"A Maximalist Hypothesis: ChatGPT and other large language models, respectively, contain and understand all true humane knowledge, which is averaged out of their training data, as conceptual abstractions, concistent causal systems, and models of the world. This wealth of knowledge, however, materializes only when prompting. I'm not interested in the simplest problems a language model fails, but the most complex ones it succeeds or fails."

(Sivusto on rakenteilla, pahoittelut pienistä puutteista siellä täällä)




Janne P. Hukkinen

Robot-AGI Researcher, independent

Working on [random notes]

  • 2024 prospecting PhD studies
    • research question, tentative: how slow explicit system 2 reasoning and problem solving [in LLM space] can be automatized to fast system 1 'scaffoldable' capacities?
    • research motto and orienting big question: how AGI robot could be implemented?
    • for a possible explanatory goal, what causal pattern offers explanation – explanatory pluralism FTW
      • [x] how to reason and solve problems in system 2 using LLM's and a set of ML modules?
      • [x] how to implement cognitive scaffolding?
      • [x] what kind of representations constitute whats-it-like reality?
      • [ ] goal: consciousness. explanation: aligment of biological/human and artificial consciousness
      • [ ] ...
    • [pass] cognitive architectures
    • [x] OODA / Sense Think Act Loops
    • [ ] sensory, motor, and cognitive capacities
    • ...
    • big & interesting questions:
      • what are the first principles of cognition?
      • labor-division and orchestration problems of cognitive architectures: for a particular cognitive tasks/work..
        • what kind of set of cognitive modules are needed: amount, individual roles, and capacities?
        • how the processing of modules should be organized in order to fulfil a set of desired high-level goals/tasks/work/end states?
    • interesting questions:
      • grounding mechanisms in science and philosophy
  • 2023 AGI (artificial general intelligence)
    • systemic underpinnigns, constraints, assumptions, and cognitive design tools
    • how much world knowledge can be pumped out of large language models?
    • how to orchestrate a modular cognitive agent / system?
    • world knowledge & understanding: representations in latent embedding space, how far we can get with tera byte language data, and respective Transformer/GPT-3 language models? When is embodied, enacted, and situated grounding needed, if ever?
    • do we need cognitive theory any more? how can it inform us?
      • theory, cognitive architecture, design & evaluation framework
      • what is known and unknown?
      • curriculum learning / critical developmental periods

Hands on



  • MA (cognitive science. minors: computer science)
  • BA


Twitter @Hukkinen. ActivityPub @HuK@mas.to. Youtube. Linkedin



  • email: Janne (a) inrobotico.com