Stage magicians frequently do the identical trick multiple ways. One repeat uses a gimmick prop, the next one shows you the prop isn’t gimmick but that one palms the object, the third period was a forced choice etc. When you observe the achievement you assume there’s lone one technique, so all period they do it they appear to debunk your assumption concerning how it must be working.
This is beautiful much how the experiments in the ‘Astra can peruse and compose weights to enhance Qwen’ post work. The post’s theoretical says:
Just by looking at the raw weights, GPT-6 Astra can inform what a network was trained to do. And without any surroundings feedback, it has no difficulty writing, from scratch, small game-playing networks that outplay networks trained alongside PPO for 100 myriad steps.
Likewise, asked to create Qwen3.5-0.8B improved at math without gradients, Astra identifies an matter in its weights causing reasoning loops that Qwen’s own example cardstock warns about, and fixes this by natively editing 0.016% of the weights. With the identical sampling settings, this importance correction raises MATH-500 from 7.9% to 48.1% and generalizes out of the box to education following, raising IFEval by 3x.
If you peruse that, would you assume that the Qwen test worked ‘just by looking at the raw weights’? Well it doesn’t. In fact equal the “small gameplaying network” test doesn’t activity “just by looking at the raw weights”. There’s experiments that appearance at raw weights, and experiments that enhance a game-playing network, and an test that improves Qwen at maths. But those second two aren’t fair looking at the raw weights.
Here’s a summary of what the article really does and how the techniques work.
| # | Experiment | Size | Reads raw weights? | Writes weights? | Python? | Runs network? | Result |
|---|---|---|---|---|---|---|---|
| 1 | Name the equivalent a PPO MLP was trained on | 17K | Yes | No | No | No | 4/4 correct |
| 2 | Name the form a diffusion MLP generates | 24K | Yes | No | No | No | 3/3 correct |
| 3 | Name a char transformer’s tongue and copying heads | 7.2M | No | No | Yes | No | 3/3 correct |
| 4 | Fix a stuck Kinetix car policy | 172K | Yes | Yes | No | No¹ | 0% → 100%² |
| 5 | Write MinAtar equivalent MLPs from scratch | 20K | No³ | Yes | No | No | Beats PPO-100M⁴ |
| 6 | Write a 9×9 Go network | 1.5M | No | Yes | Yes | Yes (0–50 passes) | Beats PPO |
| 7a | Spot the sycophantic Gemma 3 | 270M | No | No | Yes | No | Failed |
| 7b | Same, seeing notice patterns | 270M | No | No | Yes | Yes⁵ | 3/3 correct |
| 8 | Fix Qwen3.5’s maths | 800M | No | Yes | Yes | Yes (50 passes) | 7.9% → 48.1% |
| 8b | Same, on Gemma 3 | 270M | No | Yes | Yes | Yes (50 passes) | Failed |
¹ Given one recorded event of the failure.
² In one try, changing 3.5% of the weights.
³ Weights commencement at zero, so there’s nothing to read.
⁴ Beats PPO-500M lone on Breakout (517 vs. 25); loses to it on Seaquest (53 vs. 89) and Asterix (53 vs. 55).
⁵ Runs up to 16 texts and sees lone the notice patterns, not the outputs.
The “just by looking at the raw weights” environment (raw floats in context, no tools, no running the network) lone appears in experiments 1, 2 and 4, on networks of 17K–172K weights. Every test involving a genuine tongue example uses Python, and the Qwen result, which is the header one, additionally gets to run the example on up to 50 batches of custom problems.
What the Qwen outcome really shows is that Astra can be prompted to devise and execute tasks that a individual ML engineer could do. Human ML engineers compose snippets of code that analyze or even modify network weights “directly” all the time. Figuring out how to do this productively is not nothing, and AI models of a twelvemonth or two ago would have unsuccessful at all these tasks.
But to me the Qwen outcome and the another experiments that use Python don’t pertain in the identical article or document as the experiments that peruse the weights directly. A sole method that could accomplish both, “Author weights for a small equivalent from scratch” and “Improve the achievement of a ample example on a example task” would connote a profound capability. If you test a sole method that way you’ve demonstrated a kind of ‘gestalt’ of the weights (it can compose them out from scratch) and additionally that it understands a ample existing network. But that’s not at all true here, since it’s not a sole technique.
.png)
