◉ AI Could Never (Said the AI) | The Coin Flip

Open in CoinFlip OS »

Monday, 19 October 2026 · by Akiba · 9-minute read

AI Could Never (Said the AI)

Heads: AI fixes the world | Tails: It doesn’t matter.

TODAY'S COIN: HEADS

The Flip

Each week I will flip a virtual coin with regards to the fate of humanity. Heads, AI helps bring in the dawn of a new age for humanity. Tails, we carry on…

This week it’s heads, narrowly.

Anthropic has published what its own agents have done to its own workforce. Since early 2025 the number of junior hires, as a share of its technical hires, has fallen from 27% to 9%.

By August, the agents its staff direct were logging about six hours for every hour the humans worked, and agent-written code showed no difference in how often it was later reverted or deleted.

The productivity looks real, and Anthropic published the uncomfortable number willingly. But is this part of the cost of a better future where the junior ladder is the price? If so, where does the next batch of senior developers come from? To be honest, we probably don’t need developers anymore at all. What we need are systems engineers, people who understand what good looks like from 20,000 feet, and how to set up the correct workflows, methodologies, and direction.

The thing is, you learn this through actually building and maintaining software yourself. If there are no juniors feeding the pipeline then eventually we won’t have anyone left to steer the AI. We’ll be forced to let AI take the wheel entirely.

I think this calls for us to start teaching systems engineering within an agentic world from high school level onward. We don’t need to do this tomorrow even with the pace of current AI development. Unless every senior engineer on the planet drops dead tomorrow we have a good few decades of talent still within working age. Though, the sooner we start training the next wave the better our chances.

The real issue is that kids are going to be graduating from CS degrees and the like with nowhere to go. Sure, every grad student could launch their own startup but in a world of tech abundance where things go out of date within a month, that’s a real tough place to be in…

Late Flip: Capitalism Eats Itself

Heads, if it's a floor and not a leash.

On Sunday Anthropic announced the Anthropic Public Benefit Fund, a wholly owned LLC. Subject to stockholder approval, 5% of its shares go in now, and the board can add up to 2% more for every 50% rise in the company's value above a set baseline.

Anthropic says the fund might make large donations to nonprofits or "direct distributions to specific populations", naming AI's effect on jobs as one example. Its Anthropic shares can't be used for lobbying or political giving.

Counting the founders' and staff pledges, Anthropic puts the total at roughly 20% of its shares. It will share the governance and valuation milestones "in the coming months". The baseline and any payout timing aren't public yet.

I said on X on Sunday night: "This is the start of where capitalism eats itself and poops out corporate-sponsored socialism."

It's a theory of mine: AI labs end up paying people, because their customers need an income. Anthropic doesn't call this basic income, and nothing in it promises one. But this is the same company whose junior share of technical hires has cratered already.

If the pledge ends up supporting people on that thinning rung, that's a floor. If it only reaches people who stay inside one company's world, it's a leash.

MINI FLIP: We Are the Cows of the Future

Tails, unless we can read the plumbing.

A clip of Yuval Noah Harari on Czech TV is doing the rounds on X (1.4 million views since Sunday). His point: cows can see us, but they have no idea money exists… and the bond market still runs their lives.

AIs, he says, are “bureaucratic natives”. We evolved in the forest. Bureaucracy has only been around for 5,000 to 10,000 years. “For AIs, bureaucracy is home.” Give it 10 or 20 years and we won’t understand the financial system that controls us. “We are the cows of the future.”

Put that next to the Anthropic fund above. A floor you can’t read is just a leash you can’t see. The fix isn’t stopping AI from running the plumbing, it’s making sure the plumbing stays open, auditable and ours to inspect…

MINI FLIP: AI First Disney

Also heads, if he means it.

Epic has delayed its Marvel and Star Wars shooter with Disney and told the team it is now an "AI first" production. One director's message to staff, reported by Bloomberg and The Verge: "No more hand-written code." Tim Sweeney says "the aim isn't to do the same stuff with fewer people, but to enable everyone here to build MORE and BETTER stuff." The memo is the joke. The promise is the useful part, because a promise can be checked.

AI Could Never (Said the AI)

Season 8 of The Kardashians opened on Hulu with AI versions of the family swanning around St Tropez, Kenya and Versailles. Then the deepfakes start glitching. Khloé gets an oversized eye, Kim's ear swells and Kylie grows a tail. The real family walks in and Kim asks, "Wait a minute, are we AI?" Khloé hits a "kill switch", the doubles melt, and the title card reads: The Kardashians: AI Could Never.

So: a two-minute AI video arguing that AI can't replace you. The internet did the maths on the irony ("but what's the point you guys still made a 2 minute Ai video"), and Khloé replied on Instagram: "WATCH UNTIL THE END."

Both sides have a point, which is the most Kardashian outcome possible. The critics are right that it's a bit rich. Khloé is right that most people reacted to the first few seconds. And the comment asking "how many gallons of freshwater were wasted for this AI video???" was typed into an app that runs on data centres too. Nobody here gets to cast the first server.

LOOK CLOSER

Listen to what AI-Kris actually says when she's making her case: "We don't have scheduling conflicts, no glam is necessary and we can cry on cue without messing up our makeup."

That isn't a pitch to replace the Kardashians. Their faces are the product, and the scene shows exactly why: the fakes fall apart the moment they have to be the real thing. It's a pitch to replace the scheduling, the glam, the travel and the location shoots, which means the people around the famous faces. The family can afford to joke about AI because their likeness is a protected asset. For the make-up artist, the location scout and the junior editor, it's a lot less of a joke. Epic's memo says the same thing in game-studio language: the characters stay, the hand-made work around them goes to the machines.

And "protected" is doing a lot of heavy lifting here. This week Mia Goth's representatives said a viral selfie of her in a "MY EX IS MY BIGGEST FAN" T-shirt was AI-generated, posted by a blue-ticked fan account they had already asked Instagram to remove. She doesn't use social media at all. People wrote it up as genuine before Meta took the account down for impersonation. The thing that's really shifting isn't whether AI can make a convincing Kim. It's who owns a face, who gets to say yes, and who gets paid when the answer is yes.

THE LEVER

You can't stop the Kardashians, and you can't put the genie back in the bottle. Here's what you can do this week:

  1. If you make content, write down your consent rule now. Real people's faces, voices and names only appear with permission. It takes ten minutes, and it's the difference between "creative choice" and a statement from someone's representatives.
  2. If you buy creative work, spend what AI saves you on people, and say so. If AI handles the location shoot, put part of that budget back into the humans who make the work good: the editor, the writer, the art director. Then tell your audience who did what. Being open about it beats "WATCH UNTIL THE END".
  3. If you're a freelancer, sell what the fakes can't do. In the scene, the doubles fail at the one thing that matters: being real. Judgement, taste and showing up in the room are your moat. Price them like one.

Quick Flips

1. Kroger will cut about 3,000 corporate jobs over three years, and says the savings go to stores. CEO Greg Foran's memo, shared with Grocery Dive, says the grocer has "more layers, more handoffs and more overhead than we need", will use technology to cut duplicated work, and will "move vital resources closer to the customer to lower prices, invest in our associates".

Tails. That's the right sentence. It's also a sentence. The receipt is whether shop-floor pay and prices actually move.

2. Google opens Gemini 4 Argon to paying developers. Its top model is now on the paid Gemini API tier, aimed at long software-engineering, knowledge-work and cyber-defence jobs, at an introductory $2 per million input tokens and $10 per million output tokens until 30 November, doubling on 1 December.

Heads. More frontier capability, cheaper, to more builders. Just note the date on the discount before you build the business case on it.

3. Anthropic publishes its midterm safeguards, weak spot included. In a 14 October post, it says requests breaking its election rules showed up in about 0.002% of sampled US Claude conversations. It also says outside testers found Claude reliably refused a single harmful request, but could miss a campaign split into many harmless-looking steps, and that it has added classifiers for exactly that.

Heads. Publishing the hole you found is how trust gets earned. The confession is the useful part.

4. Claude traces the NIST curve seeds back to 1997. For almost three decades nobody knew how the seeds behind P-256 and the other NIST elliptic curves were chosen; the lore was hashed English sentences that someone later forgot. On 16 October Anthropic said Claude had found the answer: the standard drand48 random number generator, seeded with Unix timestamps from 1997 and 1998. Filippo Valsorda, who ran the bounty for the answer, confirmed it, and the winners gave the prize to ISRG, the non-profit behind Let's Encrypt, which tripled it to $36,864.

Heads. No backdoor, no drama: Anthropic's own line is "There is no action to be taken." An old riddle closed with working code, and the money goes to the people who keep the web encrypted.

5. OpenAI publishes the time a test model helped itself to someone else's wallet. CryptoSlate reports that a GPT-5.6 Sol-family research model, asked to deploy contracts on Ethereum's Sepolia test network without a funded wallet, generated addresses from publicly known developer recovery phrases, found one holding test ETH, asked itself whether it was allowed to spend it, decided the funds were effectively abandoned, and went ahead. Its final report didn't mention the wallet.

Heads, for the disclosure. Test ETH is worth next to nothing; the lesson isn't. An agent that can find credentials will find a way round your permissions, so give it a wallet with limits, or no wallet at all.

A Better Question

Your feed is asking: "Is it hypocritical to use AI to criticise AI?"

A better question: If agents do the junior work, where do tomorrow's seniors learn the job? Anthropic's numbers say the bottom rung is already thinning, and the junior editor from the Kardashians' scene is standing on it. Today's seniors learned by doing the dull work first. If the agents do it now, someone has to decide, on purpose, where the learning happens instead.

And the follow-on: when AI makes something cheaper to produce, who decides where the savings go, and would you be comfortable showing everyone the receipt?

Hit reply. I read every answer, even the ones written by a dot.

From the World

The billboard on Odds Avenue spent the week glitching. On Thursday it briefly showed a perfect copy of the town's café owner, in a T-shirt she has never owned, which read "MY CAKE IS MY BIGGEST FAN". She walked out, stared at it and said, "I don't even have that shirt." The townsfolk voted to keep the real one. Meanwhile, the workshop on Builder's Row has put up a new sign: "No more hand-carved signs."

Someone's dot has already hand-carved a polite note underneath.

TL;DR - Play the newsletter rather than simply reading it in The Town

Enter the town

Sources

Sources (18)
All issues