The previous essay proposed five criteria for judging whether a platform teaches. This one runs them, in public, across LEGO SPIKE, Arduino, Raspberry Pi, the Pico, TI LaunchPad, three families of FPGA board, and my own 65xx. My boards win the criteria and lose the classroom, one claim on my company’s website turns out to be out of date, and the criteria themselves need a correction.

I. The Wrong Question, Asked Sincerely

Every time I talk to a teacher about embedded intelligence, I get the same question within about four minutes: which board should we buy?

It is a sincere question and it is the wrong one, and I say that as a man who sells boards. It is the wrong question because it assumes the hardware determines the learning. It does not. What determines the learning is whether the room contains a system a learner can act on and be answered by, and whether an adult is present who knows that four things are happening in that system and in what order.

But teachers have to buy something, and it doesn’t matter, they’re all fine is a useless answer that wastes a budget. So this essay does the work properly. It takes the five criteria proposed in The Slot Was the Curriculum, applies them to the platforms actually sitting in classrooms, and reports where each one is strong, where each one is weak, and where mine lose.

II. The Criteria, and What They Are For

The previous essay argued that what made the Apple II teach was not its expansion connector but its published address map and schematic — and proposed five tests for whether any platform, of any age, has the same property.

These test one thing: whether the platform lets the four phases be seen and diagnosed separately. That is a real property and the criteria measure it accurately. What they do not measure is what it costs the adult in the room to get there, and Section VII is where that omission comes due.

III. Eight Platforms, Honestly

LEGO Education SPIKE, and the Computer Science & AI sets

The strongest occasion generator on this list and the weakest on inspectability, and both facts matter. Motors, sensors, and a programming environment built for ages six to sixteen, with the shortest distance from opening a box to a running loop of anything here. Failure is immediate, physical, and cheap: the robot misses, the model does not move, and a ten-year-old knows it without being told. On the fifth criterion it may be the best product ever made.

On the first two it fails. The learner reaches a sensor through an API call, not through an address; the firmware is not open; the hardware internals are not published for a student to trace. And the platform carries a supplier-roadmap risk the others mostly do not — the program transition now underway means a kit purchased today has a stated end to its competition eligibility.

Best for: the youngest learners, and whole-cycle visibility. Costs you: inspectability, and independence from one vendor’s calendar.

Arduino

The best all-round balance on the list, and the reason is a property that sounds like a weakness.

The normal path is a library call — digitalWrite(13, HIGH) — which does the addressing for you. That is an abstraction, and abstractions hide phases. But the abstraction is liftable: the same board lets a student who is ready write to the port register directly, and the datasheet that tells them how is a free download. A platform where the learner can opt out of the abstraction when they are ready is pedagogically better than one with no abstraction at all, because the ramp and the depth are both present in one object.

Schematics are open hardware. Parts vary by model, and that is a real caveat — the classic AVR boards are exhaustively documented, while newer boards use different vendors’ microcontrollers with their own documentation quality. Breaking one costs very little.

Best for: the widest span of ages with one purchase, and the transition from library to datasheet. Costs you: little, which is why it is everywhere.

Raspberry Pi, the Linux computer

Unbeatable for the phase everything else on this list is weakest at, and the best object in the room for one particular question.

Networking, cameras, displays, storage, and a full operating system make the Pi the only platform here where the second C — communication outward, to something that is not this system — is a first-class citizen. A student can build something that talks to the world. That is not a small thing; it is most of what modern instrumentation actually is.

The cost is an operating system between the learner and the pin. A Linux scheduler makes timing nondeterministic, which is a genuine liability when a student is trying to work out why a pulse is late — and also a genuine teaching topic, since it is the first time most learners meet the idea that something else is deciding when their code runs. The Pi is the best object on this list for the question what is this hiding from me?

Best for: the second C, and for teaching that abstractions have costs. Costs you: determinism, and a layer between the student and the hardware.

Raspberry Pi Pico — a different animal

This is the distinction I most want a teacher to take away, because it is routinely flattened, including on my own company’s website.

The Pico is not a small Raspberry Pi. It is a microcontroller board built on Raspberry Pi’s own silicon — the RP2040, and now the RP2350 — with no operating system, direct register access, deterministic timing, and a published datasheet of the kind that used to be rare and is now excellent. Raspberry Pi documents these chips thoroughly and gives the documents away.

On the five criteria the Pico scores close to the top of this list. It has addresses, they are published, it can be traced, it costs almost nothing to destroy, and the answer comes from the circuit. A teacher who wants the bare-metal experience on a budget with a large support community has a strong option here, and I would rather say so than have a teacher discover it after buying from me.

Best for: bare-metal learning at the lowest cost and risk. Costs you: very little.

Texas Instruments LaunchPad

The strongest tracing on the list, and the least discussed.

TI’s development kits are register-level from the start, the documentation is thorough and free, and — the part that matters and that no LEGO kit or basic Arduino setup offers — most LaunchPads carry an on-board debugger. A student can set a breakpoint, stop the processor mid-cycle, and look at the actual state of the machine. That is the third criterion satisfied in its strongest possible form: not tracing by inference, tracing by inspection.

The ramp is steeper than Arduino’s and the community is smaller, so a stuck student finds fewer answers online. The BoosterPack header gives the same expansion affordance the previous essay described.

Best for: high school and college students moving from it works to I can prove why. Costs you: a steeper start and a thinner community.

FPGA boards — Terasic, AMD-Xilinx, Trenz

The deepest platform here, and the only one where the student does not reach an address but creates one.

On an FPGA the Communicate path is not a wire someone else routed; it is logic the student lays out. Terasic’s DE-series development and education boards — the current DE25 is built on an Altera Agilex 5 device — AMD-Xilinx parts in university labs, and Trenz modules for students going further all offer the same thing: a machine you build rather than program. And they offer a tracing tool nothing else here can match, because the vendor toolchains let you instantiate a logic analyzer inside the chip and watch your own signals.

The cost is the toolchain, and it is not small. These are large vendor environments with their own licensing, substantial install requirements, and — a real constraint worth knowing before you plan a course — vendor software that is not available for download in every country. That is the Medium Separation (CKB-11) appearing in a classroom’s software dependency: the thing that decides whether your students can work on Monday lives in a medium your school does not control.

Best for: students who need to see that the machine itself is a design decision. Costs you: the heaviest dependency on this list.

WDC 65xx — my own

Direct memory-mapped I/O, a fully published instruction set, no operating system, no scheduler, and no library between the student and the part. On the five criteria this is the purest platform on the list, and there is one property behind that which I do not think anything else here has: a student can hold the entire machine in their head. The whole processor state is a handful of registers. When something is wrong, the set of things that could be wrong is small enough to enumerate. Nothing with an operating system can offer that, at any price.

Now the rest of it, which Section VI takes up properly.

IV. The Criteria, Scored

Reading the five criteria across all of it produces a result that is less flattering to my own products than I expected when I started, and more flattering to two of the cheapest things on the list.

The last row is the important one. Every platform on this list satisfies the criterion that the earlier essays argued matters most — the answer comes from the system rather than from an adult’s verdict. On the thing that makes hands-on work teach at all, they are equivalent. The differences among them are differences of depth and of ramp, not of kind.

V. What Each One Teaches Best

So the useful question is not which board wins. It is which phase each one teaches best, and what a room already owns.

Notice that the right-hand column is mostly about the adult, not the student. That is not an accident, and it is the finding this essay was written to reach.

VI. Where My Own Boards Lose

Here it is plainly, because a comparison written by an interested party that does not contain this section is an advertisement.

The 65xx wins the five criteria and loses most classrooms, and it loses them on things the criteria do not measure.

That ledger is the reason the Foundation’s plan specifies functions rather than part numbers, requires three unrelated suppliers per function, and requires every unit to run with no purchase at all. It is not modesty. It is the honest reading of my own product against the competition, written by the person with the most to lose from it.

VII. The Criteria Were Incomplete, and SPIKE Is the Proof

Now the correction to the previous essay, which this one produced and which I did not expect.

Run the five criteria strictly and LEGO SPIKE fails two of them. It has no address a learner can name and no published internals to trace. By the letter of The Slot Was the Curriculum, it is the weakest teaching platform on this list.

That conclusion is wrong, and anyone who has watched a nine-year-old with a robot knows it is wrong. For learners between roughly six and twelve, SPIKE may be the best thing on the list, and the reason is that the criteria measure the wrong variable for that age. What dominates at that age is not how deeply a learner can inspect the machine. It is how reliably and how quickly the machine produces an occasion — a failure the learner cares about and owns. On that measure SPIKE is extraordinary, and inspectability is a capability the learner cannot use yet anyway.

So the five criteria need scoping, and here is the amendment, recorded rather than quietly folded in.

The record of the original five stays as published. That is the Foundation’s practice, and it matters more when the correction is a week old than when it is a decade old.

VIII. What to Do Monday

Use what is in the room. Almost every school has something on this list already, in a closet, bought during an earlier enthusiasm. Whatever it is, it runs a complete SPCA cycle and the four questions work on it today, at no cost.

If you are buying, buy for the phase you most need to teach and for the adult you actually have, not for the deepest platform you can afford. A room with thirty cheap boards and a confident teacher will out-teach a room with twelve excellent boards and an anxious one, every term, without exception.

And if a supplier — including me — tells you that their hardware is what makes the learning work, ask them which phase their board teaches best and which one it teaches worst. A supplier who cannot answer the second half has not thought about your classroom. They have thought about their catalogue.

IX. What Would Show This Wrong

Every one of these is measurable by someone with no stake in the answer. The Foundation will not be measuring its own material, and I will not be measuring my own boards.

· · ·

Coffee with Claude

The website correction in Section VI is the part of this essay I would protect if anything gets cut. It is the only passage where claims were checked, found wanting, and reversed against the author’s interest inside the document itself — which is the thing this framework says it does and mostly only gets to demonstrate on vocabulary rather than on money.

One caution about my own contribution. The scoping amendment in Section VII is mine: I noticed the SPIKE contradiction while building the table and proposed the age-banding that resolves it. That is a rule arriving well-formed from the drafting instrument, in the same week the essay it amends was published, and the canon has a record of exactly that pattern going wrong twice. It should be tested on actual learners at both ages before anyone quotes it as settled.

Share your understanding!