🧠 Pathway BDH-CQ May Have Learned How to Think Without Talking to Itself
AI MAY HAVE DISCOVERED HOW TO THINK WITHOUT TALKING TO ITSELF.
Most leading generative AI systems today are built on Transformer-based or Transformer-derived architectures.
Transformers power systems such as ChatGPT, Claude and Gemini. They process sequences of tokens and can solve difficult problems by generating long chains of intermediate reasoning.
The harder the problem, the more computation the model may need.
And often, the more tokens it produces.
Researchers at AI company Pathway have introduced a radically different model called Pathway BDH-CQ. Rather than reasoning primarily through written tokens, it performs iterative computation inside a high-dimensional mathematical space, using numerical relationships to represent the underlying structure of a problem.
In simple terms, it attempts to think without first turning every thought into a sentence.
The researchers tested the model on ARC-AGI-1, a benchmark of visual reasoning puzzles designed to measure whether AI can discover unfamiliar rules rather than simply repeat patterns it has seen before.
BDH-CQ achieved 29.5% pass@2 on the public ARC-AGI-1 evaluation set, meaning it was allowed up to two candidate answers for each task.
That result is not close to the highest scores recorded on the benchmark, but the model achieved it with only 150 million parameters.
Pathway estimates that BDH-CQ costs about 11 times less per task than OpenAI’s GPT-5.6 Luna running at low reasoning effort, while Luna scored only modestly higher on the same comparison. The cost figures are not a perfect hardware-to-hardware comparison, but they illustrate the potential efficiency advantage Pathway is trying to demonstrate.
In artificial intelligence, that is the equivalent of entering a bicycle in a Formula One race and discovering it is only slightly slower while using almost no gasoline.
The difference comes from how the model remembers and reasons.
Traditional Transformer systems typically preserve previous context through structures such as a KV cache. As conversations and reasoning traces grow, those stored representations also grow, increasing memory requirements and inference costs.
BDH-CQ takes a different approach.
It stores relevant relationships in an evolving fixed-size recurrent state and reasons through repeated computational loops in latent space.
The model can be given additional recurrent reasoning iterations when more reasoning effort is desired, without producing pages of intermediate language.
It does not think out loud.
It circles the problem mathematically until something fits.
Pathway says its public ARC-AGI-1 benchmark result was reproduced in a separate black-box evaluation. Transformer co-author Łukasz Kaiser also reported reproducing the result.
Pathway now plans to scale BDH-CQ and test it on harder reasoning tasks. Separately, the company says early pretraining experiments with the broader BDH architecture have shown Transformer-like scaling behavior from 1 billion to 600 billion parameters.
That is the unanswered question.
A clever hatchling is impressive.
A 600-billion-parameter dragon is something else entirely.
✨ Why Pathway BDH-CQ Is Important
The AI industry has largely pursued greater intelligence by building larger Transformer-based systems, training them on more data and supplying them with more computing power.
That approach works.
But it is enormously expensive.
Advanced reasoning can consume substantial amounts of:
Computing power
Memory
Specialized chips
Electricity
Cooling capacity
If every improvement requires larger data centers and longer reasoning traces, the economics of intelligence may eventually become as important as the intelligence itself.
Pathway BDH-CQ suggests another possibility: the next breakthrough may come from changing how machines reason rather than simply making today’s architectures larger.
If this kind of latent recurrent reasoning can scale, future AI systems could potentially solve difficult problems with less memory, energy and expense.
Powerful reasoning might eventually move from giant data centers into smaller computers, industrial equipment, robots and everyday devices.
But there is also a deeper issue.
Written reasoning can at least be inspected to some degree. A model operating through abstract numerical loops may reach conclusions through processes that are faster and cheaper but more difficult for humans to interpret.
We may be approaching a trade:
More efficient intelligence in exchange for less visible thought.
BDH-CQ is still a specialized research model focused on visual reasoning. It has not yet demonstrated that it can replace general-purpose systems such as ChatGPT, Claude or Gemini.
The next question is whether the same efficiency survives when the architecture is pushed into broader language, mathematical and real-world reasoning.
🎯 Action
Stop assuming that the Transformer will remain the permanent foundation of artificial intelligence.
Businesses evaluating future AI systems should look beyond benchmark scores and measure the total cost of reaching those scores, including:
Computing power
Energy consumption
Memory requirements
Inference speed
Reliability
Cost per task
Researchers must also develop better ways to monitor latent reasoning systems whose internal processes cannot be translated neatly into words.
Efficiency matters.
So does accountability.
A machine that thinks cheaply but cannot explain itself may save money right up until the moment it makes an expensive mistake.
❓ The Bigger Question
If artificial intelligence no longer needs language to reason, how will humans know what it is thinking before it acts?
Enjoyed this article?
Stay ahead of the curve by subscribing to NewBits Digest, our weekly newsletter featuring curated AI stories, insights, and original content—from foundational concepts to the bleeding edge.
👉 Register or Login at newbits.ai to like, comment, and join the conversation.
Want to explore more?
AI Solutions Directory: Discover AI models, tools & platforms.
AI Ed: Learn through our podcast series, From Bits to Breakthroughs.
AI Hub: Engage across our community and social platforms.
Follow us for daily drops, videos, and updates:
And remember, “It’s all about the bits…especially the new bits.”
Comments