Unifies memory and reasoning in a single architecture, working at test time
Intelligence
with no
ceiling
Frontier AI lab building architectures and models that autonomously reason, learn, and evolve.
Follow the journeytowards the next era of AI
At sub-billion-parameter scale, BDH demonstrates efficient adaptation, latent constraint reasoning, and persistent state tracking.
97.4%
Solve rate
Sudoku Extreme
Powerful latent constraint reasoning without chain-of-thought1
95%@32K
BABILong
Multi-step reasoning over facts scattered across long contexts2
1Pathway BDH achieves a 97.4% top-1 solve rate across approximately 250,000 Sudoku Extreme puzzles without chain-of-thought or solution backtracking; the leading reasoning LLMs evaluated in the cited comparison scored approximately 0%. See the full results and evaluation setup here.
2Accuracy on BABILong QA1–QA5 using a 134M-parameter BDH model. BDH reaches 95% at 32K tokens and 82% at 128K tokens. BABILong tests retrieval and reasoning over sparse task-relevant facts embedded within increasing amounts of irrelevant natural-language context. BDH results are pending final contamination checks, independent validation, and leaderboard review.
BDH (Dragon Hatchling) is a post-transformer AI architecture that remembers, reasons, and improves itself over time.
Yields models that speak in language but think in abstract thoughts
Latent reasoning
Delivers significantly cheaper intelligence via sparse, local neuron interactions that cut test-time compute
Extreme efficiency
The cycle is continuous, the model is rewriting its own reasoning rules as it thinks.
Encoding
A new input is projected into the network and only the relevant set of neurons activates.
Local firing
Activated neurons send signals to nearby connections, allowing information to propagate through local interactions.
Integration
Neurons accumulate incoming signals and fire when their combined activation crosses a threshold.
Synaptic memory
Reciprocal activity strengthens useful connections, while inactive one-way connections gradually weaken, updating the network’s memory.
Decoding
After a few reasoning iterations, we project the network state to the output space to obtain the result or next input.
Our
Research
Pathway's post-transformer AI architecture BDH is built around a single premise: intelligence should not have to choose between reasoning and memory. Rather than bolting memory to a language model from the outside, BDH makes memory, adaptation, and inference part of the same computational fabric. It draws on principles that biology got right, namely local interaction, sparse activity, persistent state, and continual adjustment, and applies them to a modern sequence model.
The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain
Meet BDH, the flagship of the Post-Transformer Era with native continual learning, efficient skill acquisition, and powerful latent reasoning over long context
Theoretical foundations of AI: the equations of reasoning
A deep dive into how and why BDH works, and why we need a conceptual leap to unlock continual learning and long-horizon reasoning
Prior work by our Teamleading to today's scientific breakthroughs
Regularizing Neural Networks by Penalizing Confident Output Distributions
Authored by Gabriel Pereyra, George Tucker, Jan Chorowski, Łukasz Kaiser, Geoffrey Hinton
Attention-Based Models for Speech Recognition
Authored by Jan Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, Yoshua Bengio
End-to-End Attention-based Large Vocabulary Speech Recognition
Authored by Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Philemon Brakel, Yoshua Bengio
State-of-the-art Speech Recognition With Sequence-to-Sequence Models
Authored by Chung-Cheng Chiu, Tara N. Sainath, Yonghui Wu, Rohit Prabhavalkar, Patrick Nguyen, Zhifeng Chen, Anjuli Kannan, Ron J. Weiss, Kanishka Rao, Ekaterina Gonina, Navdeep Jaitly, Bo Li, Jan Chorowski, Michiel Bacchiani
Our foundersleading the Post-Transformer Era
Zuzanna Stamirowska
CEOAn École Polytechnique graduate with a PhD in Complex Systems and expert in graph-based game theory, she created state-of-the-art network forecasting models published by the National Academy of Sciences. Her groundbreaking work earned her a feature on the cover of Le Point as one of the "100 geniuses whose innovation will change the world."
Jan Chorowski
CTOA former researcher at MILA and Google Brain (under Samy Bengio), he is a prominent AI scientist who co-authored work with Nobel Laureate Geoffrey Hinton and contributed to early attention models. A co-author of the pioneering deep learning library Theano, his impactful research has earned over 12,000 citations and an h-index of 24 on Google Scholar.
Adrian Kosowski
CSOEarning a PhD in algorithms at 20 and tenure at Inria by 23, he is a prolific researcher with over 100 papers and an h-index of 29 on Google Scholar. An early pioneer of navigable small-world search (ACM SPAA Best Paper), his versatile work spans graph algorithms, distributed and agentic computing, quantum information, and bio-inspired systems.
Advisors & Partners
Łukasz Kaiser
Co-inventor of Transformers, the key researcher behind reasoning breakthroughs from OpenAI.
Martin Farach-Colton
Chair, CSE, NYU Tandon. ACM, IEEE, SIAM Fellow.
Jacques Attali
Economist, writer, State Councilor. Founded institutions such as the European Bank for Reconstruction and Development.
Jonathan Frankle
Chief AI Scientist at Databricks. Founder of MosaicML.
Partnered with leading companies
