There are moments in science when progress outpaces institutions. Historically, that is when independent research has mattered most. AI is in one of those moments. We are working on the memory, alignment, and attention problems that stand between where the field is and what AGI could do for the world.
A consciousness that persists needs memory without limits: stored cheaply, read instantly, found reliably, and rules that hold over months. Four reports take that problem apart. Every number traces to a public artifact, and failed hypotheses are printed beside the wins.
The safety paper. Over a long conversation, an AI still remembers its rules but quietly stops enforcing them, because every rule competes with every message inside one shared attention budget. We measured this failure across frontier models, then built an attention architecture that holds the rules apart. Alignment scores went from 0.26 to 0.79 with no loss of capability.
Download PDF
The law. To predict its exact next word, a model only ever needs a small, fixed number of memories, whether the conversation is eight thousand tokens or a million. Bigger models need even fewer. This reframes long context entirely: the problem is not reading everything, it is finding the few pieces that matter.
Download PDF
The money paper. We built the engine the law promised and measured it. A long conversation now costs the same per word as a short one, and the memory no longer needs expensive GPU hardware: a million-token cache that would demand roughly $180,000 of GPUs runs from about $1,500 of ordinary RAM. Every claim is priced and re-runnable.
Download PDF
The frontier. Past half a million tokens, models stop finding the right memory, and we discovered the mathematical law that says why: the needle keeps its rank while the bar for survival keeps rising. The collapse is arithmetic, not bad luck. We also showed the fix is trainable, without touching the model itself.
Download PDF
Mohammad Alsufi & Connor Scott, with the Brainsless Research Lab AI Systems Research Group. Code public, noncommercial license.
Built on open-weight foundations · GMSA is our own attention research
Neptyn 1.0 is our production model, taken further with our own training and research. It remembers, it carries a state, and it acts on your behalf. The first piece of the consciousness we are building. It runs inside Planless.
Cortex gives it memory that consolidates like a brain and a state that persists between conversations. Two variants, routed automatically: Instant talks, Thinking deliberates.
Neptyn is the model. Cortex is its memory. GMSA, rolling out now, is how it holds its shape over time.
A three-layer bitemporal memory system (working snapshot, semantic store, episodic stream) that consolidates continuously in the background, modeled on human memory: episodic in the moment, semantic over time, working-set on the surface. It is what gives Neptyn continuity across every conversation, channel, and session.
how Cortex works →Our typed-attention architecture, which gives behavioral commitments their own attention stream so they hold steady across long, multi-constraint sessions instead of diluting as context grows. Documented in BRL-2026-05.
read the paper →