# metr's interval runs from -55% to +193%, and the footer stays blank

> a metr reanalysis opens the effect from -55% to +193%, libredb prefers an empty field to a false 143, and mythos 5.1's cb risk call rested on three experts.

- edition: Saturday, September 5, 2026 (2026-09-05)
- notebook: dev & ai
- topics: llm · papers · evaluation · brazil
- items: 7 from 7 sources
- original: https://tonho.wtf/en/daily/2026-09-05/
- portuguese edition: https://tonho.wtf/diario/2026-09-05/
- authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf)

---

Four of today's items are the same thing seen from different angles: the number the tool hands back is not the number you think you're reading. The bank client's footer says 143 and the table has 500 rows, because the field measures partitions, not rows. The reanalysis of the METR study says the effect of AI assistance on productivity fits anywhere between -55% and +193%, a range that swallows both the -19% of the original result and the "easy +50%" that became common sense. The Mythos 5.1 system card concludes the model doesn't cross the CB-2 threshold, leaning on evaluations where three experts gave their opinion on chemical weapons. And `.gitignore` has always measured the wrong thing about what goes into the repo, except that now the typical leak is no longer `.DS_Store`: it's `CLAUDE.md`.

What separates the four is what each one does with the weakness of the measurement. LibreDB deleted the field. The METR reanalysis widened the interval until it stopped asserting anything. The system card review asked for a third-party audit. And the marketing for the week's two models keeps saying "the most powerful in the world", from both sides of the table, in the same month.

## research

**[Modelling variation in the METR Uplift Study](https://www.lesswrong.com/posts/XjB4w3zDWLH8xQudq/modelling-variation-in-the-metr-uplift-study)** — METR's July 2025 RCT measured devs 19% slower using AI tools. This reanalysis redoes the math under a Bayesian hierarchical model, treating the single study as one draw from a distribution of possible studies: if between-study variation were what you observe in medical (Cochrane) and economic (Askarov) RCTs, the 95% credible interval runs from -55% to +193%, and that's without even including study-specific effects. The population mean effect lands between 10% and 20% slowdown, with heavy tails, and the probability of a real slowdown lands between 66% and 79% — under a weak prior, the shrinkage is 1 to 2 percentage points and the odds fall to about 2. The point that matters for people who write software isn't the sign, it's the width: the same data is compatible with speeding up a lot and with slowing down a lot, which explains why personal reports diverge so much. And the author notes the inconvenient part: at current adoption, this measurement probably won't be repeated.

## brazil

**[Why we don't show row counts in Cassandra](https://www.tabnews.com.br/libredb/por-que-nao-mostramos-contagem-de-linhas-no-cassandra-500-vira-143)** — LibreDB Studio would open a table with 500 rows and the footer showed 143. It wasn't a bug: the client took the usual shortcut — instead of `count(*)`, which in Cassandra is a scan of every partition on every node and takes down a large cluster, it read `system.size_estimates`. Except the column is called `partitions_count`, not `rows_count`: with clustering columns, 500 orders spread across 143 partitions come out to exactly 143. And even as a partition estimate it wobbles, because it's computed from SSTable metadata by a periodic task, so data still in the memtable hasn't landed and deleted rows keep counting until compaction purges the tombstone. The call was to leave the field empty: with a blank field the user gets suspicious and goes digging, with a confident-looking 143 they write it into the report. Best Brazilian technical post of the week, and the reasoning holds for any UI that hangs a ready-made number in a place where the number doesn't exist.

## world

**[US and Iran trade retaliatory attacks on ships](https://www.bbc.co.uk/news/articles/cj64rrne643o?at_medium=RSS&at_campaign=rss)** — The US said it had permanently disabled two tankers linked to Iran, one of them near Kharg Island — through which roughly 90% of Iran's exported oil leaves — and destroyed a third in the Gulf of Oman; Iranian state media says the IRGC responded against six ships, three linked to the US and three tankers on an "unauthorized" route through the Strait of Hormuz. The strait has been effectively closed since the war began, in February. It goes in here for the bill that sits underneath every data center: energy. Not to draw a price conclusion today, but to remember where the variable comes from.

## who wrote

**[Claude Mythos 5.1 and Fable 5.1: Capabilities](https://www.lesswrong.com/posts/QHoF3tJvryRtmAmMg/claude-mythos-5-1-and-fable-5-1-capabilities)** — Zvi opens by acknowledging the absurd situation of writing a capabilities review when two models were presented as the most powerful in the world days apart, and his read, with wide error bars, is that the jump from Sol to Astra is bigger than the one from Fable 5 to Fable 5.1. The cache cut we covered on Tuesday shows up here with the source's number: cache reads from $1 to $0.25 per million tokens, which Boris Cherny translates into up to 38% cheaper in a typical Claude Code session, plus the classifiers stepping in 85% less on benign biology requests and ~60% less on cyber per session. On benchmarks the gain is modest and without a clear pattern (ECI 162.0 against Opus 5's 160.7; FrontierSWE v2 0.57 against 0.52), with one large exception in Terminal-Bench-Science, from 24.7% to 52.6%. Two things are worth the click: Fable 5 never got past ~11% of Anthropic's dollar spend on Ramp even while being the best model, and the only consistent negative feedback on 5.1 is token consumption — which is literally the 2026-09-02 item, the cheaper token and the more expensive task, now confirmed from the side of the people using it.

**[Review of the CB risk determination in the Claude Mythos 5.1 System Card](https://www.lesswrong.com/posts/vnJ2PyabE4dquJwft/review-of-the-cb-risk-determination-in-the-claude-mythos-5-1)** — Parv Mahajan and colleagues agree with Anthropic's conclusion that Mythos 5.1 probably doesn't cross the CB-2 threshold, and that's exactly why the criticism lands: the problem is the method, not the verdict. The determination rests mostly on human red-teaming with fewer than ten experts, three of whom assessed uplift on chemical weapons, and no new automated CB-2 evaluation has been introduced since May 28, 2026. There's also a concrete case of possible underelicitation: on the black-box RNA sequence design task, humans get two to three hours and the model gets two hours of tool calls and one million tokens — the review estimates 2x to 10x more resources for the human, comparing $50 for the million tokens with $150 to $250 per hour of an ML-bio researcher. And unlike autonomy risks, no third party got a checkpoint to evaluate CB before launch.

**[.gitignore Everything by Default](https://packagemain.tech/p/gitignore-everything-by-default)** — Flip the logic: `*` on the first line and then a `!` for each thing you actually want under version control (`!.gitignore`, `!*.go`, `!go.mod`, `!go.sum`). The author's argument is that the local-junk list has grown — they point at typescript-go's 207-line `.gitignore` — and that today it includes docs and agent subfolders, with `CLAUDE.md` in the same sentence as `.env` and `node_modules`. 155 points on Hacker News and three sources converging, which says something about how much this bothers people. The cost is real and the post doesn't hide it: with deny by default, a new file simply doesn't show up in `git status`, so you trade an accidental commit for a forgotten file. My take: worth it in a single-language repo with allowlist discipline; in a monorepo with ten kinds of artifact, the exception list becomes the same swamp as before, just with the sign flipped.

**[Using Blender with coding agents on macOS](https://simonwillison.net/2026/Sep/5/blender-coding-agents-macos/)** — Willison installed Blender from the official site and told ChatGPT Codex to render a scene with a prompt pointing at the app's path: no MCP, no plugin. The model writes against Blender's Python API, produces a `.blend` you can edit by hand, renders images and assembles video by stitching frame sequences together with ffmpeg. It's the simplest item of the day and the most reproducible this afternoon: the "integration" was a binary at a known path and an agent with a shell.

## stalled sources

Anthropic Engineering (104 days), Karpathy (128), Interconnects (19), Sebastian Raschka (15) and seventeen more feeds without publishing.