Banbury Road released a system consisting of 32 models Kardashev-0.7
BIBIBI
AT A GLANCE
Banbury Road released a 32 model system Kardashev-0.7, claiming that it can develop complementary specialties and reduce inference resource requirements.
Article
Banbury Road lets 32 models train together and develop different specialties on their own
Beating AI Brief:AI Laboratory Banbury Road released Kardashev-0.7, consisting of 32 different models. The team trains these models together using reinforcement learning, without predefining what each model is responsible for, allowing them to develop different specialties and complementary capabilities during training. This approach is called RL for Population Scaling（RLPS).
Previously, the RLPS Laboratory experiments reached at most 16 models. In the 8 model experiment, the jointly trained model group scored 81.70%, higher than the 8 independently trained models’71.04%, and also higher than the 8 repeated samples from the same model’s 72.65%. In the 16 model experiment, there were 22 questions that only 1 of the models answered correctly, showing that the different models had indeed learned different capabilities.Kardashev-0.7 expanded the scale to 32 models. Banbury Road says it can achieve frontier-model-level performance at 0.007 to 0.02 times the inference cost and 0.03 times the memory. Currently,API Beta still requires joining the waitlist.
However, in the previous 8 and 16 model experiments, the group scores were obtained retrospectively by selecting the best result from multiple model outputs using the standard answers. In actual use, there are no standard answers, and Banbury Road has not yet publicly disclosed a complete solution for how the system automatically selects the correct answer.
Original link https://m.theblockbeats.info/flash/370480
Key points
01
Banbury Road released Kardashev-0.7, consisting of 32 different models.
02
The team jointly trains these models using reinforcement learning, without predefining what each model is responsible for.
03
Previously, in the 8 model experiment, the jointly trained model group scored 81.70%, higher than the 8 independently trained models’71.04%and the same model’s 8 repeated samples’72.65%。
04
16 In the model experiment, there were 22 questions that only 1 of the models answered correctly.
05
Banbury Road says Kardashev-0.7 can achieve frontier-model-level performance at 0.007 to 0.02 times the inference cost and 0.03 times the memory.
06
Kardashev-0.7 The API Beta still requires joining the waitlist.
07
Previously, the group scores in the 8 and 16 model experiments were obtained retrospectively by selecting the best result from multiple model outputs using the standard answers.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is an approach in which multiple models are trained together and develop different specialties on their own. This may allow the model group to complement one another, but in the currently public experiments, the best answer was selected retrospectively using standard answers; how to automatically select the correct answer in actual use has not yet been fully disclosed.
Why it matters to readers
If the cost, memory, and performance metrics disclosed by the project team can be reproduced, multi-model collaborative training may reduce the resource requirements of high-performance inference.
The key point is not merely an increase in the number of models, but enabling different models to develop distinct roles; however, experimental results do not mean that the system can already automatically select the correct answer in real-world use.
Risks and unknowns
Kardashev-0.7 The performance, inference cost, and memory metrics for the system mainly come from Banbury Road’s statements.
In actual use, there are no standard answers, and a complete solution for how the system automatically selects the correct answer has not yet been disclosed.
A complete technical report, evaluation setup, and reproducible experimental results have not yet appeared in the evidence.
API The launch date and actual inference cost and memory usage of the Beta remain unclear.
Related Developments
Loading event timeline…
Related concepts
Reinforcement learning
This term is not in the glossary yet. Browse related concepts in the glossary.
