Vector-Based AI Could Make AI Reasoning 11x Cheaper
A new vector-based AI model could offer a cheaper way to handle complex machine reasoning. Researchers say its design could reduce computing costs. AI company Pathway developed the new BDH-CQ model using a different approach from mainstream transformer systems. Researchers published their findings on arXiv on August 10.
BDH-CQ contains about 150 million parameters. That is far smaller than many leading AI models, which can contain billions of parameters.
A Different Way to Reason
The model scored almost 30% on ARC-AGI-1, a benchmark designed to test abstract reasoning. Several larger AI systems have achieved higher scores on the same test.
However, researchers highlighted BDH-CQ’s lower operating costs. They said a lightweight OpenAI reasoning model achieved slightly better accuracy but required around 11 times more token costs.
BDH-CQ represents information through numerical vectors rather than long text sequences. It also uses an internal reasoning engine that repeatedly processes information. This method could reduce the need for lengthy written reasoning chains. Traditional transformer models often generate more tokens when they need additional time to solve difficult problems. BDH-CQ can instead perform additional internal reasoning loops. Researchers say this could provide more thinking time without increasing memory requirements in the same way. Pathway describes its approach as a “post-transformer” architecture. The company believes the design could offer a more efficient path for future AI systems.
Reproduced The Model’s Benchmark Results
Independent researchers have reportedly reproduced the model’s benchmark results. Pathway now plans to scale the architecture to as many as 600 billion parameters. The company also plans to test it against harder reasoning benchmarks. It wants to build a full language model using the same technology. If larger versions maintain the same efficiency, the approach could lower AI deployment costs. It could also support demanding uses in cybersecurity and industrial operations. The research highlights a growing AI trend: better architecture may matter as much as bigger models.
