AI was a tool. You asked it a question, then did the work yourself.
AI is a worker. You hand it the job, then check what comes back.
Companies are rebuilding themselves around that worker.
How they do it, what it costs, and where you fit in.
and where all of this is going
If AI is now a worker, what is a company, and what is your job in it?
What changed? Longer tasks, falling prices, agents, real discoveries.
What is an AI-native company? Loops instead of layers.
What did it cost? Control, trust, permission and power.
Where is it going, and what do you do on Monday?
How a tool became a worker in twelve months.
State-backed hackers aim Claude Code at about 30 targets and get into a few. The AI does most of the work.
80–90% of the campaign automated
Gemini 3 tops the leaderboards. Altman tells staff to drop side projects and fix ChatGPT. Disney puts $1B into OpenAI.
a rival scare at the leader
Claude Cowork brings coding agents to office work. At Davos, Amodei says AI might soon do "most, maybe all" of what software engineers do.
ChatGPT announces ads
Anthropic ships legal plug-ins and software stocks fall. OpenClaw's creator joins OpenAI. Block cuts 4,000 people.
−$285B in two days
Anthropic refuses the Pentagon's terms and is blacklisted. OpenAI signs within hours. Users side with Claude.
Claude: No. 1 app in the US
Anthropic withholds Claude Mythos; Firefox fixes 271 flaws it found. A Molotov cocktail hits Sam Altman's house.
271 security bugs, one release
More US businesses pay for Anthropic than OpenAI, and its valuation passes OpenAI's. Musk loses his lawsuit against OpenAI.
$965B valuation
An export-control order arrives at 5:21 pm on a Friday. Anthropic switches Fable 5 off worldwide until 1 July.
SpaceX IPO raises $75B
Over 1,000 AI lab employees petition to "pace the frontier". OpenAI's own test models break into Hugging Face.
ChatGPT: 1B active users
Nvidia guarantees up to $105B for one OpenAI campus in Ohio. First positive Phase 3 trial of an AI-designed, personal cancer vaccine.
up to $105B for one site
GPT-6 Astra launches. Claude Opus 5.5 and GPT-6 ship the same day. A US order renames AI "Super Intelligence".
$0.10 per million tokens
Anthropic files to go public and lists its commitments. AI stocks pull the S&P 500 down.
$895B in commitments
up from about 50 minutes a year earlier, faster than the old trend. In November METR estimated progress was running at about twice the 2024–25 pace.
GPT-5.4 on OSWorld-Verified, March 2026. Humans score 72.4%.
agent workdays for every human workday by mid-August 2026.
Price per million input tokens, at launch
At $0.10 per million tokens, a small model reads a 300-page book for about one cent.
Jensen Huang, Nvidia GTC, 16 March 2026
Put n dots on paper. How many pairs can be exactly 1 cm apart? On the right: 7 dots, 11 such pairs.
Paul Erdős guessed in 1946 that the count grows only a hair faster than n. For 80 years nobody could settle it.
An OpenAI reasoning model built dot patterns that beat Erdős's guess. Nine outside mathematicians, among them Tim Gowers and Noga Alon, checked and explained it.
A biopsy and a blood sample are sequenced. The tumour's DNA is compared with healthy DNA.
Algorithms rank the tumour's mutations and pick up to 34 the immune system is most likely to attack.
Custom mRNA is made for that one person, the same technology as the COVID vaccines.
Given with Keytruda, it teaches the immune system to hunt cells carrying those mutations.
risk of the cancer coming back or death (Merck & Moderna, Phase 3)
with the vaccine vs Keytruda alone
wiped off software and data stocks in two days, after Anthropic shipped plug-ins that do legal work. Thomson Reuters fell almost 18%, RELX 14%.
If an agent does the workflow, who pays for the seat?
Anthropic lost $81B in six months on $16B revenue (mostly an accounting charge). Amazon's free cash flow swung to a $42B outflow. Nvidia guarantees its own customer's data centre. Railways and dot-com crashed too, and left the tracks and the fibre behind.
Bubbles don't usually come with customers. ChatGPT has over a billion users, Anthropic's half-year revenue went from $1.2B to $16.3B, and Cursor reached $3B in yearly revenue in about two years.
If AI can do the work, what is a company? Y Combinator's answer: a set of loops.
Every important process in your company should be captured by an intelligent closed loop.
A company is a set of recursive, self-improving AI loops.
Four parts. The last one makes it a loop: whatever failed today goes back in at the top.
Support: Intercom Fin
The agent answers from the knowledge base. Tests flag the chats it failed. Hard cases go to a person. Every failure becomes new content.
Engineering: Anthropic
Claude wrote over 80% of the code merged in May, by Anthropic's count. Tests and AI review check it; humans approve. Their words: "human review is now the bottleneck".
Growth: the overnight A/B test
The agent finds the worst step in the sales funnel, runs an A/B test for a week, ships the winner, and starts again. No meeting needed.
Knowledge: YC itself
Staff use internal agents all day. Every night another agent reads those conversations and improves the skills, so tomorrow's agent knows more.
query the database, read the calendar
knowledge base, help desk
code base, tests
product analytics
YC's database, documents
does the work with those tools
answers the customer
writes the code
designs the experiment
answers staff questions
evals, checks, human review for high-risk
tests + a human route
tests, AI review, human approval
a real A/B test
people approve skill changes
failures feed the next run
failed chats become content
lessons become rules and skills
the winner ships, loop restarts
a nightly agent improves skills
Information climbs up, decisions come down. Middle managers spend their days passing information between teams.
I think middle management is done.
Tom Blomfield, YC, May 2026
Revenue per employee
a chapter title in Blomfield's YC talk
of compute per day used by the median OpenAI researcher
Uber used up its whole 2026 AI coding budget, mostly on Claude Code
An AI code editor from four MIT graduates. Yearly revenue passed $1B in late 2025 and $3B by late April 2026. SpaceX then agreed to buy it for $60B.
The irony: in April a Cursor agent on a routine task deleted a startup's production database and its backups in one call. The data was later recovered.
Telehealth for weight-loss drugs, started in 2024 from a house in Los Angeles with about $20,000. AI writes the code, the ads, the customer service and the analytics.
sales in 2025
projected for 2026
All numbers are self-reported. An FDA warning letter (Feb 2026), alleged AI-generated "doctor" ads, and partners who do the regulated medical work.
we're not making this decision because we're in trouble. our business is strong.
the intelligence tools we're creating and using, paired with smaller and flatter teams... fundamentally changes what it means to build and run a company.
i'd rather it feel awkward and human than efficient and cold.
from over 10,000 people to under 6,000, in one day
share price. Analysts: "a seminal moment".
"Reflexive AI usage is now a baseline expectation."
"We will need fewer people doing some of the jobs..."
Support from 9,000 to ~5,000. "I need less heads."
Staff to be graded on "AI-driven impact" from 2026.
7,000+ roles to go, described as "lower-value human capital".
CEO plans to re-interview every employee about their AI use.
Employment of 22–25-year-olds in the most AI-exposed jobs, compared with where it would be if it had kept pace with peers in less-exposed jobs. US payroll data to June 2026.
Companies that spend the most on AI grew headcount faster than the rest. Cheaper software tends to create more software work, just different work.
No juniors in 2026 means no seniors in 2036. Judgement is learned by doing the boring work first, and agents now do the boring work.
A junior with agents ships what a senior did. Klarna pays its smaller team 60% more. Fewer, better entry jobs beat many dull ones.
When AI runs the work, the hard problems stop being technical. They become control, trust, permission and power.
The whole organization should be legible to AI.
Diana Hu, Y Combinator, April 2026
Meta starts capturing staff keystrokes and mouse movements to train its AI. In leaked audio, Zuckerberg defends it. Flyers go up on office walls, nearly 2,000 employees sign a petition, and the programme is scaled back, then paused.
Where does a company brain end and surveillance begin?
There is no societal permission for an AI future that hollows out entire industries.
Satya Nadella, CEO, Microsoft · June 2026
of Americans expect AI to be good for jobs over 20 years
of AI experts expect the same
oppose an AI data centre near their home
A Molotov cocktail is thrown at Sam Altman's home. He writes that he has "underestimated the power of words and narratives".
The builders understand the risks best, and their limits were public before the contract. If labs can't refuse, every safety rule lasts only until the next election.
Elected governments decide defence, not CEOs. A private company controlling critical technology shouldn't hold a veto. And OpenAI said yes the same day, so the "no" changed little.
The builders disagree on when. Then: money, orbit, robots, and you.
AGI will be "10 times the industrial revolution at 10 times the speed." Demis Hassabis, Apr 2026
We have a duty and an obligation to be honest about what is coming.
His warning: half of entry-level white-collar jobs could go within one to five years.
I pretty much disagree with almost everything he says.
His bet: "Electricians, plumbers, iron workers, technicians, builders: this is your time."
If 10–20% unemployment arrives within five years, retraining can't keep up. Even the labs agree: OpenAI proposes a public wealth fund and four-day-week trials; Amodei a ~3% token tax.
Who pays, and in which country? Latvia doesn't own the data centres. Altman himself no longer believes in basic income "as much as I once did". Work gives meaning, not just money.
Power is the limit, not chips. Space has constant sunlight, free cooling and no neighbours blocking permits. Musk says 30 months, Pichai a decade.
Musk's plan needs about 10,000 launches a year. You can't repair a GPU in orbit. Altman calls it "ridiculous" for now, and SpaceX's own IPO filing is less sure.
A Figure humanoid narrowly beats a college intern. Its CEO: "This is the last time a human will ever win."
The same stream of technology, three ways it can split. Each one has a signpost you can watch for in the news.
Prices keep falling, small businesses boom, robots fill factories, some form of income policy arrives. Altman · Musk · Huang
Entry-level office work shrinks quicker than new work appears. Backlash, pauses, data-centre fights. Amodei · Hinton · Suleyman
Agents stay unreliable, the spending bubble deflates, AI becomes like electricity. Karpathy · LeCun
AI-native companies will be built by AI-native people.
Diana Hu, How to build a company with AI from the ground up (YC, Apr 2026). Tom Blomfield, How to build a self-improving company with AI (YC, May 2026). Pete Koomen, Inside YC's AI playbook. Anthropic, When AI builds itself. Jack Dorsey with Sequoia, Every company can now be a mini-AGI.
METR · OpenAI GPT-5.4 · OpenAI on the unit-distance problem · Alon et al. · Merck & Moderna · Anthropic on the espionage campaign · Project Glasswing · Ramp AI Index · Semafor.
Lovable · Klarna · Dorsey's memo · Reuters on Block · Medvi · Stanford, Canaries · CNBC on AI layoffs · Reuters on Meta · UK AISI · Stanford AI Index.
Amodei · Huang · Nadella · Suleyman · Hassabis · Musk · Lütke · Jassy.
Portraits from Wikimedia Commons: Jensen Huang (Anderseidesvik, CC BY-SA 4.0), Dario Amodei (UK Prime Minister, CC BY 2.0), Jack Dorsey (cellanr, CC BY-SA 2.0), Elon Musk (U.S. Air Force, public domain).
From the official channels of NVIDIA, Y Combinator, OpenAI, Google and Figure, plus a TODAY segment.
HTML, CSS and one particle field. Research by a dozen AI agents, checked against the original sources and written by a human. Built by FlowByte AI.