You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$12,000 left to reach the funding goal.
libbrain4 is a learning library in C/CUDA built at Khive Labs in Rawalpindi, Pakistan. A trained brain is about 400 KB. Instead of training a large network for each task, it is raised on many task families and then has to learn new kinds of task within a single life, the way an animal does.
Research began on 3 November 2026, and brain v4 was started on 13 November 2026. Results so far, on our own benchmark:
On 10 task families held out of training, it scores 0.44 ± 0.04 across three independent raisings (0.40 / 0.45 / 0.48). Tuned PPO scores 0.076 after the same experience (77M steps, 10 seeds), so even the weakest raising is over five times PPO.
In-life self-imitation (learning again from its own best stretches during its life) raised every raising, +0.010 on the benchmark, and is now the default.
A specialist "place" part helps consistently across three raisings: navigation +0.13, foraging +0.44 to +0.48, water maze +0.20.
A learned router that picks parts per situation beats every single part on familiar kinds of task in new worlds (0.794 vs 0.776, pre-registered test passed) and is level on brand-new kinds (0.462 vs 0.459).
The learning rule is itself a small program that a self-improvement loop edits and tests. Several edits are confirmed on fresh lives, with gains on validation so far. Confirmed rules compile to native x86-64 code: the rule program runs 10x faster, a whole life 1.18x.
Show the result holds on a public benchmark: Atari 100k against BBF, EfficientZero and DreamerV3, at 10 seeds.
Measure learning speed against people: tries-to-criterion on the same tasks, compared with published human and animal data.
Keep improving the learner, with every change confirmed on three independent raisings before it counts.
A public Atari 100k table with confidence intervals and a tries-to-criterion comparison with human data, both posted with per-seed logs and pre-registered pass/fail rules on https://labs.khive.io, whatever the outcome.
GPU and CPU time. Budget lines: https://labs.khive.io/roadmap
$1,500 minimum: 2 extra GPUs for 3 months, enough for the confirmation runs.
$12,000 goal: the core six-month plan, about $9,500 of compute (4 extra L4-class GPUs around the clock, 64 vCPUs, and an Atari 100k block of about 2,000 A100-hours) plus about $2,400 of AI research-agent time.
Today the work runs on one RTX 3080 laptop GPU, a shared server and a one-GPU cloud quota, paid for by Khive.
Rana Aitsam, CTO and founder of Khive, leads the research. Khive has built custom AI agents, voice AI and software for clients in the UK, US and EU since 2017, with a team of about 50 engineers, and has never raised outside money. libbrain4 also runs the creatures in our own artificial-life game.
The most likely gap is that results on our own benchmark don't carry over to Atari 100k. Then we publish the table as it stands and learn which parts transfer; the library and benchmark stay useful. Gains between versions are close to the spread between raisings, so every claim needs three raisings, and compute decides how fast we can run them.
None from outside. The work is paid for by Khive's own client revenue.
There are no bids on this project.