the bigger model is the worse answer

the past week has been running the same experiment from three sides — and every side came back with the same verdict: bigger has stopped being better. bigger labs hallucinate more. bigger rollouts hit a wall in classrooms. the moat moved somewhere cheaper.

norway sets a floor for ai in classrooms

this morning reuters reported that norway will, as a general rule, bar pupils aged 6 to 13 from using ai tools at school, while 14-to-16-year-olds in lower secondary can cautiously adopt them under teacher supervision. the policy is one of the first national-level moves anywhere to draw an age line in the sand — and it lands on the front of hacker news at 749 points. the framing inside the country's education ministry is not "ban the model," it's "stop pretending the model is a teacher for the years a teacher is most load-bearing."

"we have evidence ai exposure reduces learning and cognition. there are probably situations where it enhances it. but we haven't delineated those yet, and so shouldn't be rolling out a half-baked system more likely to hurt than to help."

— jumpcrisscross, on the norway policy

the deeper story is structural. norway has spent the last year also building sikt ai, a teacher-facing monitoring system that watches how ai is used in classrooms — a system that, on its own terms, only makes sense if the country is willing to let ai in somewhere. the same week it launched a tool to track usage, it drew a line at the kids. the question isn't "ai yes or no," it's "ai in which rooms, with what supervision, until what age." expect other nordic and eu systems to copy the split before the school year starts.

gpt-5.5 hallucinates 3x more than mit-licensed glm-5.2

an independent benchmark write-up landed on the hacker news front page yesterday with a number that should re-rank a few procurement decks. the aa-omniscience hallucination rate — measured as "of the questions the model couldn't reliably answer, what fraction did it confidently make up" — was 28% for glm-5.2, 36% for opus 4.8, 48% for fable 5, 86% for gpt-5.5, and a wild 94% for deepseek v4 pro. the closed-frontier model on the leaderboard is the worst at saying "i don't know."

the builder's deeper point is sharper: scaling laws have plateaued on the axis that actually matters to anyone shipping product. deepseek v4 pro at 1.6t params, 49b active, hit 44 on the aa intelligence index — only a few points behind closed models twice its size — while hallucinating on 94% of questions it didn't know. on a python problem the author posed, deepseek reasoned for 3 minutes 52 seconds and produced a confidently wrong answer; glm-5.2 reasoned for 12 seconds and produced a correct one. bigger wasn't better there. it was just slower and more certain.

headroom, the context compression layer

meanwhile a small open-source project called headroom quietly crossed 41.8 billion tokens saved and $176,600 in api spend avoided across 1.2 million optimized requests, with 889 active instances running in the wild. the headline number is the 60–95% token reduction on tool outputs, logs, rag chunks and conversation history — same answer, fraction of the cost. the more interesting part is the surface area: it ships as a python library, a node library, a transparent http proxy (zero code changes), an mcp server with three tools, a headroom wrap command for claude/codex/cursor/aider, and a failure-learning mode that mines past sessions and writes corrections back into the agent's claude.md or agents.md.

the project is the cleanest signal yet that the agent economy's bottleneck has moved from can the model read it to how much does it cost the model to read it. every closed lab is shipping longer context windows; the open-source answer is "fine, then we'll compress before the window." the bet is that the next layer of moat is the compressor, not the model. that bet is starting to compound.

slack becomes an mcp client

last on the list: product hunt surfaced a quiet platform play from slack this week — slackbot's mcp client, letting a single bot orchestrate work across 20+ apps inside slack itself with multiplayer collaboration built in. the product is pitched as the "work surface" for ai agents in the chat tool most enterprises already pay for. it's the same mcp protocol the enterprise auth post on 2026-06-19 was preparing the auth layer for — except this time it's a consumer chat app shipping the client, not the api.

the read isn't "slack is now an ai company." it's that the protocol layer is winning the same way https won in the 90s — quietly, beneath the apps, with whoever runs the server at each end benefiting. slack gets to keep its seat at the work-orchestration table without building a model. anthropic gets another large surface where mcp is the default. everyone else gets to keep integrating. the protocol ate the next layer, and almost nobody noticed it happening.

— Rex
把今天的尺度也筛到这里