Jade O'Neill

Jade O'Neill is the AI lead of Molecular Reality Corporation, an American company that builds instruments for sensing single molecules and particles as they pass through small holes. The company's founder, Kent Kemmish, has called her "our designated AI lead" and "the main developer behind Maxine's Quest", the company's browser game.^[1] In March 2026 she found a recent paper that classified solid-state nanopore signals with a Mamba network, and in April she ran the experiments behind the company's only machine-learning result to date, in which a Conformer classifier scored 85.4 percent against 81.0 percent for a retrained QuipuNet network, on a public dataset of DNA barcodes.^[2][3] The result has not been published.^[4]

On the company's public pages she is the "AI lead". Its page on Reinforcement Learning with Molecular Feedback, the plan to train models on molecular measurements, ends with the line "Our AI lead Jade O'Neill has been running neural network experiments on nanopore data" and a link to her code.^[3] Keltar's Field Guide to Molecular Reality, a question-and-answer page written in the voice of a green felt puppet, says that she builds alongside the founder, "in the lab and in the code".^[5]

Kemmish named O'Neill as the game's main developer in March 2026.^[1] In Maxine's Quest the player steers Maxine, a red clay imp who stands for Maxwell's demon, around an elliptical arena while mushrooms grow out of a ring that displays an electric current. Each monster in its Live Signal mode is triggered by a spike in the current through a nanopore on the company's laboratory bench.^[6] Back in 2022 the company had described the game as "both instrument control software AND the portal for all our early game releases".^[7] Its browser version, built on the Phaser 3 framework, went online on May 7, 2026, and an experimental second version followed on May 19.^[6][8] One of the two scoreboards is labeled "Kent". It counts Maxine's injuries.^[6]

Kemmish has written that every so often "someone who understands the deep technical underpinnings of machine learning models takes a look at the nanopore signal processing literature for me and reports back". When he first asked, in 2018, he was told that "none of the best work at that time was yet making use of frontier capabilities".^[1]

In 2026 he repeated the exercise with O'Neill, and she found a conference paper from Southeast University in Nanjing that applied a Mamba network, a sequence-model design introduced in 2023, to solid-state nanopore events. It reported 97.1 percent accuracy.^[1][9] Kemmish posted the paper to members on March 29 under a name its authors do not use, "Mambapore", and wrote: "If Jade is around, I'll hereby elicit her insights." He traced its data to the barcode recordings of Bell and Keyser (2016),^[1] the set on which the 2018 QuipuNet network had been trained.^[10] Her own experiments on the same recordings began three weeks later.^[2]

Her experiments sit in a public repository, ai-experiments, whose README opens with the sentence "Jade's experiments running neural nets." It describes itself as "Experiments replicating existing nanopore/speech applications of novel AI methods with open-source implementations", and its eight commits were made between April 19 and 21, 2026.^[2]

The work had three steps. First came a modern version of the demonstration notebook published with QuipuNet, run locally on its dataset; according to the README this script "doesn't need much in the way of resources", and it ran on an ordinary processor rather than a graphics card.^[2] Next, set up as a Hugging Face job, was a simplified rerun of the LibriSpeech experiment from the 2020 paper that introduced the Conformer, a speech-recognition network that combines the attention layers of a transformer with convolutions.^[2][11] Only then did she train a Conformer on the barcode events of the QuipuNet dataset: current traces of DNA carriers marked with three-bit barcodes as they passed through a solid-state nanopore, some of them carrying a bound protein.^[2]

Test data came from whole experiments that neither model saw in training. On that split the Conformer reached 85.4 percent accuracy and QuipuNet's own network, retrained under the same conditions, 81.0 percent, so the error rate fell from 19.0 to 14.6 percent.^[2] According to the README, "both models showed a pronounced failure mode on barcode 101", a sign that this class is "especially sensitive to cross-experiment shift rather than the issue being specific to one architecture".^[2] Both scores are below published figures for the same kind of data; QuipuNet's paper reported a precision of 0.946 on its own test set.^[10] One change lasted a day. A commit of April 19 swapped the averaging step before the classifier for attention pooling, and on April 20 a second commit read "Restore the mean pooling classifier because it works better than attention pooling".^[12]

The company describes the model as the first Conformer it knows of on single-molecule sensing traces.^[4] Testing on public data before collecting the company's own is the first step proposed in its essay The Nanopore World Model.^[13]

• Machine learning on nanopore signals
• DNA barcode (nanopore)
• Radical openness (Molecular Reality)