Newest AI model can read the inside of other AI and write it itself

An AI model normally learns by practising. A neural network is made of a huge number of numbers, known as weights. During training, these numbers are adjusted a tiny bit millions of times, until the network does its task well. This takes a lot of time and computing power. No human can read such a row of numbers and see what the network does, let alone fill in the numbers by hand.

Four researchers at CSAIL, the computer lab of the American university MIT, now show that OpenAI's newest model, GPT-6 Astra, can do this to a fair extent. They gave the model the weights of small networks as text. From these, Astra could work out what a network does, repair mistakes in it and even write new networks, without any training. The researchers published their results on 2 November and say themselves that they are still very early.

What Astra could do

Astra was given only the weights of four small networks, each of which had learned to play a different computer game. It named the right game for all four. It also recognised which shape each network drew in small networks that make pictures of a circle, triangle or square.

Then it went a step further. In a simulation, a network had to steer a car up a slope, and it failed in every attempt. Astra changed 3.5 per cent of the weights. After that, the car reached the top in all 1,024 attempts.

The most striking result: Astra wrote networks entirely by itself for three simple versions of well-known computer games, including the brick-breaking game Breakout. According to the researchers, all three networks played better than networks that had practised in the usual way for 100 million game steps. In Breakout, Astra's network scored 517 points, against 16 points for the practised network.

Repeated by another researcher

Researcher Alex Krentsel repeated the Breakout result himself. According to him, only the very newest models can write out such a working network in a single answer. In his test, the network had 20,486 weights. He calls it a clear jump in what AI can do, but also points out that this is tiny: the largest AI models today have about a thousand billion weights.

A small language model made better

The researchers also had Astra improve an existing small language model: Qwen3.5-0.8B from the Chinese company Alibaba. On arithmetic problems, this model often went round in circles and never reached an answer. Astra changed 131,072 weights, about 0.016 per cent of the whole model. After that, the score on a well-known set of maths questions rose from 7.9 to 48.1 per cent. The model also became better at following instructions: that score went from 21 to 62 per cent.

A large part of that gain came from the model now finishing its answers. Before the change, 93 per cent of the answers got stuck at the maximum length; afterwards it was 33 per cent. Co-author Alex Zhang says the team first thought Astra had put a new arithmetic skill into the model. In fact, the model could already do most of the sums but lost its way.

Caveats

Researcher Lucas Beyer pointed out that the instruction given to Astra contained a hint: if the model repeats itself, finding and fixing the cause counts as part of the job. Co-author Jayden Teoh replied that the team had given a short instruction to the programming assistant Claude Code and that it had added the hint itself. The team also tried a shorter instruction without that hint. In a smaller round of tests, the model reached the same score as with the hint.

With another small language model, Gemma 3 270M from Google, Astra did not manage to improve its arithmetic. According to the researchers, adjusting weights can free up a skill that is already in a model, but cannot add a skill that is missing. They are now working on a fixed test to measure this new ability of AI models, and have made all the log files of their experiments public.

What does this mean for the road to the singularity?

Until now, the weights of a neural network have come from training: a calculation method that adjusts them step by step. This research shows that a language model can also read, repair and write those numbers directly, although the networks are still small. If that also works with larger models, AI can improve other AI in a more targeted way, without expensive training rounds. That is an early step towards AI that works on AI itself, one of the most important ideas behind the singularity. The next step is to show how far this grows as the networks get bigger.

What is the technological singularity? Read the explanation →