Brainsless Research Lab · Planless
Mohammad Alsufi is the co-founder and CTO of Brainsless Research Lab and Planless. Before that, he built NoNerds, an AI learning platform he exited in 2026, the technology now runs in front of 2.1 million students.
Brainsless Research Lab is an independent AI research group studying how AI can work reliably beside people over time, publishing its findings openly, and deploying them through Planless.
Ground News
Cairo Scene
Deal Room
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.
It starts with memory. A consciousness that persists needs memory without limits: stored cheaply, read instantly, found reliably, and rules that hold over months. The lab’s four reports take that problem apart. Every number traces to a public artifact, and failed hypotheses are printed beside the wins.
The lab publishes its work openly as four technical reports, and the reports double as the engineering diary of its production systems: Attention Has A Type (BRL-2026-06), on the Constant-Support Law; Attention Pays Its Bill (BRL-2026-07), on the measured cost of long-context memory; Attention Finds Its Keys (BRL-2026-08), on the findability frontier past half a million tokens; and A Measured Decomposition of the Trillion-Parameter Serving Step (BRL-2026-11), on the trillion-parameter serving record: 511.6 tokens per second single-stream on four B200 GPUs, lossless.
All four reports, with full methodology and DOIs, are on brainsless.com.
Neptyn is the lab’s production model, built on open foundations and made the lab’s own through its training. Its memory comes from Cortex; its serving speed comes from Fovea, the lab’s drafter program. It serves live requests inside Planless, which means the lab’s papers are tested where research is hardest to fake: in production, beside real users, across the long sessions the research is about.
Alsufi incorporated his first company at sixteen, in Dubai, while still in school. At eighteen he rebuilt NoNerds, a Jordanian student platform founded in 2022, into an AI-native learning platform, and he exited in 2026; within a month the technology was operating across Jordan and Iraq, in front of a base of 2.1 million students.
Today he builds Planless with his co-founder and co-author Connor Boone. Founder profile at Planless →
“AI is becoming a cognitive layer for humanity rather than a tool: infrastructure that reshapes how people think, learn, build, and decide.”
He intends to spend his career on the problems that outlast any one person: curing disease, ending hunger, opening the road off this planet. The lab funds the journey by shipping real products.
Four technical reports through Brainsless Research Lab: “Attention Has A Type” (BRL-2026-06), on the Constant-Support Law; “Attention Pays Its Bill” (BRL-2026-07), on the cost structure of long-context memory; “Attention Finds Its Keys” (BRL-2026-08), on the findability frontier past half a million tokens; and “A Measured Decomposition of the Trillion-Parameter Serving Step” (BRL-2026-11, July 2026, DOI 10.5281/zenodo.21217452), on the trillion-parameter serving record.
A failure mode documented in the lab’s first report: in long sessions, a model can still recall its rules but stops enforcing them, because all tokens compete for one shared attention budget and commitments dilute as context grows. It is distinct from forgetting; the information is present, the routing fails.
BRL-2026-11 measures the decoding step of Kimi-K2.6, a trillion-parameter open model, term by term, and serves it at 511.6 tokens per second single-stream on four B200 GPUs, lossless — the fastest measured GPU serving of the model. The record was set at 505.9 and raised to 511.6 by blind re-runs of the public release; anyone can re-run it for about $15.
A finding from the lab’s second report: a frozen transformer’s exact next-token prediction depends on a small, constant number of attention keys regardless of context length, shrinking with scale toward a floor near sixteen. Long context becomes a finding problem rather than a reading problem.
Neptyn is Brainsless Research Lab’s production model: a large AI system built on open foundations and developed through the lab’s own research. It powers Planless in production.
Brainsless Research Lab is an independent AI research group co-founded by Mohammad Alsufi and Connor Boone. It studies how AI can work reliably beside people over time, publishes its findings openly, and deploys them through Planless.
All four reports target the same core problem: can AI be trusted over long sessions, at reasonable cost, across huge contexts? The first proves that correct answers require finding a small set of the right information. The second shows a million-token memory can run from $1,500 of RAM. The third maps where finding breaks down past half a million tokens and shows the fix is trainable. The fourth decomposes the serving step of a trillion-parameter model and serves it at 511.6 tokens per second single-stream. The lab’s products inherit all four results.
That AI is becoming a cognitive layer for humanity rather than a tool: infrastructure that reshapes how people think, learn, build, and decide. His research aims to make that layer trustworthy enough to think with.
He co-founded Brainsless Research Lab and Planless with Connor Boone, who co-authors the lab’s research.