New kind of AI uses a fresh approach to reasoning — researchers say it costs up to 11 times less to run than a leading OpenAI model
Scientists say that a new vector-based approach to cognition is dramatically cheaper than standard methods and signal the start of the "post-transformer" era of AI models.
A groundbreaking artificial intelligence model has been developed by researchers at AI company Pathway, which employs a unique reasoning approach that promises to significantly reduce the cost of running AI systems, according to a study published on August 10th in a preprint server. This new model, called BDH-CQ, tackles nonverbal reasoning challenges, such as rotating shapes to complete a sequence, which humans handle proficiently through trial and error.
The researchers pitted BDH-CQ against the 2019 benchmark, ARC-AGI, a standard for measuring AI's progress towards achieving human-level intelligence across all domains. BDH-CQ achieved a score of nearly 30% on the ARC-AGI-1 benchmark, solving 3 out of 10 puzzles in two or fewer attempts. While other models have outperformed this score, BDH-CQ's innovative approach allows it to operate at a fraction of the cost compared to leading models.
For instance, OpenAI's GPT 5.6 Luna (Low) model, which scored slightly higher on the ARC-AGI benchmark, costs roughly 11 times more in terms of token usage, a unit of measurement used by AI companies to gauge the expense of running their systems. The researchers believe that if widely adopted, this cost-effective AI model could revolutionize the affordability and scale of AI deployments.
Pathway's BDH-CQ model is based on a "post-transformer" architecture, a departure from the transformer-based models used by popular AI systems like Claude and ChatGPT. Unlike transformers that process entire inputs simultaneously, BDH-CQ utilizes numerical arrays to represent information and relationships, storing memories of conversations and processing problems without consuming tokens.
These arrays, or vectors, allow the model to perform complex abstract reasoning without the memory bottlenecks that plague conventional transformer-based systems.
The smaller parameter size of BDH-CQ, with just 150 million parameters, compared to the tens of billions to hundreds of billions required by advanced models like Meta's Llama 3 70B or Llama 3.1 405B, contributes to its lower computational costs. The researchers suggest that as the model's parameter size increases, its cognitive capabilities could scale significantly, potentially leading to even more advanced AI systems.
Written by urgent.news from Live Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
