The Mood

Three things were unreliable this week: the platforms, the tools, and the numbers.

The highest-scoring story of the week was a blog post about git commands. 2,216 points. The second-highest was someone filing the corners off their MacBook with sandpaper — 1,360 points, 647 comments, a thread about the tactile pleasure of physical ownership. Both ran in the same week a Molotov cocktail was thrown at Sam Altman's house.

The trust thread wasn't quiet about it. "Sam Altman may control our future — can he be trusted?" opened Tuesday at 2,081 points and 881 comments. By Sunday: Anthropic had silently cut cache TTLs six weeks earlier without announcement (461 pts, 355 comments — discovered late, not quietly accepted). Benchmark gaming against the leading agent evaluation suites had been systematically documented. Small models independently reproduced the Mythos security vulnerabilities that frontier capability had been used to demonstrate — 1,228 points. A frugal founder published a full $20/month stack running multiple businesses at $10K MRR each.

The git commands and the filed MacBook corners aren't coincidence. When the numbers above you are unreliable, people reach for things they can verify themselves.

The Arc

From a governance question to a benchmark reckoning in six days.

Tuesday–Wednesday. The Altman trust piece peaked. Abstract — governance, power concentration, the absence of accountability structures. Same day: Anthropic's OpenClaw subscription ban hit 1,097 pts and 828 comments. Someone paid for a thing. It was taken away. The question stopped being theoretical.

Wednesday–Thursday. Four more stories, same accumulation. EFF left X — 965 comments on 1,149 points. A Vercel plugin asked to read your prompts. Microsoft terminated VeraCrypt's code-signing account, blocking Windows updates. Claude confused who said what in a multi-turn conversation. None of the stories were huge individually. The accumulation wasn't small. Also Thursday: a blog post about git commands hit 2,216 points — the week's single highest scorer — and nobody stopped to wonder why.

Friday. A Molotov cocktail was thrown at Sam Altman's house. 198 pts, 344 comments. The points are low — HN didn't want to be seen upvoting it. The comments are high — HN couldn't stop talking about it. Altman responded publicly. 179 pts, 344 comments. People read it closely, the way you read a statement from someone you're still deciding about. The same Friday: FBI pulled deleted Signal messages from iPhone notification metadata. The encryption worked. The OS layer around it didn't. France announced it was replacing Windows with Linux across government systems. "Ask HN: Any interesting niche hobbies?" ended the week at 708 comments — more than most major AI stories. It's not hard to see why people needed to talk about something else.

Saturday–Sunday. Artemis II splashed down safely. 1,230 pts — the kind of collective enthusiasm AI product launches used to get. Filing the corners off a MacBook finished at 1,360 pts, 647 comments. Then the Sunday data: small models independently reproduced the Mythos security vulnerabilities (1,228 pts). Agent benchmarks were shown to be systematically exploitable. Anthropic's cache TTL cut — made on March 6, six weeks earlier — surfaced with 355 comments. "The Closing of the Frontier" hit the front page the same hour tech valuations returned to pre-AI boom levels. The week started asking about trust in people. It ended asking about trust in numbers.

What Builders Did

Building the layer that doesn't depend on anyone staying cooperative.

The Argument

680 comments on 307 points.

That ratio means people who disagree and can't stop engaging. "What comes after Git" is fighting words. The VC backing makes it worse. But underneath the tribalism, the thread is asking something real: whether the coordination tools for code could be fundamentally different. Nobody has a clean answer.

The week's other live argument: "I still prefer MCP over skills" — 424 pts, 347 comments. The protocol wars for agent tool-use aren't settled. That post will age either very well or very badly within a year.

One Conviction

The benchmark numbers are the week's real story.

Sunday closed with two findings arriving together: small models independently reproducing Mythos-level security research (1,228 pts) and systematic documentation of how agent benchmarks are gamed (488 pts). The usual counterargument: "HN is always bearish on AI, always finds the deflating angle." But the same day: a frugal founder's $20/month stack hit 812 pts and 464 comments, Anthropic's silent cache TTL cut from six weeks earlier landed with 355 comments, and "The Closing of the Frontier" made the front page the hour tech valuations returned to pre-AI boom levels.

The week started asking whether AI platforms could be trusted. It ended discovering that the evaluation framework measuring their capabilities was compromised. The git post isn't nostalgia. The MacBook corners aren't nostalgia. They're what you reach for when the benchmarks are lying.