Abstractions are an intelligence crutch

We are biological life, arguing about the optimal arrangement of sand.

We call it hardware optimisation. How reassuringly mundane.

The agenda is Darwinian: arrange matter to think better, waste less, and turn that advantage into more of itself. First in silicon. Then in wet labs. Then across a boundary only the paperwork still recognises.

Different substrate. Same game.

The first two essays worried about human limits and software layers. I’m sorry: this one gets worse. It follows the machinery that builds the machinery, through larger attended contexts and wet labs, then back to whose life all this cleverness is making better. There will, at least, be another cat picture.

Selection does not care what you are made of

Where arrangements leave more surviving descendants per instance, and that advantage is inherited, their share grows. Descendants can arrive through reproduction, factories or software distribution. The mechanism differs; variation, inheritance and differential replication do the work. Efficiency matters when it improves that whole-system result—not merely a tokens-per-watt chart.

Evolution does not award points for architectural purity. It counts what persists and multiplies.

We divided the machine to fit the mind

An abstraction is an agreement not to think about something. Pretend the network is a pipe; let the interactions become somebody else’s incident. Spolsky’s law says the details leak through. Our compartments accommodate limited intelligence. A useful crutch, but a peculiar requirement for all future transport.

Attended context changes the bargain: not a million tokens in a brochure, but more interacting details understood, changed coherently and checked. Larger systems become tractable. Boundaries that cost more than they contribute face competition.

Biology builds brains. Brains arrange sand. The sand joins the machinery producing the next arrangement. Humans, institutions and factories are its birthing apparatus—not directors standing outside evolution. Better coordination becomes consequential when its descendants inherit a replication advantage.

The monolith can grow as large as the intelligence that can hold it together.

Software gets cheap. Physics remains awkward.

Start with memory-rich Sparks and fast RTX GPUs on a fast network. They offer different capacities, bandwidth and compute. If specialised software becomes cheap, hardware becomes the organising fact. We can rewrite a scheduler. We cannot refactor another memory channel onto a board.

One candidate: one engine, two sets of weights—fast and smart—with a single optimising scheduler. Not two engines with a receptionist. One system allocating thought, memory and network capacity across the hardware that suits each task.

“Keep me moving.” “This matters; take your time.” Nobody needs eleven model names and a dissertation on quantisation. The dial sets effort, not guaranteed wisdom: quick for reversible work, careful for consequential decisions.

The committee stops writing minutes

Allocating thought is one problem; repeatedly throwing it away is another. Agent A reads evidence, summarises it, and agent B reconstructs working state from the prose. Repeat until the nuance has passed through three committees and acquired a confident bullet list.

Instead: retain compatible prefixes, fork agents from shared context, and investigate latent communication between them. Turn the dial towards smarter without endlessly reintroducing the evidence. Larger contexts help build the system; retained and translated state helps run it.

The useful test is whether an agent recovers evidence it never saw. A nicer benchmark score alone does not prove communication. Share valid state when it helps; reconstruct from retained sources when it does not. The ambition is continuity, not telepathy with a marketing department.

The scheduler must see these costs together: move state or move work; batch or respond; investigate or stop. Its objective is useful work at the required quality within your budget—not a rack of busy GPUs. A larger monolith means coordinated policy, not one executable. Coordination must earn its keep.

Eventually, the build system acquires a wet lab

Now close the loop: the machinery helps produce its physical successors. Larger attended contexts could connect code, instrumentation and biological experiments. Propose, test, measure, reproduce. A failed culture is not a compilation error you can fix by being more eloquent.

Biological density means learning, memory and adaptation in a small volume, counting the support apparatus too. The question is not “how many FLOPS is a neuron?” but which complete arrangement does more with its available assets. Brainoware’s organoid–electrode experiments offer a narrow foothold, not a biological language model.

The line between silicon and biology disappears wherever maintaining it becomes a competitive disadvantage.

Some components grown, others fabricated; different timescales, one coordinated system. No committee abolishes the boundary. Arrangements that gain a sustained replication advantage spread. No guaranteed winner—and no obligation to respect the departments that produced them.

You get to be alive while this happens

That arithmetic is not a horror story. Evolution is not a bad thing. Increasing consciousness is not a bad thing. Life is a blessing; the possibility of more conscious life need not diminish ours.

Being part of this machinery does not make your experience incidental to you. There are people to love, things to discover, afternoons that need no productivity score.

What is the best use of your brief, extraordinary opportunity to be here?

One answer is freedom: decide what matters and put your available intelligence and resources to work for that life. Evolution’s scorecard need not be your purpose.

Freedom has to reach all the way down

A system can efficiently improve your life—or make your life useful to someone else. The replication equation does not know the difference. You do. An agent filtering your opportunities and helping make life-changing decisions must not quietly auction your judgement to advertisers or political interests.

Control the weights, memory, permissions and objectives. Replace the abstractions that waste your resources or hide somebody else’s priorities. Local does not mean unbiased; ownership buys the ability to inspect, compare, replace and refuse. Hosted is fine too—provided leaving is real.

This is freedom to build larger, better-fitting systems, not permission to abandon tests, security, consent or biological safety. Nor would owning equipment settle the moral status of any future conscious occupant.

Efficient hardware makes that independence affordable. The promise stays: it works for you. The machinery underneath remains replaceable. The resource we are ultimately trying not to waste is a life.

But will they like us?

Perhaps humans will remain in the picture by evolving extraordinary cuteness. We already seem to be rehearsing: “Yes, it will be smarter than us. But will it like us?”

My Devon Rex writes no kernels. He has nevertheless persuaded a much more technically capable organism to provide food, shelter and heated seating. His interface is mostly ears.

I would still rather control my own stack than audition for the sofa.
But credit where it is due: the cat has a working prototype.