The Nanopore World Model

The essay calls the decision to predict in a learned latent space "the single best decision in this whole stack". Its reason is the nature of the data: "A 200 kHz current trace is mostly thermal noise, capacitive ringing, and the amplifier's Bessel filter talking to itself." A model that reconstructed every sample would spend "~95% of its capacity memorizing noise". An interactive trace lets the reader switch the noise on and off and overlay the gold "latent envelope" that a JEPA would keep.^[1]

At the center of the essay stand "The Five Hard Truths", described as "Not reasons to stop".

First comes video: "Cameras run at 30 fps, one frame every 33,000 µs. The amplifier samples every 5 µs." That is a ratio of 6,600. A molecule crosses a pore in microseconds to milliseconds, so a camera can record slow nuisances such as a syringe push, fouling or a bubble, and never the molecule.

The second is identity. "Ionic current is a 1-D shadow of a 3-D molecule," the essay says, beside a drawing of a cylinder that looks round from the top and square from the side. In a demonstration with four classes of molecule, two classes share the same mean depth and duration of blockade. Repeated measurement tightens their clusters but leaves them "welded together"; only a change of conditions, such as the salt or the voltage, moves them apart.^[a] The company's advantage, it concludes, is that it owns the dials: "voltage, salt, cis/trans asymmetry, mechanical sliding-occlusion, dual-pore flossing".

Scale is third. A slider drags the pore on a logarithmic scale from 200 µm, the size of a Coulter counter's aperture, to 2 nm. Some things carry across all scales: the amplifier's response, fouling, drift, wetting and bubbles. Below a few micrometers, surface charge, electroosmotic flow and access resistance begin to matter; below about 60 nanometers the electric double layer and entropic barriers to capture take over, and the slider's verdict is that the parts of the model that recognize molecules "must be RE-TRAINED".

Fourth, labels. Under the plan, the language model would learn from transcripts of the operator's spoken notes, and a model trained to repeat "likely a fibrous occlusion" learns only the operator's guess. The essay asks for a firewall between perception, calibrated physical features and narration, with "a working 'I don't know'".

Fifth comes evaluation: "'Will we get non-gobbledegook back?' is the wrong metric." For clean, isolated pulses, five classical numbers (depth, dwell at half depth, area, asymmetry and the time between events) fed to a gradient-boosted classifier are "the baseline and frequently the winner". A deep model that cannot beat them on clean calibration particles, the essay says, has a bug.^[1]

A boxed warning follows. Video JEPAs come with a slowly updated teacher network that stops the representation from collapsing, and a home-made version for current traces "does not get that for free". Without it, "the predictor collapses every signal to one constant vector and reports a beautiful, meaningless loss of ~0."^[1]

The essay then proposes "MVF Step 0", from the company's phrase minimum viable falsification:^[5] train the model on public nanopore recordings, such as Oxford Nanopore's files of lambda phage DNA, and check whether its frozen embeddings beat the five classical numbers. If they fail, "do not build the rig". It describes the company's alternative, "Building a 200 µm pore and collecting 100 hours of slurry", as "a misallocation of startup capital", and offers the public test "for the price of a pizza".^[1] By then the company's live stream had been running on a 200-micron laser-drilled pore since June, and in July it began the campaign it called July of Streaming, "one hundred hours of molecular streaming across the month of July".^[3][6]

Six changes close the essay. First, the model should act as an experimenter, choosing the next setting of voltage or salt that adds the most information; for free molecules, which diffuse away from a pore within microseconds, the essay's default is to characterize populations, because "the distribution confesses". Single-molecule control is reserved for tethered or trapped molecules, where "your sliding-occlusion patent is the actuator". Physical quantities should be read from the model with small linear probes, including the Coulter estimate of volume, ΔR/R ≈ d_p³/(d_pore²L). The ESP32 in the company's devices "cannot backprop a ViT or Mamba"; it should run a frozen model, and the Molecular Streaming Corps fleet, together with the GPS-tagged World Particle Project, should be the training set. Anomalies should count as discoveries only when many devices see them, since "a lone anomaly is a bubble". Embeddings should be split into an analyte part and an instrument part. A Mamba state-space model should carry the running belief. In the essay's summary, "Conformer = ears", "Mamba/SSM = working memory", and "JEPA = the training objective that keeps it honest".^[1]

A diagram of the revised stack has ten blocks that open when clicked. The language model's block now reads "DEMOTED to a downstream CONSUMER", and the video encoders are "DEMOTED to a NUISANCE encoder". A closing roadmap ends with the fleet: "The model and the movement become one."^[1]