Roadmap & support
Where brain v4 goes next, what it costs, and how to help. Research by Khive Labs, Islamabad.
Summary
Our core plan needs about $9,500 of compute over 6 months (~17,300 GPU-hours on L4-class GPUs plus a ~2,000 A100-hour benchmark block) to turn brain v4 from a promising single-machine result into a tested, reproducible learner compared fairly against the best sample-efficient methods.
Separately, about $2,400 over 6 months covers the Claude research agents that design, build and analyse our experiments; this fits AI-credit programs and cloud credits spent through Bedrock or Vertex AI. The plan scales from $1,500 to $100,000+; see What more funding unlocks.
- What brain v4 is: a small (~400 KB) learning library in C/CUDA that learns new kinds of tasks within one life, built by Khive Labs.
- Where it stands: 0.44 ± 0.04 across three independent trainings on our 10 held-out task families (with self-imitation, now the default) against 0.076 for tuned PPO after the same experience (77M steps, 10 seeds); 0.458 with in-life self-imitation, our best confirmed way of living; specialist parts confirmed across three raisings, with a learned router that beats every single part on familiar kinds of task and is level with the best single part on brand-new ones (−0.001); a self-improvement loop with several confirmed rule improvements.
- Why money: results vary by about 0.03 between trainings of the same recipe (measured over 3–4 raisings), larger than many gains we test, so every claim needs at least 3 trainings. More GPUs let us run those trainings side by side instead of one after another.
- What it buys: confirmed improvements at 3x the rate, a public comparison on Atari 100k against BBF, EfficientZero and DreamerV3, and a human sample-efficiency comparison.
- What funders get: confirmed improvements run three trainings deep, a public Atari 100k result for a 400 KB in-life learner, and open code and minds others can build on.
Why this is worth funding
Brain v4 is a different bet from scaling up: a tiny learner that gets better at learning, built and run on almost no compute, with every gain measured honestly.
- A new kind of learner, not another big model. A 400 KB brain that adapts to whole new kinds of tasks within a single life, the way animals do, instead of needing millions of steps per task. Its lives run on a laptop CPU.
- It improves itself. Its learning rule is a small program; our loop proposes edits, tests them on unseen lives and keeps only confirmed winners, which then run as native machine code (10x faster for the rule program, about 1.18x for a whole life). Several improvements have been found and confirmed this way, so far mostly on validation tasks.
- Big results from very little. Everything here, including tuned baselines and a self-improving search, was produced on one laptop GPU, a shared server and about $10–12 of cloud a day. A modest grant multiplies what already works.
- Rigour funders can trust. Our main claims are judged by a rule written before the run and confirmed on fresh seeds, and failures are reported with the wins. When a gain does not survive a second training, we say so.
- A team and a business behind it. Khive is a 50-engineer AI company with revenue from client work; brain v4 is our in-house research, run with production engineering habits (automated cloud jobs, reboot-proof runs, cost caps).
- Clear uses. Small fast learners for game characters (our own god game already runs creatures with it), testing bots, on-device assistants and robots: places where big models are too slow, too costly or cannot learn on the spot.
- Open by default. We will publish code, trained minds and verdict logs, so the field gets a reusable learner and benchmark either way.
Built in Pakistan. Khive is an Islamabad and Rawalpindi AI company of about 50 engineers, and brain v4 is serious research built here on very little compute. Funding keeps this research, and the talent doing it, in Pakistan.
Where it can be used
Brain v4 is the part that learns on the spot: it adapts to a new home, owner, level or body from a few tries, offline, on cheap hardware. Each use pairs it with the sensors and body that use needs.
Now
Game characters and creatures
- What the brain learns on the spot
- to hunt, flee, find food and adapt to the player
- Why a small fast learner fits
- runs hundreds at once on one machine; no scripting
now: running in our own god game
Game-testing bots
- What the brain learns on the spot
- each new level or rule change in a few plays
- Why a small fast learner fits
- finds bugs without hand-written test scripts
now
Near
Robot pets and companion toys
- What the brain learns on the spot
- its owner's habits, games and routines
- Why a small fast learner fits
- cheap chip, offline and private, grows a personality
near: simple sensors and motors
Smart-home helper
- What the brain learns on the spot
- the household's daily patterns (lights, heating, reminders)
- Why a small fast learner fits
- learns each home on-device; no cloud data
near
Personal desktop assistant
- What the brain learns on the spot
- the user's preferences and workflows
- Why a small fast learner fits
- small enough to live on the user's machine
near: alongside a language model
Mid-term
Drones and small robots
- What the brain learns on the spot
- to fly or walk again after damage, wind or a new load
- Why a small fast learner fits
- adapts within one life instead of retraining
mid-term
House-helper robot (tidying, fetching)
- What the brain learns on the spot
- where things go in this home and how this family likes them
- Why a small fast learner fits
- each home is different; learning per home is the bottleneck
mid-term: needs a camera vision front end and a robot body
Cooking assistant robot
- What the brain learns on the spot
- a kitchen's tools and a family's recipes and tastes
- Why a small fast learner fits
- learns from a few demonstrations and corrections
mid-term: needs a camera vision front end and robot hands
Later
Assistive and rehab devices
- What the brain learns on the spot
- how one person moves and what helps them
- Why a small fast learner fits
- adapts to the individual quickly and safely
later: needs medical-grade validation
Vision in our game: our creatures see their world through evolvable eyes that cast real sight rays. On MinAtar Breakout, a small game given to the learner as a grid of symbols, brain v4 reached the score threshold in 60k steps against PPO's 85k. A robot still needs a camera front end and real-world testing.
The common thread: wherever learning has to happen per home, per person or per body, a learner that needs fewer tries is worth more than a bigger model.
A live evolving world. Once the game's new bodies are stable, we will stream our world live on YouTube: creatures with brain v4 evolving bodies, hunting, healing and forming species, with viewers voting on god powers from the chat. Clips of the best moments go to Shorts. The stream links back to labs.khive.io, building an audience that is itself a signal for funders. Running it around the clock needs one cloud GPU machine (about $250–350 a month at current spot prices) plus a small promotion budget.
Where we are
The core learner works and keeps improving; what is missing is confirmation at scale and a fair public benchmark.
| Result | Number | Status |
|---|---|---|
| Head-to-head on 10 held-out task families, same experience (77M steps) | 0.44 ± 0.04 over 3 raisings (0.40 to 0.48, with self-imitation) vs tuned PPO 0.076 (10 seeds) | pass by the rule set before the run; tuned RecurrentPPO running |
| Best way of living (in-life self-imitation) | 0.458 vs 0.441 without it; on the benchmark +0.010, all 3 raisings up | confirmed on fresh blocks (6 of 6) and on the benchmark |
| Specialist parts (place part, 3 raisings) | navigation +0.13, foraging +0.44 to +0.48, water maze +0.20, every raising | confirmed; as a whole mind only 1 of 3 raisings beat the base |
| Learned router over the parts | familiar kinds, new worlds: 0.794 vs 0.776 (+0.017, 3 of 3 fresh blocks); brand-new kinds: 0.458 vs 0.459 | familiar: confirmed by its pre-set rule; brand-new: level, not yet ahead |
| Self-improvement loop (rules edited, tested, kept) | several confirmed improvements, +0.005 to +0.014 validation | running; gains on unseen tasks not yet shown |
| Self-written memory rule (srwER) | +0.007 held-out (2.06 se over 9 blocks) | stacking test running |
Closed or negative (published as such): self-made curriculum, gain/flicker line, both episodic-memory designs, always-on thinking pauses, strong self-imitation in training. Several of these keep local wins and stay as specialist modules for the router.
Roadmap
Four phases over six months, each ending at a gate judged by a rule written in advance; the self-improvement loop runs underneath all of them.
Month 1 · Confirm
- 3 trainings a claim
- new best mind
- router, srwER
- determinism fix
- tuned baselines
Months 2-3 · Router
- plug in D/EVO, planner, memory and SIL specialists
- place curriculum
Month 4 · Benchmark
- Atari 100k, 10 seeds
- vs BBF, EfficientZero, DreamerV3
Months 5-6 · Humans
- tries-to-criterion vs human data
- brain in game creatures; preprint
Month 1 spends most of the GPU time on confirmation, because single trainings vary by about 0.03; later phases build only on what passed.
Compute budget
The ask is $9,500 for 6 months, about 19,300 GPU-hours plus CPU. Prices are approximate Google Cloud spot rates.
| Line | What it runs | Amount | Rate (approx., spot) | Cost |
|---|---|---|---|---|
| Core research GPUs | 4 L4-class GPUs around the clock: 3 trainings per experiment, router and module tests, curriculum, confirmations | 17,280 GPU-h (4 × 24 h × 180 days) | ~$0.30 per GPU-h incl. host | $5,200 |
| CPU lives and self-improvement loop | +64 vCPUs around the clock: bit-exact life evaluations, loop lineages, baselines' evaluation | ~276,000 vCPU-h | ~$0.007 per vCPU-h | $1,900 |
| Benchmark block | Atari 100k: brain v4 plus re-runs of baselines where published numbers can't be reused (BBF, EfficientZero, DreamerV3), 10 seeds | ~2,000 A100-h (~1,000 H100-h equivalent) | ~$1.20 per A100-h | $2,400 |
| Total | ~19,300 GPU-h + 276,000 vCPU-h | ~$9,500 |
Smaller tiers if a program caps lower:
- $3,500: 2 extra GPUs for 6 months plus CPU; doubles our trustworthy-experiment rate; no Atari block.
- $1,500: 2 extra GPUs for 3 months; enough to confirm the current candidates and the router.
Our own machines keep running everything that fits on them first; cloud covers only the overflow.
Research agents (separate line, ~$2,400 for 6 months). Claude runs our research agents: it designs experiments, writes and fixes the C/CUDA code, and analyses results under pre-registered rules. This line fits AI-credit programs (Anthropic research credits) and cloud credits used through Amazon Bedrock or Google Vertex AI, where Claude is available. It is an estimate of about $400 a month and will be re-measured from real usage.
What more funding unlocks
Every level is useful, and progress scales with it: more GPUs mean more trustworthy experiments a day, and each confirmed improvement compounds into the next.
| Funding (6 months) | What it buys | What we can deliver |
|---|---|---|
| $1,500 | 2 extra GPUs for 3 months | confirm the current candidate mind and the router |
| $3,500 | 2 extra GPUs for 6 months + CPU | ~2x today's rate of confirmed improvements |
| $12,000 (core plan) | 4 extra GPUs + CPU + Atari block + research agents | ~3x the rate; tuned-baseline and Atari 100k results; human comparison |
| $25,000 | ~16 GPUs + CPU + more agent time | 30–40 trustworthy experiments a day; full benchmark suite (Atari 100k, MinAtar, Procgen) at 10 seeds; scaling study of larger brains |
| $50,000 | the above + a robot-pet prototype and a smart-home pilot | a physical demo learning a real owner's routines on cheap hardware; public SDK preview; the evolving world streamed live 24/7 |
| $100,000+ | the above + 2 dedicated researchers for a year | a published, reproducible general learner; game, robot and assistant products built on it |
Compute is the bottleneck today: ideas and code are ready faster than 2 GPUs can test them. Each step up converts directly into more confirmed results.
Expected outcomes
Every outcome is judged by a rule written before its run, and losses are published with the wins.
| Outcome | How we measure it | Our estimate | If it fails |
|---|---|---|---|
| A confirmed new best mind (3 trainings, fresh seeds) | held-out families ≥ v4_meta3 + 0.035 (3 raisings a side) | open: two candidates won on one raising and lost on a second | we report the trading-off specialists instead |
| A working module router | one mind beats both fixed settings on held-out and keeps training gains | likely within 2 months (level on brand-new kinds now, not yet ahead) | modules stay as specialists; the chooser is redesigned |
| Fair head-to-head vs tuned PPO/RecurrentPPO | held-out score at matched experience (77M and 300M steps), 10 seeds | PPO done (pass); RecurrentPPO running | we publish the gap and its causes |
| Atari 100k comparison vs BBF, EfficientZero, DreamerV3 | human-normalised score at 100k steps, 10 seeds | a public result for a learner this small; we expect it strongest where few samples matter | published as a baseline for the next version |
| Human sample-efficiency comparison | tries-to-criterion on matched tasks vs published human/animal data | first numbers in month 5 | published with its limits |
| Rate of confirmed improvements | confirmed changes per month | ~3x today's (2 GPUs to 6) | — |
What everything produces: open code, raised minds, verdict logs and a write-up on labs.khive.io; a preprint when the benchmark results are in.
Support this research
Cloud credits, GPU time or a grant at any of the three tiers above ($1,500, $3,500 or the full $9,500) turn directly into confirmed experiments, and everything they produce is published openly. Any level helps: see what each level unlocks, from $1,500 to $100,000+.
To help, or to ask a question, contact us through khive.io.