curve, the same people, separated by time.
At Anthropic, Claude writes almost all the code today. Engineers feel superhuman, which is exactly why the endpoint hides. The window to plan for the second half is open now, while it still reads as good news.
Track before the turn:
▫️ Which roles you augment today
▫️ Which roles raise output while headcount holds flat
▫️ Whether your workforce plan accounts for the second half at all
Why it matters: the companies that survive the second half built for it during the first. Plan for the jobs turn ahead of the curve announcing it.
2. The super weapon Anthropic held back
“Some of the early companies that we gave this to said things like, this is a super weapon. You should have to own a gun license to use it. Please don’t release this.”
A project called Mythos sits under every public Anthropic story. It found thousands of cybersecurity vulnerabilities and mapped exploitable flaws across every major operating system. Anthropic’s own read: released openly, it could hack banks, reach state secrets, and hit critical infrastructure.
Dario calls it “a particularly large jump,” one that arrived ahead of any internal schedule. The response was Project Glasswing: controlled access for a few trusted organizations, federal agencies included. It is the first live test of whether a private company can hold that much power responsibly.
Why it matters: enterprise and government buyers now bet on which vendors hold the governance to handle a Mythos-level event. The labs with pre-set release criteria set the precedent the rest follow.
3. One engineer, zero lines of code, 17x growth
“Now I talk to my Claude and it writes the code. And then while it does that, I talk to the next Claude and it writes some code. And at any point I have either a few Claudes running and up to a few thousand Claudes running, doing things.”
Boris Cherny, the engineer behind Claude Code, spent six months coding through Claude rather than typing syntax. This is parallelized AI labor: dozens to thousands of instances running at once on complex tasks. Claude Code’s internal volume at Anthropic is up 17x year over year, and one Code with Claude attendee put it plainly: “I have the confidence of a 22-year-old with VC money.”
The job changed. The winners orchestrate the most parallel agents against the hardest problems: ▫️ Understanding what is worth building
▫️ Judgment on direction and priority
▫️ Directing agents toward the right goal ▫️ Knowing when to override the output
Why it matters: treat this as a tool and you go 10x faster. Treat it as a replacement for thinking and you become replaceable. Judgment survives, execution commoditizes.
4. A 25% chance of collapse, and the airline test
“If there was a 25% chance of an airplane crashing, you wouldn’t get on that plane. That’s right. 25% is too high. We’re trying to make that probability much, much lower. That is the goal.”
He puts collapse odds at 10 to 25 percent and holds the number under pressure. His logic: the technology exists, many actors will build it regardless, and collective action is the hardest problem class in history, so one player opting out shifts little. The better move is to sit at the frontier with the most safety-focused team.
The person building the thing is telling you the risk is genuine. That is the data point. Reducing that number is half of Anthropic’s operating budget, and the same person who calls 25 percent too high builds harder, faster.
Why it matters: a 10 to 25 percent risk estimate belongs in scenario analysis and board conversations as a planning input, from the CEO of the most valued AI company by revenue.
5. Why he left OpenAI, and what actually drove it
“When you feel that you can’t trust someone, when you feel that their values are not what they say they are, when you feel that they’re not honest, that makes it very hard to continue to work with the company.”
Coverage compresses Dario’s exit into safety disagreements. The deeper driver was trust. Safety frameworks invite honest disagreement among smart people. A gap between stated values and actual behavior is a different category, and once it opens, the argument ends.
The number that says the rest: every Anthropic co-founder still works there, which at this scale and under this much Pentagon and competitor pressure is rare. That cohesion is a competitive asset.
Why it matters: trust is the load-bearing wall of every partn
6. The Pentagon blacklisted them, and the lines held
“What does winning this fight actually look like? I won’t even call it a fight. This is more a debate about what the proper use of AI by the government is.”
Anthropic wrote its red lines before the meeting: bans on mass surveillance, human oversight required on any targeting. The Pentagon pushed to strip them, Anthropic held, and the blacklist followed. Lawsuits came, Trump called Dario “an ideological lunatic,” and the ban later lifted with the lines intact.
The lesson sits in the sequence: specificity before the meeting beats philosophy after it.
Why it matters: the companies that built Washington credibility during calm periods gain leverage when the next capability reveal forces a policy response. Define your red lines before the room.
7. A school was hit, and Claude was in the chain
“What we’ve seen here is Claude assists, but a human makes the final call. So a human made that final call, not Claude. Imagine if you had a world in which the AI model just makes the decision and the human never sees it. That’s what we were standing up for.”
A US missile hit a girls’ school in Iran, more than 150 people died, and Claude may have been in the targeting chain. Dario calls the strike terrible and separates two questions most coverage fuses: was the outcome acceptable, and was the decision architecture acceptable.
The counterfactual is the hard part. The alternative to Claude assisting under human oversight is AI from less careful actors making lethal calls with a human fully outside the loop.
Why it matters: “human in the loop” is the load-bearing phrase in high-stakes AI. Every org using AI in consequential calls needs a written answer to one question: who makes the final call, and how do you verify it happened?
8. The bet everyone called crazy
“At that point in time, I know it sounds crazy now looking back, not a lot of people believed, hey, scale-up is the way that these models are going to get smarter and better. That was sort of an unusual, countercultural, scientific perspective.”
Before ChatGPT made scaling obvious, Dario argued it would work: add compute and data to large models and they improve, predictably, on the same algorithm. In 2016 that read as fringe, and the founding team held it anyway. They were right.
The rarer thing is the conviction held under social pressure for years until it paid out. To spot the 2025 version: find the belief researchers hold privately, consistent for two-plus years under pressure, where the holder stays put when the room pushes back.
Why it matters: Anthropic’s whole position traces to one technical conviction held under resistance. The next decade-long return hides inside a belief today’s consensus still calls wrong.
9. $285 billion gone in a day, and why the pie still grows
“The existing incumbents may be smaller in relative terms. Some of them may go down in value. Some of them may even go out of business if they don’t adapt. But I would guess that the software industry gets larger, not smaller, although there will be some big losers.”
Claude Cowork launched, and $285 billion in software market cap vanished in a day. Traders called it the SaaSpocalypse. Dario’s actual view is more useful: total software grows while distribution shifts hard. Those are two statements, and fusing them is how sharp operators misread the moment.
Which side you land on:
▫️ You can name what your product does that AI struggles to replicate
▫️ That answer is specific, defensible, and hard to build around
▫️ Your users would lose something irreplaceable if you vanished
▫️ Your moat lives in the workflow, the data, the relationship, or the judgment layer Try this: complete the sentence “AI can do everything we do except ___.” A specific, defensible ending is your moat. A blank is the work to finish before the next launch ships.
10. Silicon Valley yo-yoed on regulation. Dario held the third path
“They started with a position of, you know, even having transparency around this technology is all just totally apocalyptically destroy our potential. And then as soon as they see the first real danger, there’s all this talk of nationalization and the government should just seize it. Come on, folks.”
Silicon Valley swung from “any regulation kills us” to “nationalize AI” inside one capability reveal. Dario stayed put. His position is specific: the technology rides a smooth exponential, it keeps getting more capable, and governance belongs in place ahead of the crisis.
His mechanism:
▫️ Pre-release testing and auditing before any major model ships
▫️ Accountability that skips government ownership
▫️ Transparency enough that outsiders verify what labs claim
▫️ A line between “the government should know” and “the government should own”
Why it matters: the policy fight stays open, and the next reveal forces another response. Show up with consistent, specific positions. The ones who held them help write the rules.
What this means for you #
Dario runs the most valued AI company by revenue, believes civilization faces a meaningful chance of collapse, and builds harder anyway, because handing the lead to someone less careful is the worse outcome.
Three principles to carry forward:
The automation hump reads as productivity until it turns. Build for both halves while it still feels like good news.Trust is the actual reason partnerships break. Values that drift from behavior end things faster than any strategy fight.The smooth exponential feels like stillness until it goes vertical. Most people are standing in that flat stretch right now.
- Founders:* build something that compoundson the right side of the curve. Physical world, human judgment, or directing AI at scale: pick one lane deep enough that the next Claude launch makes you
[harder to replace](https://www.thevccorner.com/p/what-top-vcs-look-for-2026-founder-playbook?r=1krivi). Write your moat answer down before the next reveal.
* Investors:* Anthropic went underdog to
[trillion-dollar](https://www.thevccorner.com/p/the-most-valuable-vc-backed-startups?r=1krivi)in four years on two contrarian bets. The next decade-long return hides in a belief the
[current consensus](https://www.thevccorner.com/t/investor-lists?sort=top)still calls fringe. Back the person who holds it and stays put.
* Operators:* the
SaaSpocalypserepeats. Map where your users lose irreplaceable workflow context if you vanished. That is your moat.
Quiet, quiet, quiet.
And then zoom.
Plan for the zoom.
Keep reading #
Anthropic, Claude, and the power fight
▫️ [Claude and Anthropic library](https://www.the-ai-corner.com/t/claude-and-anthropic?r=1krivi)
▫️ [Inside the Fable 5 and Mythos government ban](https://www.the-ai-corner.com/p/anthropic-fable-5-mythos-5-government-ban-2026?r=1krivi)
▫️ [Anthropic passed OpenAI at $30B ARR](https://www.the-ai-corner.com/p/anthropic-30b-arr-passed-openai-revenue-2026?r=1krivi)
▫️ [Dario Amodei on safe AI and AGI](https://www.thevccorner.com/p/dario-amodei-safe-ai-agi-anthropic?r=1krivi)
Agents, Claude Code, and the new job
▫️ [The Claude Code system that replaces your dev loop](https://www.the-ai-corner.com/p/the-claude-code-system-that-replaces?r=1krivi)
▫️ [Loop engineering for coding agents](https://www.the-ai-corner.com/p/loop-engineering-coding-agents-2026?r=1krivi)
▫️ [Claude managed agents guide](https://www.the-ai-corner.com/p/claude-managed-agents-guide-2026?r=1krivi)
▫️ [The Anthropic AI jobs report](https://theaicorner1.substack.com/p/anthropic-ai-jobs-report-2026?r=1krivi)
Moats, markets, and the SaaS shift
▫️ [SaaS defense playbook for the AI era](https://www.the-ai-corner.com/p/saas-defense-playbook-ai-era-survival-guide-2026?r=1krivi)
▫️ [Marc Andreessen on the AI moat](https://www.the-ai-corner.com/p/marc-andreessen-ai-moat-not-the-model-2026?r=1krivi)
▫️ [The biggest VC-backed startups](https://www.thevccorner.com/p/the-most-valuable-vc-backed-startups?r=1krivi)
Full documentary
If this breakdown saved you 47 minutes, share it with one founder or investor who needs to see it. They will thank you later.