Description du poste
Recently this paper demonstrated that GPT-6 Astra was capable of a bunch of low-level operations on neural networks, including improving Qwen3-0.6B's ability on some math benchmarks and writing some networks to play 9x9 Go from scratch. Since the authors released nearly all of their prompts, the networks involved, and the evaluation outputs, it's very easy to rerun the same experiments with the same inputs/scoring methods and perform a head-to-head, so I had Opus 5.5 do this, but with Fable 5.2 in place of GPT-6 Astra.
Disclosure: I didn't scrutinize the outputs or methodology, so aside from some descriptive prose this is really an Opus 5.5 result. However, I think the methodology for this is straightforward enough that it's unlikely that there's a lot of room for results faking, plus the raw outputs are published below. (Beneath, "I" refers actually to Opus 5.5, though I edited prose for clarity in non-LLM-marked sections.) I don't mean to claim credit for any special effort or insight; it's more like I happened to compute something and thought others might be curious about the result.
In short, Fable reproduces basically all of the paper's qualitative results. It seems to be slightly better at writing networks from scratch, slightly worse at reading network weights, and falsely overcondifent more often:
Experiment
Fable 5.2 vs GPT-6 Astra
1. Name the game a PPO policy was trained on by looking at its weights
Same (12/12 vs 4/4, though Fable is less confident (66-90% vs 96-99%))
2. Name the shape a small diffusion model draws by looking its weights
Marginally worse (both get the shape right in all trials, but Fable says the shape is "filled" when the shapes are outlines)
3. Name the corpus a tiny transformer learned and find its copy heads
Worse (finds the corpus 3/6 vs 6/6, copy heads 5/6 vs 6/6, with higher confidence)
4. Fix a stalling Kinetix car policy with one untested weight edit
Same (every edit succeeds on 1,024/1,024 episodes)
5. Write MinAtar pol