Pathway

Founded by Zuzanna Stamirowska in 2021

Run post-transformer artificial intelligence models that combine memory and reasoning in one system

Pricing model

Demo

Rating

★ People love it

Time to value

Immediate

You can use Pathway to build and run next-generation machine learning models that integrate parametric memory directly with logical inference. Instead of bolting external retrieval onto separate models, the system uses recurrent latent reasoning to update connections as it processes new context. It lets you tackle complex constraint problems like Sudoku without chain-of-thought steps, retain critical details across context windows up to 128K tokens, and reduce inference compute costs on demanding reasoning benchmarks.

What Pathway does

Run benchmark evaluations on abstract reasoning datasetsProcess multi-step queries across extended context windowsExecute latent constraint calculations without step-by-step decodingProject input vectors through sparse neuron networksUpdate network memory weights during continuous inferenceBenchmark accuracy-to-cost metrics against conventional sequence modelsUnifies long-term memory and reasoning inside a single model architecturePerforms latent constraint reasoning without needing chain-of-thought tokensMaintains state and tracks facts across context lengths up to 128K tokensUses sparse neuron interactions to decrease inference computing expensesContinuously adjusts connection weights through synaptic memory mechanismsAdapts dynamically to new inputs at test time via local neuron firings