The commercial labs aren't building what they say they are.

AGI is invoked like a destination but kept conveniently vague. Meanwhile, the cognitive partner people actually imagine when they say "AI" requires architecture the labs aren't really pursuing.

I’m actually building the system-level approach this memo hints at.

I’m not from a lab or an academic group, but I’ve been working on a post-transformer inference method where you extract a low-rank “meaning field” from a frozen Llama-70B layer and train a small student model to generate those fields directly.

The idea is similar to what this memo describes, but with an empirical implementation.

I just open-sourced the reference version here:

GitHub: https://github.com/Anima-Core/an1-core Paper + DOI: https://zenodo.org/records/17873275

It isn’t about bigger models. It’s about reorganizing the system around meaning and structure, then treating the transformer as a teacher rather than the final destination.

I’d genuinely appreciate critique or replication attempts from people here. HN tends to give the most honest feedback.