Figure 01 · Comparative Cognition · Illustrative, not measured

The brain is the existence proof.

Hyper is searching for a computer system with the brain's properties — sample-efficient learning, continual update, energy economy, coherent long-horizon agency — without giving up the LLM's scale. Such a system is physically possible: evolution found a point in the design space. Whether that region is reachable by search is a separate argument, made on the architecture page.

Brain and frontier LLM are different instances of intelligence. They dominate opposite axes, and neither is a better version of the other.

All domains simultaneously active - no context-switching cost between code, chemistry, and Cantonese. Knowledgebreadth 7 plus or minus 2 working-memory items vs. hundreds of thousands of tokens held in view at once. Contextwindow Words per second produced. An LLM writes a page while a human drafts a sentence. Speed Coherent goal pursuit across days, weeks, years - not just the length of a context window. Long-termagency Embodied contact, intervention, intuitive physics - causal understanding earned by dropping things. Grounding Humans update online from lived experience; LLM weights are frozen after training. Continuallearning The brain runs on roughly 20 W; training and serving consume orders of magnitude more. Energyefficiency A child learns a new word from one exposure; LLMs need billions of tokens. Sampleefficiency 2 4 6 8 10

Brain — wins on efficiency (data, energy, continual update) and situatedness (grounding, long-term agency).

LLM — wins on scale (knowledge breadth, context window, speed).

A rough silhouette on eight public-facing axes. Scores are ordinal, debatable, and deliberately coarse — the shapes are the point. Explainability, reliability, and controllability are tracked in the research docs but left off this radar because they need more careful definitions than a single visual axis can carry. Hover any axis label for commentary.

Reading the two crescents

Clustered like-for-like, the eight axes fall into two opposing groups.

The brain wins on efficiency and situatedness. Sample, energy, and continual update are efficiency properties — biology does more per unit of training signal and per unit of energy by four orders of magnitude. Grounding and long-term agency are situatedness properties — a grounded world model, coherent goals pursued over years. These are not properties a bigger transformer gets for free.

The LLM wins on scale. Knowledge breadth, context window, speed — the cluster of "hold everything at once and move through it fast". These are properties of the digital substrate as much as of the architecture: any mind running on silicon gets throughput for free, and any mind trained on the whole internet gets breadth.

The two crescents face each other. That opposition is the thesis: these are not better-and-worse versions of one thing. They are complementary organs. The region of design space that picks up some of the brain's efficiency without giving up the LLM's scale — that region is what Hyper is searching in.

What we’re trying to find

Hyper is searching for a shape that sits on top of both species. The minimum bar: match brain where brain wins, match LLM where LLM wins — the envelope of what exists today. If efficiency gains compound, scale axes grow too: a smaller model at matching capability is faster, more polymathic per unit compute, and holds context longer. The dashed outline in each diagram is the target; solid shapes are the measured reality reprised from the radar above. Hover any axis label for commentary.

Match LLM (10) - knowledge breadth comes from broad training data.Knowledgebreadth Match LLM (9) - context window is an architectural property.Contextwindow Match LLM (9) - speed comes from the digital substrate.Speed Match brain (8) - long-term agency comes from stable goal representations.Long-termagency Match brain (9) - grounding comes from embodied contact.Grounding Match brain (9) - continual learning is biology's home ground.Continuallearning Match brain (10) - energy efficiency is biology's peak.Energyefficiency Match brain (9) - sample efficiency is biology's peak.Sampleefficiency 2 4 6 8 10
The envelope. What we are searching for: digital system with the combined strengths of brain and LLM.

Efficiency compounds — a smaller model at matching capability is faster, more polymathic per unit compute, and holds context longer.

Knowledge breadth to 11: data-efficient training deepens mastery per domain - more learning per exposure.Knowledgebreadth Context window to 10: efficient architectures hold more coherent context per unit compute.Contextwindow Speed to 12: a smaller model at matching capability gives an order-of-magnitude throughput gain.Speed Long-term agency to 9: continual learning retains task state - goals persist beyond context windows, though architectural work remains.Long-termagency Grounding holds at 9: grounding requires embodied intervention; efficiency doesn't substitute for a body.Grounding Continual to 11: addressing the architectural residual past biology's forgetting ceiling.Continuallearning Energy to 12: efficient learning and efficient inference beat biology's 20 W by orders of magnitude.Energyefficiency Sample efficiency to 11: compounds on itself - efficient systems learn faster from less data.Sampleefficiency 2 4 6 8 10
Compounding. Sample, energy, and continual-learning gains push past r=10. Scale axes follow: a more efficient learner trains smaller models, which are faster and hold context further.

What this argues, and what it doesn't

What it argues. The efficient-learner region of design space is non-empty — evolution found one — and under-searched by a field whose published architectures cluster tightly around transformer variants. If that region is reachable by any search procedure at all, owning the search procedure is the bet worth making.

What it does not argue. The chart is not a benchmark. The scores are ordinal, the axes are negotiable, and the claim is not "brains are better". The claim is weaker and harder to dismiss: the two shapes are different, and difference is evidence of non-emptiness in design space.

Caveat we owe the reader. Continual learning partly reduces to efficiency — cheap fine-tuning would approximate it. What resists is catastrophic forgetting: an architectural, not economic, problem. The chart groups them because the shape is what matters, but the scientist reader should know the split.

Read further