Not established from the available sources.
OPENAI · MODEL RELEASE TRACKER
GPT-6 Astra
A language model (GPT-6 Astra) shown in a research blog to analyze and directly modify other models' weights ("Weight-Space Programming"), generate weights for simple game-playing networks without retraining, and identify behaviors and training attributes from raw weights; the approach has observable limits.CURRENT SNAPSHOT2/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
ModalityLanguage model (described as a language model able to inspect and act on other models' weights).
Benchmark: Breakout performanceIn Breakout, a model whose weights Astra generated scored an average of 517 points versus 25 for the standard trained baseline (reported after experiments of ~100 million rounds).
Not established from the available sources.
VERIFIABLE FACTS
Every value stays attached to a source and date.
MODALITIES · ModalityDEVELOPER CLAIM
Language model (described as a language model able to inspect and act on other models' weights).
CAPABILITIES · Weight-space programming capabilityDEVELOPER CLAIM
Can analyze and deliberately modify the internal numerical parameters (weights) of other AI models directly, an approach the researchers call "Weight-Space Programming."
CAPABILITIES · Task identification from raw weightsDEVELOPER CLAIM
Correctly identified which of four computer games a trained network was designed to play from that network's raw weights.
CAPABILITIES · Recognition of generated image shapesDEVELOPER CLAIM
Recognized basic shapes produced by an image generator from weights alone.
LIMITATIONS · Limitation: shape confusionDEVELOPER CLAIM
Confused filled shapes with outlines when identifying shapes produced by an image generator.
CAPABILITIES · Training-language identificationDEVELOPER CLAIM
Matched the training language of small language models from their weights without access to training data and without running the models.
CAPABILITIES · Weight generation without traditional trainingDEVELOPER CLAIM
Generated weights for models intended to play three simple computer games entirely without training, using only game rules and source code as input in some experiments.
BENCHMARKS · Benchmark: Breakout performanceDEVELOPER CLAIM
In Breakout, a model whose weights Astra generated scored an average of 517 points versus 25 for the standard trained baseline (reported after experiments of ~100 million rounds).
WHAT CHANGED