Jay from the Monastery of Panfi
QbitAI is public.
Astra modified the small model weight? MIT's explosion study was publicly questioned by the Dalai Lama.
Two days ago, Astra had just used a remarkable brush-screen AI circuit to directly modify the weight of another AI model without training or even gradient.
The MIT team claimed that Astra modified only 0.016% of the weight of Qwen3.5-0.8B, bringing the mathematical score from 7.9% to 48.1%, an increase of more than 6 times from the original.
Today, things seem to have reversed.
One of the leading figures in the circle posted a lengthy blog post, directly blasting:
There is a serious misunderstanding about this study.
The questioner is Matthew Honnibal, author of SpaCy and co-founder of AI company Explosion.
The Magician's Play
What happened behind this?
Honnibal used an interesting analogy to parody what the MIT Institute is doing: the ophthalmology.
Stage magicians often perform the same trick in multiple ways.
The first time, he may use a special prop; the next time, he will show you no special props, but he hid something himself; the third time, he may use some forced choice.
When you watch a performance, you assume there is only one technique, so every time a magician performs it seems to overturn your assumption of how it works.
Honnibal believes that the experiments in the Astra study are similar to this kind of magic.
I'm going back to Paper, which actually shows three different capabilities:
Weight readingJust look at the raw weight figures to identify what a small neural network is doing.
Writing Weight: Doesn't rely on environmental feedback, directly generating or modifying parameters for small networks from scratch, allowing it to play mini-games such as scrolling and scrolling.
Repair of large modelsBy modifying the Qwen's partial weight, it significantly improved its mathematical test performance.
But the paper's summary gives people a central impression of what it means to be gay.
Astra can understand and modify another model "only by observing the original weight".
It's hard not to understand, if you just glance at it:
Both the ability to recognize the function of a small network and the mathematical ability to repair a large language model are derived from the same "brain-reading" technique.
But is this true?
Honnibal carefully analyzed the eight groups of experiments in the paper and found that the experimental conditions for achieving these three things differed greatly.
That's what I mean.
Purely pure is the exchange of concepts.
Two technological pathways
What kind of little bushy feet did he show up with?
First, in experiments such as identifying small network tasks, repairing a stuck racing model, Astra did "just look at the raw weight".
This is because the experimental conditions are very demanding:
Directly input the weighted numbers (raw floats) into the context of the Astra.
Astra is not allowed to run Python code.
Astra does not allow target networks to observe their behaviour.
In this model, the size of the model parameters that Astra handles are very small.Between 170,000 and 172,000.This is great, but there is still a long way to go to really understand a modern language model.
But the most remarkable and exciting thing about Qwen's mathematical ability to repair experiments is that it doesn't work that way.
When repairing the Qwen model with 800 million parameters, AstraNo direct reading of the original weight。
Instead, Astra is licensed to use Python programming and can perform a maximum of 50 forward-to-back tests. Each time the model is allowed to process a set of exercises to analyze and modify weights by observing the model's performance.
It's a two-way street.
Honnibal pointed out bluntly that packing the two approaches together makes it easy for readers to think wrongly:
Well, the kind of "brain-reading" that's done on small models can also be applied on a large scale to 800 million parameters.
But the truth is that when it comes to dealing with a truly large language model, Astra is more like an "off-the-field coach" who needs to intervene through a series of engineering tools such as programming, running, and observation.
True value
So, does this mean that MIT's research is worthless?
Of course not.
It must be admitted that Astra's ability to read is still remarkable, even when the "read heart" ring is stripped away.
Honnibal also admitted that getting an AI model to design and execute the task of human machine learning engineers to write snippets of code to analyze or even directly modify the network weights would be a huge advance in itself.
The AI model from a year or two ago will fail at all of these tasks.
This shows that today's top models already have the sophisticated tools and reasoning skills to diagnose and repair other models as an "AI engineer".
The real problem is the way the results are presented.
To blur these two distinct abilities into a unified "superpower" that can read and write in the AI brain.
In fact, I thought it was A-Company....
It also reflects the hype frenzy of today's AI circle.
What is the meaning of this word?