HIPPOCAMPAL

Video models start to reason.

Infrastructure for video reasoning AI — data engines, training pipelines, inference APIs, evaluation suites, and foundational models. Plus a long-term neural hardware vision.

Thesis

Why video reasoning

Real reasoning is temporal, spatial, and causal. Video keeps the world intact. Text-first systems compress the world into descriptions — that breaks when meaning depends on motion, geometry, or multi-step causality. Video reasoning keeps state across time.

Language reasoning is a sub-task of video reasoning.

If you can reason over world state, text becomes just another surface.

Embodied AI is a sub-task of video reasoning.

Perception, state, prediction, action — that is the loop. Video reasoning is the core.

Platform

One platform. Data to deployment. Generate reasoning data, train models, ship inference, and measure progress with the same primitives.

Data Engines

  • Parameterized generators for reasoning tasks, reproducible by seed.
  • Distributed execution at scale; centralized storage as source of truth.
  • Dataset versioning built into the workflow.

Generate a dataset:

import videomodel as vm
gen = vm.data.Generator(task="mental_rotation", params={"difficulty": "medium"}, seed=7)
ds = gen.run(samples=50000, workers="lambda", output="s3://.../rotation-medium/")
print(ds.manifest_url)

Training Pipelines

  • Evaluation-first training loops for reasoning, not aesthetics.
  • Distributed compute support with restartable runs.
  • Tracked configs and artifacts for repeatability.

Launch training:

import videomodel as vm
run = vm.train.Run(model="hippocampal/base", dataset="s3://.../rotation-medium/", objective="reasoning")
run.launch(accelerator="h100", nodes=8)
print(run.dashboard_url)

Inference APIs

  • Low-latency video reasoning endpoints.
  • Batch and streaming interfaces for product use.
  • Optional traces for auditing and debugging.

Run inference:

import videomodel as vm
client = vm.Client(api_key="YOUR_KEY")
result = client.reason(model="hippocampal/base", video_url="https://.../clip.mp4", task="maze_solve", return_trace=True)
print(result.answer)

Evaluation Suites

  • Standard suites plus custom task sets.
  • Outcome metrics now; process metrics where available.
  • Progress you can measure and compare.

Run evaluation:

import videomodel as vm
report = vm.eval.run(model="hippocampal/base", suite=["chess","sudoku","rotation","ravens"])
report.save("./eval_report.json")
print(report.scorecard())

Models

hippocampal/base — private preview

Reasoning-first training mix: curated and generated tasks. Strengths include state tracking, constraint satisfaction, planning, and transformation.

Glasses

Neural AI Glasses

A long-term terminal for video reasoning: perceive the world, build state, and act — with the human still in the loop.

Video reasoning is the next operating system.

Operating systems translated intent into actions on hardware. Video reasoning translates perception into state, prediction, and decisions — grounded in the world.

A neurotech stack for real-world intelligence

High-bandwidth perception. Reliable intent signals. Feedback that supports learning and action.

  • AI Glasses in development
  • EEG research
  • Mandible EMG research
  • Neural Wristband in development
  • Vagus Nerve Stimulation research
  • tDCS research
  • Nasal Peltiers research

Research

We build with evidence-first rigor and publish what we can. Core thesis: video reasoning is a new infrastructure layer. Demos and benchmarks will be shared as they are validated. Private preview updates are available via access request.

Get access

Request private preview access to the platform.

Get access