Our goal
Raise human intelligence to superintelligence with artificial neural networks.
We want people to remember, learn and think far beyond today's limits, by linking the brain to artificial neural networks. No one can do that yet.
Every step toward it depends on software that reads the brain, called a decoder. That is what we build. The replay shows ours reading a real recording.
Step one, working today: reading movement from the brain
of the variation in hand velocity recovered on reaches the decoder never saw. 1.0 would be a perfect match.
median gap between the decoded and the real hand position at the end of a reach. The reaches end a median 129 mm from where they start. A standard linear decoder misses by 55 mm.
unseen reaches used for scoring. All come from one monkey and one recording session, with 182 neurons recorded.
The path from today to superintelligence
No one can make a person superintelligent today. What exists is a sequence of steps, each one resting on the step before. The first is proven in people. The last is our goal.
-
Read the brain
Decoders turn brain activity into movement and speech. A man with ALS has used a speech implant at home for nearly two years.
Shown in people -
Run on the implant
The decoder has to fit on the implant's own chip and keep working as brain signals drift from day to day.
Our work now -
Repair memory
A decoder reads the brain's state and times stimulation to help it store a memory. Early studies in patients show modest gains.
Early patient studies -
Extend healthy minds
The same loop, used to add capacity: more working memory, faster learning, a direct link to AI. No one has shown this yet.
Not yet shown -
Superintelligence
Human and artificial neural networks working as one mind, able to do what no person can do today.
The goal
Solid steps are shown in people or under way. Dashed steps have not been shown by anyone. Nobody has raised the intelligence of a healthy person with a brain implant, and we do not claim to have.
Why artificial neural networks
The brain works through the firing of billions of neurons. Artificial neural networks are the strongest tool found so far for reading that activity. In our own test, on the same 182 neurons, a linear model recovered 0.60 of the movement and a neural network recovered 0.91.
The same kind of network could later decide when and how to stimulate the brain, which is what the memory step needs. That part is still ahead of us.
What has to be solved first
- Writing to the brain. Today's electrodes can nudge a brain region. They cannot write a specific memory or idea.
- The brain's code. Decoding works for movement and speech because there is a clear output to learn from. Abstract thought has none yet.
- Safety. Brain surgery cannot be justified in healthy people, so progress has to come through patients first.
What we have shown so far
This is step one. We trained three decoders on the same public recording and scored each on 574 reaches it had never seen. The neurons and the data stayed the same. Only the software changed.
Accuracy of three decoders
Share of the variation in hand velocity recovered (R²)
-
Linear decoder0.60
The standard starting point: a weighted sum of each neuron's recent firing rate.
-
Linear decoder with history0.73
The same kind of model, given the previous 250 ms of spikes.
-
Our neural network0.91
A small recurrent network that reads the spikes in 5 ms steps and keeps a running memory.
An R² of 1.0 means the decoded velocity matches the real one exactly. 0 means no better than always guessing the average.
How we checked the result
- Three training runs from different random starts scored 0.914, 0.917 and 0.914.
- Given the brain activity from a different reach, the network scores about −0.8, far worse than guessing. Its output depends on the neurons it is shown.
- The network only uses spikes from before each moment, which is what a real-time device has to do. We tested this directly.
- Its settings were chosen on reaches set aside from the training data. The 574 test reaches were not used to tune anything.
- Our linear decoder scores 0.60. The benchmark paper for this recording reports 0.62 for the same method.
Eight unseen reaches: real path against decoded path
Show these numbers as a table
| Gap at the end of the reach | Linear decoder | Neural network |
|---|---|---|
| Median of all 574 unseen reaches | 55 mm | 21 mm |
What this does not show
- It is one monkey and one recording session, analysed after the fact. Nothing here ran live.
- Decoders of this kind are already in the published research. Our numbers reproduce known results and are our starting point.
- Nothing here has run on a chip or in a person.
Why we start with the decoder
Every step on the path runs through the decoder. Implants that read the brain already help people. In a study published in 2026, a man with ALS used a speech implant at home for nearly two years and communicated at about 56 words a minute. Decoders are held back by three things.
Accuracy
How much of the intended movement the software recovers from the neurons.
Where we are: 0.91 on one public recording.Size
An implant runs on very little power, so its decoder has to be small. When we were choosing the network's size, a version with 64 units scored 0.90 and one with 256 units scored 0.92 on the reaches used for tuning.
Where we are: an early sign only. Not yet measured on a chip.Stability
Brain signals shift from day to day, so a decoder trained today can be worse next week and has to be recalibrated.
Where we are: not yet tested. This is our next step.What we do next
Three steps, in this order.
-
Get scored by an independent benchmark
NextFALCON is an open benchmark that tests decoders on later recording days with very little recalibration data. Scoring uses hidden test data and includes recordings from a person with an implant. On 11 November 2026 the top scores were 0.74 on the human movement task and 0.78 and 0.60 on the two monkey tasks, with about 20 entries each.
-
Shrink the decoder for a chip
After thatMeasure how accuracy falls as the network's numbers are stored with fewer bits, from 8 down to 2. Then run the decoder on a microcontroller and measure its speed and power. In our search so far we have not found a published study that maps this below 8 bits for this kind of decoder. We will confirm that before calling it new.
-
Move from movement toward memory
LaterUse public recordings from 280 epilepsy patients memorising word lists to predict whether a word will be remembered. In a 2018 study, brain stimulation timed by this kind of prediction raised the odds of recalling a word by 18%.
What we are raising money for
We are looking for a first grant or early investment to carry out the three steps above, the first part of the path. The amount and terms are not set yet.
- Working timefor the benchmark entry and the chip study.
- Hardwaremicrocontroller boards and a power meter for the chip measurements.
- Computing and storagethe human memory recordings alone are over 600 GB.
- Expert timepaid hours from a neuroscientist or clinician to review the work.
What we need besides money
- A scientific adviser from a brain-computer interface lab or clinic.
- Conversations with implant makers about what they need from a decoder.
To talk about funding or advising, reply to the person who sent you this page.
Who is behind this
threebitschip is led by its founder, a software engineer. The research is carried out with AI. Claude, an AI model made by Anthropic, writes and runs the analysis under the founder's direction.
The human team comes next, starting with a neuroscience adviser, which is part of what we are asking for above.
Sources
- Recording: Churchland, M. and Kaufman, M. (2022). MC_Maze: macaque primary motor and dorsal premotor cortex spiking activity during delayed reaching. DANDI Archive, dataset 000128. Licence CC BY 4.0. We did not collect these recordings.
- Benchmark paper for this recording: Pei and others (2021). Neural Latents Benchmark '21.
- Earlier comparison of decoders: Glaser and others (2020). Machine learning for neural decoding, eNeuro.
- Speech implant used at home: Long-term independent use of an intracortical brain-computer interface for speech and cursor control, Nature Medicine (2026).
- Independent benchmark: FALCON and its public leaderboard.
- Stimulation timed by a memory decoder: Ezzyat and others (2018). Closed-loop stimulation of temporal cortex rescues functional networks and improves memory, Nature Communications.
- Human memory recordings: Delayed free recall of word lists, OpenNeuro.