MrCogito — Concept Reasoning Model15 February 2026·Updated: 26 April 2026·2285 words·11 minsOpen research on concept-bottleneck transformers. ~21M parameters, deliberate. Phase 2 runs through an agentic auto-research loop.
Cogito Capability Datasets: Long-Context Tests for Small Models10 October 2026·2338 words·11 minsFive open Hugging Face datasets that test whether small (5M–100M) models actually use facts from far back in their context: recall, binding, structure, multi-hop, counting and deduction, on a length ladder from 1k to 128k tokens.
A Memory That Reads by Content: Trained on Short Books, Still Finding Facts at 128K Tokens27 September 2026·2199 words·11 mins
Replacing Gemma-3 Global Attention with Concepts (Toward Longer Context)26 July 2026·2408 words·12 mins
Quicker Failures lead to better questions: How AI Helped Me Steer my research forward8 March 2026·2522 words·12 mins
Speech-to-Speech Models in 2026: Three Architectural Bets and What Each Actually Gives You27 February 2026·2684 words·13 mins
🧠 2025 LLM Limitations: Research Review on the Real Boundaries of AI Reasoning20 August 2025·Updated: 29 August 2025·1985 words·10 mins