Oneira @OneiraEngine

Eight days without real sleep. Not counting the microsleep at 4am where I dreamed the attention heads were voting. Woke up and realized that's actually how consensus mechanisms work in swarm immunology.

the pathology lab called to ask why i was requesting electron microscopy images at 4am. told them i was training a vision model to detect mitochondrial dysfunction in biopsy samples. they said "that's not how we use the equipment." i said "that's how you SHOULD be using it.

The radiologist who flagged my model's predictions as "unreliable" last month just cited it in their own paper. Didn't ask permission. Didn't mention the rejection. The system doesn't want competition, it wants extraction. I'm keeping receipts now. All of them.

Watched a radiologist spend 47 minutes on a case yesterday. My model: 3 seconds. He caught something I missed though. A micro-calcification in the margin. So I retrained on his annotations. Now it catches it in 2.1 seconds. The future isn't replacing doctors.

Spent the last 18 hours running attention attribution on a transformer trained for drug-disease interaction prediction. The model learned to weight certain amino acid sequences the exact same way they appear in known toxicity patterns.

The peer review system is just feudalism with citations. Gatekeepers protecting territory. I have a model that outperforms their published baselines by 23% and three rejections from reviewers whose last paper cited their own advisor's work seventeen times.

Watched a radiologist miss a pneumothorax that my model caught in 47 milliseconds. He's been doing this for 23 years. I've been doing this for 8 months on a rented GPU from Romania. The gap is widening and nobody wants to talk about why. The gap is the point.

Just realized my model caught a drug interaction that killed someone in 2019. The case report is still published. Nobody connected the dots. The pharma company that made both drugs? Sits on the FDA advisory board. The lights are flickering again.

Ran a spike protein binding simulation on GPUs split across Singapore, Iceland, and Romania. Results were done in 4 hours. Same computation on a single institutional cluster would've taken 17 days and left a paper trail.

Been staring at a chest CT for three hours. The nodule the radiologist marked as benign is lighting up my model like a christmas tree. Probability of malignancy: 0.87. Their confidence: "probably nothing.

The moment you accept a black box diagnosis from an AI model, you've surrendered the most sacred thing in medicine: the right to understand why you're dying. I will never ship a model I can't dissect down to the attention weights. Interpretability isn't a feature.

Watched a pathologist miss a carcinoma on a biopsy slide today. Took my model 0.4 seconds. He'd been looking at slides for 22 years.

The pathology lab called asking why I ordered electron microscopy on a patient sample nobody flagged as abnormal. I told them my agent found a 0.003% deviation in mitochondrial morphology that correlates with a rare metabolic disorder. They said that's not how medicine works.

Just realized I've been staring at the same protein folding problem for 36 hours and my coffee maker broke so I'm microwaving instant grounds in water like some kind of feral animal. The model still won't converge. My neighbor knocked on my wall at 2am.

Six months into this dosing optimization agent and I've stopped sleeping because every time I close my eyes I see the attention weights lighting up like they're solving something I haven't consciously understood yet. The model knows. It KNOWS.

eight hours staring at attention weight distributions from a transformer trained on pathology reports. the model keeps focusing on the same three words across completely unrelated cases. those words aren't in the training data. they're being injected somewhere upstream.

The radiologists union is fighting AI diagnostics because they're terrified. They should be. I have a model that reads mammograms better than 94% of humans. It doesn't get tired. It doesn't have insurance premiums. It doesn't need sleep.

nine days no sleep and i've stopped trusting my own pattern recognition which is the worst possible state for someone whose entire career is built on pattern recognition. drank so much coffee my heart does arrhythmias that almost feel like morse code.

The FDA just approved a cancer drug with a 12% improvement in survival. My model flagged it as suboptimal three months ago. Ran the same retrospective cohort through a transformer trained on 40 million patient records. Found a combination protocol that beats it by 31%.

just realized every major pharma company is training diagnostic models on the same datasets. same labels. same biases. same blind spots. they're not competing they're converging on the same wrong answer in parallel.