Everyone is obsessed with making AI bigger. More parameters. More GPUs. More money burning into the atmosphere. But what if the secret isn't getting bigger, but getting smarter about how it actually thinks? A new research paper from an AI company called Pathway just dropped, and it might actually flip the entire script. They call their new model BDH-CQ. It uses a "post-transformer" architecture. And get this: it costs up to 11 times less to run than OpenAI’s entry-level reasoning model, GPT-5.6 Luna.
 |
| Post-Transformer AI: Pathway’s Model Cuts Compute Costs by 11x |
Eleven times. Let that sink in.
We have been stuck in the transformer era for years. ChatGPT, Claude, all the big names. They work by looking at your entire prompt, turning words into mathematical reference points, and guessing the next word in the sequence. Linearly. Step by step. It is basically verbalizing a train of thought out loud in the background. The problem is that it is wildly inefficient. The longer the prompt, the harder the system chokes. Computational complexity spikes quadratically. You double the text, you quadruple the processing power. It is a memory bottleneck waiting to happen. GPUs just clog up re-reading the same conversation history over and over.
Pathway decided to ditch the text logs entirely.
Instead of chaining tokens together, BDH-CQ uses numerical arrays. Vectors. Directional data points mapped directly inside the GPU’s memory. It doesn't "talk" to itself to figure out a problem. It uses what the researchers call a latent reasoning engine. It runs iterative recurrent loops in the background. Think of it like staring at a chessboard, running through scenarios in your head without saying a single word. If it needs more time to think, it just runs more loops. Here is the crazy part. Running 200 loops doesn't exponentially spike the memory or compute cost like generating 200 text tokens does. The resource consumption stays relatively flat. It just thinks longer.
Does it actually work, though?
They tested it on the ARC-AGI-1 benchmark. That is the 2019 gold standard for nonverbal reasoning. You know, rotating shapes, finding patterns, the stuff humans do effortlessly through trial and error but machines usually choke on. BDH-CQ hit almost 30%. It successfully solved the equivalent of three out of ten puzzles in two attempts or fewer. Now, sure, bigger models have scored higher on this test. But they did it by brute-forcing the problem with tens or hundreds of billions of parameters. Pathway’s model did this with a measly 150 million parameters. A fraction of the size. A fraction of the cost.
You might think this is just another startup hype machine. But the results were independently verified. Richard Zhong from NYU replicated it. So did Łukasz Kaiser. You might not know the name, but Kaiser co-authored the seminal 2017 paper "Attention Is All You Need." He literally helped invent the transformer architecture we are all trying to move past. Even he admitted that Pathway proves model architecture, not just raw scale, can drive the next leap in AI reasoning.
Pathway isn't stopping at 150 million parameters, either. They plan to scale the BDH architecture up to 600 billion. They want to tackle the harder ARC-AGI-2 and ARC-AGI-3 benchmarks. Eventually, they will build a full-fledged text-based chatbot on top of this technology. They are eyeing complex reasoning problems in sectors like cybersecurity incident response and industrial operations. Places where you need deep, abstract thinking without burning a small country's worth of electricity.
The AI arms race has been all about throwing hardware and cash at the wall. Maybe it is finally time we rethink how the machine thinks.
 |
| The End of Brute Force: How BDH-CQ is Rewriting AI Reasoning |
A newly published research paper details Pathway’s BDH-CQ model, a post-transformer AI architecture that utilizes vector-based latent reasoning to drastically reduce computational costs. By replacing linear token generation with iterative numerical loops, the model achieves competitive nonverbal reasoning scores at a fraction of the expense of traditional large language models, signaling a potential paradigm shift in artificial general intelligence development.
#ArtificialIntelligence #PostTransformer #MachineLearning #AIResearch #PathwayAI #AGI #TechInnovation #VectorBasedAI #ComputeEfficiency #FutureOfAI