Adam 15 Quanta [Pong]
Adam 15 learned Pong the way an animal would. Nobody showed it the rules or told it what a paddle was for. It felt distress every time the ball got past it, and over 2,049 seconds it worked out how to stop feeling that. Adam 15 Quanta is the part of that mind small enough to live in your browser: we extrapolated its idea and understanding clusters and wired them to a bionic arm. Play it and see how far it got.
- It was never taught the game
- No labelled examples and no reward function, just a court, a paddle, and a distress signal on every miss. Over 2,049 seconds and 1,050 balls it went from returning half of them to 84 percent. A paddle that never moves returns 19 percent.
- Three segments
- Adam 15 has a retinal cluster that sees, cortical idea clusters that learn and understand, and a motor cluster that moves the paddle. Running all three as one input-output organism is what makes him too heavy for a web page.
- Quanta keeps the middle
- Adam 15’s idea and understanding clusters were extrapolated into 540 spiking cells. The retinal cluster was replaced with calibrated position data from the court, and the motor cluster with a bionic arm, a decoder taught to read what the idea clusters put out and turn it into movement.
- The arm is blind
- It never sees the ball. It reads how fast each half of the motor pool is firing, and nothing else. Silence the idea clusters and the paddle scores 19 percent, the same as a paddle that never moves. Measured, not claimed.
- Traditional Pong physics
- Perfect reflection off the walls, return angle set by where on the paddle it struck, a little more speed each rally. No assists, no magnetism, no rubber banding for either side.
- Accuracy
- Quanta returns 87 percent of the balls that reach it, against 19 percent for a paddle that never moves. On Adam 15’s own court it returns 96, where Adam managed 82. Quanta plays this one game better. Adam 15 can learn the next one.
- Platform
- Web app. 540 cells, 44 KB, about 0.05 ms a frame. Laptop, tablet and phone, and nothing is sent anywhere.
Hindsight
As you walk, Hindsight reads the path ahead in real time, speaks directions toward safe, walkable ground, and gives early warning when a hazard appears. We built it as a downloadable web app for people who are visually impaired. Pathfinder does the seeing: our own vision model, five million parameters we trained to identify walkable pathways and flag their hazards.
- The path, in real time
- It sorts the scene ahead into safe, walkable ground and everything else, as you move.
- Guidance you can hear
- It speaks directions toward safe ground and warns you early when a hazard appears, such as a staircase.
- Adjustable by feel
- The AI speed, the spoken feedback and the visual overlay each take one tap to change.
- Platform
- Downloadable web app
- Model
- Pathfinder, our in-house vision model of five million parameters
Adam V2
Draw a digit and Adam V2 answers first. Then you tell it the truth, and the mismatch between the two, measured inside the network, becomes its learning rate. Underneath, it is V1’s seeing layer with an idea cortex and a voice grown on top. Every boot is a fresh brain that lives in your tab, learns your handwriting in front of you, and dies when you close it. Nothing you draw or teach leaves the page.
- It answers before it hears
- Every lesson has two phases. First the picture alone drives the network’s own answer, then the truth arrives. Nothing outside the network computes the error.
- Teach it live
- Draw a digit, hear its answer, correct it. One lesson is often enough to flip the next answer, and the gate value on screen shows how surprised it was.
- Each boot is its own Adam
- The brain lives in the page’s memory and is never uploaded, cached, or saved. Close the tab and that Adam is gone. Boot a blank one and raise it from nothing.
- The cortex, mapped
- The map shows all four populations live, cluster by cluster: the seeing layer, the idea cortex, the shared mouth-ear, and the fast store, drawn unconnected because it is.
- It asks when unsure
- Below a measured confidence gate it declines rather than guesses, and asks to be taught instead.
- Accuracy
- The shipped brain names 60.5% of unseen clean digits before you teach it, and its confusions are yours to fix. This is a research demo you finish training.
- Platform
- Web app, runs in the browser, nothing sent anywhere
- Model
- DENN v2 (Digitally Emulated Neural Network): a spiking system of four populations and six thousand units with continuous local plasticity, built in-house
Adam V1
Through your camera, Adam V1 reads handwritten digits, and you can watch the whole thing happen, spike by spike, in your browser. Its hundred neurons taught themselves what digits look like, with nobody labelling anything: no answer key, no supervision, just repeated exposure and competition between cells. Nothing you point it at leaves your phone.
100 neurons. No labels.
- Learned without labels
- No one told the hundred cells which digit is which. They learned by competing to respond to what they were shown, and the picture beside this is the result, one tile per neuron.
- Watch it think
- You see the frame it receives, the hundred neurons firing, and how close the pattern comes to each of the ten digits, all on screen at once.
- It says nothing when unsure
- Below a measured confidence threshold it reports nothing rather than naming the nearest digit. It answers about four frames in ten.
- Honest by construction
- Flip the control switch and the identical readout runs on an untrained network, so you can see for yourself how much of the result is the learning and how much is the readout.
- Accuracy
- 60.5% on clean handwritten digits, against 32% for an untrained network of the same size. A camera is harder. This is a research demo, not a scanner.
- Platform
- Web app, installable, runs offline
- Model
- 100 spiking neurons, trained by spike-timing dependent plasticity and built in-house