New AI Model Reduces Costs
Researchers develop a new AI model that reduces costs by up to 11 times, using a post-transformer architecture.
Introduction To Post-Transformer Architecture
A new artificial intelligence (AI) model has been developed by researchers at Pathway, which uses a novel approach to AI cognition. This model, known as BDH-CQ, has shown promising results in reducing the cost of requests, with the potential to be up to 11 times less expensive to run than a leading OpenAI model. The researchers published their findings in a paper on the preprint server arXiv, detailing the technical foundations of the BDH-CQ model.
The BDH-CQ model is based on a "post-transformer" architecture, which differs from the traditional transformer-based architecture used in most mainstream AI models. The post-transformer architecture allows for more efficient processing of information, resulting in lower costs and faster training times. The model was trained on just 150 million parameters, significantly fewer than the tens of billions to hundreds of billions of parameters used in more advanced AI models.
Nonverbal Reasoning And The ARC-AGI Benchmark
The researchers evaluated the performance of the BDH-CQ model using the ARC-AGI benchmark, a 2019 benchmark that measures the cognitive ability of AI systems using nonverbal reasoning puzzles. The puzzles involve rotating a series of shapes to complete a sequence, requiring the AI system to infer the rules of the puzzle through trial and error. The BDH-CQ model scored almost 30% on the ARC-AGI-1 benchmark, successfully solving the equivalent of three out of 10 puzzles in two or fewer attempts.
While the score achieved by the BDH-CQ model is not the highest, its underlying reasoning approach makes it significantly more cost-effective than other models. For example, OpenAI's GPT 5.6 Luna (Low) model achieved a slightly higher score, but at a cost roughly 11 times that of the BDH-CQ model in terms of relative token costs.
Transformer Models And Their Limitations
Most mainstream AI models, including those powering Claude and ChatGPT, are based on transformer models. These models transform user inputs into interconnected mathematical reference points, allowing them to infer context from position. However, transformer models have been criticized for being expensive to run, with high costs associated with training and deploying them.
The transformer model forms its responses to user queries by looking at the full prompt simultaneously and then predicting what the next word in the sequence of its reply should be. This process is sequential, meaning that the model has to work through each stage of the response before generating the next word. The post-transformer architecture used in the BDH-CQ model, on the other hand, allows for more efficient processing of information, reducing the costs associated with training and deployment.
Implications And Future Outlook
The development of the BDH-CQ model and its post-transformer architecture has significant implications for the future of AI. The model's ability to reduce costs while maintaining performance could lead to widespread adoption, particularly in applications where cost is a major factor. The researchers believe that this type of AI model architecture could have a dramatic impact on the overall cost and scale of AI deployments.
The results also suggest that the model's cognition capabilities could scale significantly when expanded to larger parameter sizes. This could lead to further improvements in performance, making the BDH-CQ model an attractive option for a range of applications. As the field of AI continues to evolve, the development of more efficient and cost-effective models like the BDH-CQ will be crucial in driving progress towards artificial general intelligence.
Conclusion And Future Directions
In conclusion, the BDH-CQ model and its post-transformer architecture represent a significant step forward in the development of AI. The model's ability to reduce costs while maintaining performance makes it an attractive option for a range of applications. As the field of AI continues to evolve, it will be exciting to see how the BDH-CQ model and other post-transformer architectures are developed and deployed.
The future of AI is likely to be shaped by the development of more efficient and cost-effective models like the BDH-CQ. As researchers continue to push the boundaries of what is possible with AI, we can expect to see significant advances in the field, driving progress towards artificial general intelligence and beyond.
Sources
This is an original synthesis by Qivorane based on reporting from the outlets below.