For decades the argument about AI was framed as symbolic against non-symbolic. We think that frame was limited. The real axes are different: whether a solution space must be differentiable to be searched, and whether learning is the passive absorption of data or active interaction with a world. Those two axes organize everything this lab works on, and everything the studio ships.
As we argued, before the LLM eraThe limiting constraint of deep learning was never its rejection of symbols, which emerge in the latent space. It is the requirement that the space of solutions be differentiable. Program spaces are discrete, compositional and hierarchical, and they admit evolutionary search: descent without gradient.
As built, todayAgents that write and execute code are the strongest version of this position. A harness defines a search space over programs, and the quality of that space, its tools, its feedback, its pruning, decides what the agent can find.
For your systems, this is why harness design is engineering, not prompt folklore.
As we argued, before the LLM eraHuman-level competence is not distilled from data alone. Infants act to produce their own data, observe causal links and test hypotheses, learning through structured interaction. The semantic layer is not in the data; it is in the process of acquiring it.
As built, todayThe agentic era settled the question. Context engineering, tool loops and human-in-the-loop protocols: what makes a modern system competent is the structure of its interactions, not just the size of its corpus.
For your systems, this is why we design the acquisition process, not just the model call.
As we argued, before the LLM eraMeaning is the internal structure that forms when an agent categorizes its interactions with the world. Language is a negotiated convention for serializing that structure between agents.
As built, todayMultimodal models, world models and tool-using agents are closing the gap between text and world. Systems that act on real environments need representations that survive contact with them.
For your systems, this is why we build around verifiable state, not just fluent output.
As we argued, before the LLM eraScaling is not enough. Neural networks excel at perception and prediction; algorithmic and formal methods are often more efficient, more general and more explainable for the rest. Serious systems combine them.
As built, todayThe modern agent stack is exactly this hybrid: a model core with planners, memories, verifiers and classical infrastructure around it. A model alone is not a system.
For your systems, this is why we architect the whole assembly, not just the model.
These positions predate the LLM era and were argued publicly, in writing. The program is laid out in the research whitepaper, and the story behind it continues on the team page.
Read the whitepaper (PDF)