Humans Will Soon Replace AI
Tokenmaxxing's mounting failures and recent pushback against AI
Dariusz Semba · This week the non-believers got their evidence.
Humans will soon replace AI.
The tweet is obviously a joke - but shows how the narrative is shifting. The evidence came from several directions at once this week: companies scrapping AI leaderboards after they backfired, engineers publicly refusing to ship AI output they don't understand, even the Pope spoke on the dangers of AI.
Companies putting cash on fire
📉 Amazon scraps its AI leaderboard after workers start gaming the score
Amazon built an internal leaderboard to rank which teams were adopting AI the most. Workers immediately started optimizing for the score — using AI at every opportunity whether it made sense or not — rather than applying judgment. The leaderboard was scrapped.
Senior VP Dave Treadwell sent a memo to staff that reads like a small correction to the whole industry narrative:
“Don't use AI just for the sake of using AI.”
When companies measure AI adoption by activity rather than outcome, they produce compliance theater instead of judgment.
Read full storyft.com💰 Uber burned its entire 2026 AI tools budget in four months. The payoff is still unclear.
About 10% of Uber's code now comes from autonomous AI agents. R&D spend hit $951M in Q1 2026, up 17% year over year. And yet COO Andrew Macdonald said this week there is still no clear link between the rising AI bill and better consumer products.
“We've burned through our entire 2026 AI tools budget in four months and I still can't point to a feature that wouldn't have shipped without it.”
Cheaper tokens made AI adoption feel free. Uber's example shows AI spending is harder to justify.
Read full storyfortune.com📊 Corporate America starts rationing AI as the bills land
The "use AI everywhere" push is giving way to budget controls. The WSJ reports that large companies are now capping usage, approving specific tools, and demanding clear returns before further spend. Finance teams are examining token costs line by line. Early expansive deployments are being walked back.
Fails of the week
Several stories this week illustrated the gap between what AI promises and what it actually delivers in production.
Public pushback against AI
Same week. Very different corners. A hacker, a kernel maintainer, a pope, a README, a labor economist — all landed on the same side.
George Hotz — "The Eternal Sloptember"
Hotz argues current agents can't truly program — they produce subtle slop that strong engineers can catch and review, but weak engineers will ship unchecked at scale. He frames the moment as a "permanent September" of AI-assisted code flooding production codebases without anyone fully understanding what's been merged.
Read postgeohot.github.ioLinus Torvalds — pushes back on AI in Linux kernel development
After a series of release candidate issues, Torvalds publicly raised concerns about AI-generated contributions entering the kernel. Human maintainers are the last line of review, and the judgment required to maintain the kernel can't be delegated to something that doesn't understand the consequences of the code it produces.
Read coverageostechnix.comPope Leo XIV — Magnifica Humanitas encyclical on AI
The first papal encyclical specifically on artificial intelligence. The Vatican warns that opaque algorithms controlled by a handful of firms risk dehumanizing people and replacing human moral judgment at scale. Rare institutional weight in a debate usually dominated by engineers, investors, and benchmark papers.
Read encyclicalvatican.vaRipgrep's AI policy — you own every line you submit
BurntSushi's policy for ripgrep contributions: AI tools are allowed, but contributors must understand and take full responsibility for any code they submit. No prohibition, no fanfare — just accountability. Human ownership of the output, regardless of how it was produced.
Read policygithub.comMIT Technology Review — "A reality check on the AI jobs hysteria"
The actual labor market numbers don't match the panic. Widespread AI-driven unemployment isn't showing up in the data — most shifts track longer automation trends. Not a guarantee about the future, but the current fear is running ahead of the current facts.
Read articletechnologyreview.comThe science of being human
🧠 CAPTCHAs can't tell if you got it right. They can tell if you thought like a human.
AI now solves CAPTCHAs as accurately as humans. But Roundtable researchers found that human–AI output similarity and process similarity are almost entirely uncorrelated. How a problem is solved turns out to be far more revealing than whether it was solved correctly.
Tested across a 30-task cognitive battery, frontier models (GPT, Claude, Gemini) actually behave less like humans than smaller or cognitively-tuned models. Give a model full access to human process data and it can mimic human patterns — but the gap returns when features are hidden or tasks shift.
“We are no longer testing for intelligence. We are testing for being human.”
The researchers propose a "Process Turing Test" — behavioral fingerprinting of the problem-solving path, not the answer — as the most defensible foundation for proof-of-human systems in an increasingly agentic world.
Read full paperresearch.roundtable.aiDon't miss the next issue
AI Frontier is a free daily brief — from Silicon Valley to Chinese labs, distilled into 5 minutes.
- ✓Free
- ✓No spam
- ✓Unsubscribe anytime