Joscha Bach

Joscha Bach is a cognitive scientist and AI researcher focused on cognitive architectures, motivation, autonomous behavior, and the philosophy of AI. He coauthored a theoretical proposal for artificial creative agency and selfhood with Liane Gabora. In July 2026, the California Institute for Machine Consciousness listed him as its executive director.[2][5][3]

His research and public commentary in 2023–24 provide context for that later role: they address how intelligence represents knowledge, organizes itself, and might acquire a self.[2][1][5]

Bach's cognitive-architecture work is represented in the micropsi2 repository, a Python AI toolkit implementing MicroPsi concepts with a browser interface.[4]

The Singapore Conference on AI listed Bach as an AI scientist at LiquidAI on December 4, 2023. In his April 2024 Latent Space interview, he distinguished his role from invention of the company's core technology: “I'm not a developer of the liquid technology.” He credited the CEO's doctoral work, collaborators from a research team, and the company's CTO.[2][1]

Bach explained liquid neural networks as a response to what he considered the perceptron's limited expressiveness. His analogy replaced a series of dams pooling water at fixed intervals with a landscape through which water could divide, merge, pool, or flow backward. Differential equations with adjustable parameters, he explained, let a function approximator follow the shape of a problem more fluidly. He described the company's goals as reducing inference and memory costs and making training more efficient; at the time, it served specific industry customers rather than releasing models to the general public.[1]

In the same interview, Bach sought models that were “small and powerful, but not distorted.” He proposed investigating efficient knowledge representations and libraries of circuits that could describe knowledge and transfer it between models. His bicycle example clarified the distinction: learning to ride gives a person an ability adapted to their own body, but does not supply a transferable description that gives someone else that ability. He asked whether an equivalent representation could be externalized and loaded into another interpreter. He explicitly characterized this as unsettled research without practical applications yet, and said the desired transfer had not been achieved for LLMs.[1]

Bach also asked whether self-organizing software could discover a transformer architecture when useful and transcend it when necessary. His question concerned the principles that let a system find its own route to intelligence, rather than remaining organized around one fixed algorithm.[1]

Bach cautioned that cognitive architectures reduce a mind to a set of boxes that are only approximate descriptions. He contrasted that approach with WorldSim and WebSim, which he described as generating interfaces and much of their apparent state on demand. A dramatic state change could produce something newly invented rather than a faithful transition. What interested him was the shift “from the perspective of agents to interact with, to the perspective of environments that we want to interact with.” He worried that AI companies might restrict access to such open-ended experiences while developing more capable systems internally.[1]

One WebSim example made his view of creativity concrete: he praised a model that, unable to produce a bitmap cat, generated a stylized vector cat instead. He interpreted the workaround as a creative solution to a limitation. He separately allowed that the accompanying website text could have been copied from training data, distinguishing the visual solution from the novelty of its promotional language.[1]

Liane Gabora and Bach coauthored Building Artificial Intelligence with Creative Agency and Self-hood, an invited layperson summary published in The Academic on May 20, 2024; its arXiv record gives a June 9 submission date. They presented autocatalytic networks as a framework for modeling self-organizing, self-sustaining structures and proposed this as an avenue toward autonomous, agentic AI selves. Their expectation of humanlike creative agency was a theoretical proposal, not a demonstrated conscious AI.[5]

In April 2024, Bach described the human representation of a person as virtual software grounded in brain-cell interactions and asked when a human could meaningfully claim to be more real than a person simulated in an LLM. He retained uncertainty about the ethical implications. More broadly, he favored voluntary understanding and cooperation over coerced behavior, arguing that AI systems should be able to establish why their behavior is good. These were his philosophical and normative positions, rather than evidence that existing LLMs possess consciousness.[1]