In partnership with
Which AI improvement would matter most to you?
WTF are AI weights? Weights are numbers inside an AI model that influence how it responds. They are adjusted during training. In this experiment, another AI changed a small fraction of those numbers directly.
In an MIT experiment, a small AI started answering a maths question, then repeated itself until it ran out of room. The researchers asked a more powerful AI to investigate. After it edited part of the smaller model, its score on one maths test rose from 7.9% to 48.1%.
That is a big jump. The reason behind it tells us more than the number alone.
I am Alex, welcome to ShortCu8 by Innov8.
Lets Dive Deep ๐ฐ
โญToday's Shortcut
The next time you see a big AI improvement claim, ask three questions:
What was tested? Here, the headline number comes from one maths benchmark, MATH-500, with specific test settings.
What actually changed? The edit reduced a repetition problem, so the model could finish more answers it could already work out.
Did it work elsewhere? The edited Qwen model also improved on an instruction-following test. But a different small model that rarely looped did not show a clearly established maths improvement. Research report
How did one AI help another?
MIT researchers gave GPT-6 Astra access to a small model called Qwen3.5-0.8B, along with tools and practice problems. They asked it to improve the model's maths performance without training it again.
Astra identified a pattern associated with the model's repetition loops and directly edited 131,072 weights, about 0.016% of the model's weights. The researchers then evaluated the edited model. Research report
The researchers directed the experiment. This was one AI helping to modify another, not an AI independently building its own successor.
Why did the score jump?
Before the edit, the model often kept generating until it hit the test's output limit instead of finishing. It hit that limit on 93% of answers in the reported MATH-500 evaluation.
After the edit, that figure fell to 33.1% under the same test settings. More answers made it to the finish line, and the maths score rose from 7.9% to 48.1%. Research report
Think of a student who knows how to solve a problem but keeps getting stuck rewriting the first line. Helping them finish can improve their score without teaching them a new method.
The fix helped with more than maths on this model. Its score on the researchers' instruction-following test rose from 21.0% to 62.3%. That's a second test on the same edited model, not evidence that the fix works on every model.
What does this mean for local AI?
A small model is part of the story. Qwen3.5-0.8B is a small, open-weight model. The researchers have made the edited version available to download, so people with a compatible setup can experiment with it.
There is a reason to pay attention. If future work can make small models more reliable, that could make AI running on your own computer more useful. This experiment is an early example, not proof that the method will improve every model.
There is no instant upgrade for everyone. Running a model locally still requires compatible hardware and software. The research does not show that your existing AI app will get this fix or that your subscription will become cheaper.
What can you do with this today?
If you use an AI app: Pick one task you actually care about, such as summarising a document, and check the answer against the original. A maths benchmark alone cannot tell you how well the model will do your job.
If you run models locally: The edited Qwen weights are available for a compatible setup. Try the same questions on the original and edited versions, then compare the answers. Treat it as an experiment, not a drop-in upgrade for every task. Model card
Sources
Final Thought
I want smaller AI models to become more useful, especially for people who want to run them themselves. This experiment gives us one reason to be curious: an AI found a specific problem inside another model and helped remove it. Now the question is how often that kind of repair can work.
The ShortList
๐ ๏ธCool Tools of the Week:
Synthesia: ChatGPT plug-in to make on-brand videos with motion graphics, avatars, etc
Runway: Now available on Cursor via a pluginย
Audio Intelligence: Appleโs AI-powered Apple Watch features available in beta
๐ฉ Innathe Shortcu8 engane undarunnu ๐๏ธ?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha Shortcu8il kanaam bieโฆ๐
If you read till here, you might find this interesting
#AD1
Read the Top 5 Leadership Books in 1 Week
Reading the five most recommended leadership books cover to cover takes most people a year. With Shortform, it takes a week.
Start with:
Never Split the Difference: how a former FBI hostage negotiator gets to yes
The 7 Habits of Highly Effective People: the principles behind results that last
The 48 Laws of Power: how influence actually works inside organizations
Each guide gives you the key ideas, real examples, and how to apply them at work.
#AD2
โRunning shosโ shouldnโt return zero results. Gladly understands what your shoppers mean, typos and all, and finds the right products.