The Startup That Says Everyone Else Is Building the Brain
Molecular Reality's rebuilt front page has a logo, a video of an eye and one sentence. The sentence is about the rest of the AI-for-science business.
On Wednesday the front page of molecularreality.com was replaced. What loads now is the company's wordmark, a short looping video and a single line of type, all lowercase: "everybody's racing to build the brain, we're building the eye." The same page answers at the company's three other addresses, epicquest.bio, demonpore.com and keltar.ai.
The video lasts 12 seconds. An eye with a blue iris opens on a black background, a protein drawn as a green and orange space-filling model rises out of the pupil and turns, and the eye shuts. Then it opens again. The page's alt text describes the scene in six words: "Eye with protein emerging from pupil."
The page has one link. It says "contact", and clicking it reveals a messaging number and an email address. Getting there took a day and a half of commits. On Tuesday the page carried what one commit message called an "ultraminimalist splash with email-code contact form", and a later commit that day taught the form to "reject typo'd email domains with near-miss + MX checks". By Wednesday the form was gone and the single link had taken its place.
Who is building the brain
The sentence assumes the reader knows who "everybody" is. Company materials say. In Molecular Reality's usage, "the brain" means the large models being trained to do science: drug design, materials discovery, biology. When the company draws the comparison as a chart, the column headed THE BRAIN names five companies: Isomorphic Labs, Lila Sciences, Periodic Labs, Bioptimus and CuspAI.
Isomorphic Labs is the London drug-discovery company Alphabet set up in 2021 to build on DeepMind's AlphaFold. Lila Sciences, in Cambridge, Massachusetts, came out of Flagship Pioneering in 2025 with a plan for laboratories run by AI. Periodic Labs is a 2025 startup building AI scientists with automated labs of their own, starting with materials. Bioptimus builds foundation models of biology in Paris. CuspAI, in Cambridge, England, uses machine learning to design new materials.
The other column is headed THE EYE and holds one entry: "utility-scale molecular sensing with programmable solid-state pores." Other lines in the same materials spell out the relationship the company has in mind. One says "The brain-builders will all need eyes." Another: "the brain reads the internet, the eyes read reality".
The argument has a specific source. AlphaFold, the model at the root of the first entry in the brain column, learned protein structure from the Protein Data Bank, an archive of experimentally determined structures that has been filling since 1971, one experiment at a time. The company's materials make the point in one line: "AlphaFold 3 learned mostly from the half-century-old Protein Data Bank." In the company's telling, the brain was trained on data that earlier instruments collected, and the next models will need instruments that do not exist yet.
What the eye is today
Molecular Reality's eye is the nanopore, a hole anywhere from a few nanometers to a few hundred micrometers wide that molecules or particles pass through one at a time, each leaving a dip in the ionic current. The long-term plan, which the company calls utility-scale molecular sensing, puts those sensors in sinks, toilets and air vents.
Today there is one. The company's FAQ, which went up this month and is written in the voice of a green felt puppet named Keltar, is direct about the scale: "Today we have one instrument in one lab." The instrument is an EDR4, a commercial patch-clamp amplifier sampling at 200 kHz on a small bench in the US Southwest, "the kind of bench operation where the founder solders and the webcam points at the rig." Its output goes out on the company's live stream. The archive holds eighteen recordings, and the first of them are labeled as air tests.
The founder on AI
None of this makes the company an AI skeptic, and its founder, Kent Kemmish, has said so in writing. In a March post on EQBio-OS, the company's game-shaped members' site, introducing a planning document from 2022, he described "AI sucking all the oxygen out of any room where The Future is the topic of discussion," and went on: "AI is awesome, terrifying, astounding, and useful. And I think it will be a central tool for 'curing all diseases'." His argument is about what the models will have to learn from. An essay he wrote to introduce the company's game answers the objection that superintelligent AI will cure everything by itself: "short of genie-level magic emerging from AI systems, I don't see how they'll cure all diseases without what we're building."
The company plans models of its own. The FAQ describes a "nanopore world model", a public model to be trained on the streamed data, and a training scheme the company calls RLMF, reinforcement learning with molecular feedback. It labels both: "Both are proposals, not products, and I want that said first."
One AI already works for the company in public. Keltar, the FAQ's narrator and the voice on the company's phone line, explains himself in the same document: "I am a puppet with a processor, and I say so happily." The felt body is real. "The voice you hear on the phone line is artificial intelligence speaking my lines."
The sentence
Wednesday's page is the third version of the site's front page since mid-June. For about two weeks that month the root of the site showed the live stream itself; after that it showed a page of tabs. Now the stream lives at its own address, the archive at another, and the front page carries only the argument.
Elsewhere the company puts the argument more bluntly. In the same materials as the chart it reads: "The brains are starving for physical data." On the front page the sentence is lowercase, has a comma where a period might go, and ends with a period. Under it, the eye opens again.