Polish scientists’ startup Pathway announces AI reasoning breakthrough

From the article 

Solving the “generalization over time” problem is among the “holy grails” of the AI world – a goal numerous top scientists around the world have unsuccessfully strived to reach for some time now. The new, groundbreaking AI architecture created by Poland’s Pathway startup seems to have done just that – creating a digital structure similar to the neural network functioning in the brain, and allowing AI to learn and reason like a human.

This article is a nightmare. Here’s the paper: https://arxiv.org/abs/2509.26507

> We confirm empirically that specific, individual synapses strengthen connection whenever BDH hears or reasons about a specific concept while processing language inputs.

ngl, still sounds a lot like any other kind of reinforcement learning. Anyway their breakthrough is to attempt to model neurons in as in the human brain.

I haven’t fully digested it, but it might not be BS.

We’ve been waiting for a while for what state space models (SSM) can do. One of the major problems with Transformer models is that they don’t maintain internal state, but these do maintain a latent state (albeit fixed size). Maintaining state has obvious effects on sequence data. Also, depending on the implementation (for example Mamba, https://arxiv.org/abs/2312.00752) state space models can scale linearly with context size, but Transformers naturally want quadratic (lots of caveats here…). SSMs are also much closer to the brain than Transformers. So there has been a lot of interest in them, but it’s not been clear that they can be better in terms of performance, so people have been looking for hybrids.

This appears to be a new architecture that is a state space model. Honestly, there is a lot of theoretical claims in there, but it’s just not clear to me that it will perform better. It’s going to take a while to unpack and turn the claims into tested demonstrations. On the plus side, there’s an implementation: [GitHub – pathwaycom/bdh: Baby Dragon Hatchling (BDH) – Architecture and Code](https://github.com/pathwaycom/bdh)

There’s also a lot of influence of recent work on Probably approximately correct learning theory, which would be welcome because it’s interesting theoretically but I haven’t seen too much practical come out of it.

Dumb. This is marketing at best. Everyone throwing their piece into this gigantic nvidia powered dumpster fire nightmare.