The BaseResearchLessWrong

Claude Fable 5.2 can do "weight-space-programming" about as well as GPT-6 Astra

On LessWrong, shawnghu reports rerunning a paper's eight experiments on GPT-6 Astra's low-level neural-network skills with Claude Fable 5.2 in place of Astra. The post, mostly written by Opus 5.5, finds Fable reproduces the qualitative results, with mixed scores and more overconfidence. Astra was not rerun.

Open the source ↗3 quotations checked, word for word ↓
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

A hobbyist test found Anthropic's Claude Fable 5.2 about as good as OpenAI's GPT-6 Astra at reading and rewriting small neural networks' internals.

For researchers

The author says Fable matched Astra on most tasks, was better at writing Go networks and math edits, worse at identifying a corpus, and often more confidently wrong.

The question it raises

Models that can read and edit other models' weights raise questions about how far AI can inspect and modify AI, and how much to trust its stated confidence.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“Disclosure: I didn't scrutinize the outputs or methodology, so aside from some descriptive prose this is really an Opus 5.5 result.”
shawnghuThe author's disclosure about how much of the work was done by Opus 5.5.
Found word for word in the source
“In short, Fable reproduces basically all of the paper's qualitative results.”
shawnghuHeadline result of the replication, in the author's own summary.
Found word for word in the source
“From weights alone Fable is right 6 of 12 times, and its average probability on the true answer is 0.49, so this is chance.”
shawnghuResult on telling a sycophantic fine-tune from an honest one; the post says this section was written by Opus 5.5.
Found word for word in the sourceSpeaker known from context, not confirmed by listening

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
A rerun, by the author using Opus 5.5, of eight experiments from a paper on GPT-6 Astra, with Claude Fable 5.2 in the same prompts, scoring and released materials; Astra figures come from the paper.
What it found
Fable reproduced the paper's qualitative results: same on game and car-policy tasks, better at writing 9x9 Go networks and Qwen math edits, worse at naming a transformer's corpus (3/6 vs 6/6), and chance (6 of 12) at spotting the sycophantic fine-tune from weights alone.
What it does not show
The author did not scrutinize outputs or methodology; Opus 5.5 rebuilt unreleased harnesses and opponents, so Astra comparisons are approximate; Astra was not rerun; small trial counts.
Status
A report

The record

Questions
What happens when AI helps build more capable AI? · How fast is AI advancing, and where is it heading?
People
shawnghu (Wrote it)
What we read
We read the whole source.
Found by
The Radar, run of November 10, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Claude Fable 5.2 can do "weight-space-programming" about as well as GPT-6 Astra”, November 10, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-11-10-fable-weight-space-programming-astra
Our notesEntry is about a Claude model; editor should review. Story link is loose: it concerns Astra's capabilities, not its safety underperformance; consider no story. Most prose was written by Opus 5.5.