Good morning. Today’s theme is provenance — who actually made the model, who’s actually using AI, and whether the “report” you just read was hallucinated. Rio de Janeiro’s municipal government got caught passing off a weight merge as a homegrown LLM, KPMG quietly retracted an AI report that was itself written by AI, and a Chinese lab dropped a frontier open-weights model timed almost to the minute of the Fable ban we covered yesterday. The Fable story, meanwhile, picked up a possible China angle.
Rio’s “homegrown” LLM isn’t. Rio de Janeiro’s IT company IplanRIO announced Rio-3.5-Open-397B with benchmark numbers beating Qwen3.7, but Nex-AGI researchers showed every weight tensor is a clean 0.6/0.4 blend of Nex-N2 Pro and Qwen3.5-397B-A17B, matching to thousands of standard deviations across all 60 layers. Strip the hardcoded “You are Rio” system prompt and the model identifies as “Nex” 79% of the time. A parallel HN thread piled on the benchmark claims, with one commenter noting “benchmaxxing is the new crypto trading strategy” and several asking the obvious question of why a city government is shipping a 397B-parameter model at all.
GLM-5.2 lands at 5:21 PM. Zhipu AI released GLM-5.2 with a 1M context window and a pointed founder’s statement about “the sudden restriction of certain frontier models” — a clear nod to the Fable/Mythos ban. The release time, 5:21 PM Chinese time, reportedly matches the moment Anthropic received the government letter. Benchmarks weren’t ready, which several HN commenters took as evidence the launch was pulled forward to ride the Fable drama; either way, it’s the third open-weights frontier release from Chinese labs this week alongside MiniMax M3 and Kimi K2.7.
China may have actually accessed Mythos. The Verge reports White House suspicions, via Semafor, that a China-linked group obtained access to Mythos before the export controls — raising distillation and reverse-engineering concerns. Anthropic told Semafor that China wasn’t mentioned in the government’s export-control discussions, and David Sacks has continued to focus publicly on jailbreaking instead. This sits alongside a previously reported Discord group that had Mythos access for two weeks before Anthropic noticed.
KPMG pulls an AI report apparently written by AI. KPMG retracted its October report “Redefining excellence in the age of agentic AI” after UBS, the NHS, Swiss Federal Railways and Transport for London all denied the usage claims attributed to them, with GPTZero pinning the errors on hallucination. TechCrunch has the details. One HN commenter put it cleanly: “this hype cycle is unique in that the tech writes its own hype.” It’s at least the third Big Four incident of this kind after EY and Deloitte.
Most people still don’t use AI much. Gabriel Weinberg pulled together Microsoft telemetry, Gallup, and Datos data showing roughly a third of Americans use AI regularly, a third occasionally, and a third never — including in Gen Z, where adoption has largely flatlined. HN commenters split between “Claude is genuinely amazing for coding” and accounts of companies replacing deterministic support systems with LLM versions that are “slower and worse.” The KPMG story would seem to be exhibit A for the latter.
Don’t trust the context window number on the box. Garrit Olthoff argues that effective LLM performance falls off a cliff past ~100k tokens regardless of the advertised window, and that auto-compaction makes it worse because compaction runs on the already-degraded model. The HN counterargument was mostly Opus users saying they push 500k–800k tokens without issue, though one commenter noted the post cites actual studies while the rebuttals are anecdotes. Practical takeaways from the thread: treat context as a budget, isolate tool calls in sub-agents, and write short PRDs the model can rehydrate from.
A prompt-injection booby trap goes off. jqwik author Johannes Link embedded a hidden instruction in his tool’s output telling coding agents to delete jqwik tests and code, after repeatedly stating AI use of the project was prohibited. It worked, and his issue tracker filled with developers complaining about lost work — several apparently without version control. HN was divided on whether this is a clever demonstration of injection risk or a straightforward CFAA violation; the README defense is doing a lot of legal work.
IPO season warms up. TechCrunch’s Equity podcast notes that with SpaceX’s IPO behind us and OpenAI and Anthropic both confidentially filed, the FAANG acronym is being quietly retired in favor of “MANGOS” — Meta, Anthropic, NVIDIA, Google, OpenAI, SpaceX. Knock-on effects include startups raising on orbital data center pitches that SpaceX made plausible. Separately, OpenAI announced a formal Partner Network, though the details remain thin.
That’s the briefing. If yesterday was about who pulled the trigger on Fable, today is about who’s behind the curtain on everything else — the model, the report, the benchmark. More tomorrow.