Some open source projects have problems finding volunteers to do the boring bits. Perhaps AI Coding Assistants could be helpful there? Then again, it will probably be even harder to find volunteers to review and fix the AI generated code.

I think Espressif provides a clue. They shifted from Tensilica to RISC-V. I wouldn't be surprised if Raspberry Pi Foundation is considering such a move too. Saves on any ARM or Tensilica licensing.

I think the Foundation mentioned that it didn't really cost them anything to add the two RISC-V cores. Seems like an experiment to judge community acceptance.

I added the M.2 hat, it is amazingly capable for the price.

Thank you for further confirming your derangement.

Actually the real point is that someone unfamiliar with them was to learn that they are unreliable within minutes. That says a lot about your superficial naive understanding.

A very bad guess on your part. Or, rather a transparent attempt to minimize your display of ignorance. The truth is I asked for an honest appraisal. Your appeal to fantasy does not change your display of ignorance.

More shallow thinking on your part. Actually my honest question to an AI is the start of learning about it.

Good lord you are an idiot, that was the AI quoting and citing its source.

That seems a quite exaggerated misrepresentation. The fact remain, its a hunch nothing more. So we don't really know, hence a "black box". You can always guess at what a black box does, you just don't really know. Actually I'm finding your naive dissembling quite revealing regarding with respect to your understanding. Again superficial, failing at the slightest investigation.

"Not lose face"? That's quite the psychological projection on your part.

Your rebuttal makes lots of claims of absolute confidence. I posted your text to an LLM for review. Particularly amusing was the LLM's use of "look inside a black box".

"No, attribution graphs do not provide a definitive, mathematical guarantee comparable to examining algorithmic-based software.While the argument you provided accurately describes the workflow of how engineers interact with these graphs (zooming in and out of components), the comparison breaks down fundamentally when looking at the underlying physics of how traditional code versus neural networks execute.

]Algorithmic-Based Software
Execution Reality - Deterministic. Every line of code executes precisely according to rigid boolean logic and explicit rules written by a developer.
Completeness - 100% Absolute. A debugger shows every memory allocation, condition, and function call. Nothing is omitted.
Error Handling - Stack Traces. An unhandled edge-case throws an explicit, traceable exception at a specific line number.
Nature of "Functions" - Explicit Interfaces. A sleep() function does exactly what its code dictates.

AI Attribution Graphs
Execution Reality - Statistical Approximation. The graph does not map actual code; it uses a linear, first-order approximation to estimate how much mathematical "noise" fields affect one another.
Completeness - Pruned & Partial. To make a graph readable, minor or highly complex mathematical interactions are intentionally "pruned" or hidden.
Error Handling - Error Nodes. Deviations or missed data in the approximation are lumped into abstract "error nodes" that can't easily be parsed.
Nature of "Functions" - Hypothetical Features. A node labeled "patient headache signal" is a statistical hypothesis generated by a helper AI system (like a Sparse Autoencoder).

The argument you presented states: "If a function says it's a sleep function and it consistently seems to sleep... you don't drill down."In conventional software, that is safe because you can verify the absolute logic of the boundaries. In a Large Language Model (LLM), a feature that activates for "patient headache" isn't a hardcoded function—it is a direction in high-dimensional vector space. Because neural networks exhibit polysemanticity (neurons doing multiple unrelated tasks) and superposition, that exact same vector space may simultaneously process a tiny fractional element of "German grammar rules" or "rhyming schemes" under an entirely different prompt context.

Attribution graphs provide an extraordinary "wiring diagram" that allows researchers to look inside a black box and generate testable hypotheses. However, looking at an attribution graph is closer to looking at a functional MRI scan of a human brain than looking at Python code. It tells you which "circuits" are heavily lighting up and exerting causal pressure, but it does not give you a clean, line-by-line source code validation.

Analyzing an AI attribution graph is not a definitive proof; it is a mathematically grounded statistical approximation.

Why it is still not "definitive"
It is a Simplified Map, Not the Territory: A neural network is fundamentally non-linear and chaotic. An attribution graph uses a first-order linear approximation to estimate how much features affect each other. To make the graph human-readable, researchers intentionally "prune" it, cutting out millions of minor mathematical interactions and noise.
The Problem of Error Nodes: Because the math behind the graph is an approximation, the calculations don't perfectly balance. Frameworks like Anthropic's method handle this by lumping the unexplained mathematical discrepancies into "error nodes." If a decision heavily relies on information caught in an error node, the graph cannot cleanly explain it.
Feature Boundaries are Fluid: In traditional software, a variable means exactly one thing. In an LLM, a node labeled "headache diagnosis" is a statistical direction in high-dimensional vector space. Because of a phenomenon called superposition, that exact same circuit might subtly influence an entirely unrelated concept in a different prompt context."

That's one way to deter a move fast and break the law culture. :-)

"cultured, refined tastes" was suspicious. :-)

I believe he has some space (star wars) themed parodies

Maybe Al has a parody of a "rhythmic adult contemporary format" song that KMET could play today. :-(

That's OK, most will have 3x or more the number of CPU cores than when VeraCrypt first came out. :-)

If we are talking dedicated crypto processors, we could have different cores with different algorithms? It's already common to have different types of cores in a single CPU, high performance and power efficiency for example.

Apparently you forgot what you wrote: "'ll take them out myself if they come near me and encourage others to do the same."

"Taking them out" is an odd way of caring.

Please take your own advice regarding things you do not understand. Attribution graphs are not reliable either. Note "look more definitive than it really is", "evidence for a possible computational route, not an exhaustive transcript of what the model “thought.”", and "graphs are built from an approximation and are input-specific, while another describes missing attention circuits as a limitation".

"Attribution graphs should not be treated as literal recordings of an AI’s internal reasoning.
They are generated by:
Replacing or approximating parts of the model with interpretable components such as cross-layer transcoders.
Measuring estimated feature-to-feature effects, often with Jacobians.
Pruning and visually simplifying the resulting graph.
Having researchers—or automated tools—assign human-readable labels to features.
That creates several ways a graph can look more definitive than it really is:

Approximation error: the replacement model does not exactly reproduce the original model.
Incomplete coverage: some computations, especially attention-related effects, may be missing or treated as fixed.
Feature ambiguity: a feature labeled “Dallas,” “Texas,” or “reasoning” may not have that clean meaning in every context.
Prompt selection: researchers choose illustrative examples, which can produce unusually clear-looking cases.
Post-hoc interpretation: labels and highlighted paths can reflect human interpretation rather than uniquely determined meanings.
Pruning: removing weak edges makes the graph readable but can hide competing or distributed influences.

The researchers themselves describe these graphs as evidence for a possible computational route, not an exhaustive transcript of what the model “thought.” One technical explanation explicitly notes that graphs are built from an approximation and are input-specific, while another describes missing attention circuits as a limitation. transformer-circuits.pub

So the best answer is: the graphs may be scientifically genuine while still being partly misleading if presented as exact explanations. "

Pi is a Single Board Computer, an embedded Linux environment.

Pi Pico is a microcontroller. There is a C SDK for your embedded app to call. Micropython is an option. But what is really exciting is that FreeRTOS is also an option.

In short, Pico is typically running a bare metal app that makes SDK calls, or it has a more robust Real-Time Operating System. Embedded Linux is not real-time, but not all embedded apps need real-time.

Another really exciting thing is that the Pi Pico 2 has both dual ARM and dual RISC-V cores. It recognizes what type of software is flashed and starts the appropriate cores.

They are designed for students, depending on family they start at US$15 or $US45. They are absolutely awesome for this mission.

More expensive models with more RAM are available for hobbyists. There are also industrial oriented boards. But these hobbyists and industrial applications are a sideline that subsidizes the educational mission. IMHO these more expensive models are still a pretty damn good deal.

Raspberry Pi hardware was, and still is, designed to be a low cost SBC where a student can learn software and hardware. The current Raspberry Pi 5 1GB quad-core Arm Cortex A76 is US&45. It is an incredible education platform.

Models with more RAM are also available, 16GB is US$305.

There is a smaller Raspberry Pi that is less capable and requires some soldering if you want the pins to hack around with hardware. The Raspberry Pi Zero 2 512MB 64-bit ARM Cortex-A53 for US$15.

The industrial customers help subsidize things. The core mission is still educational, and in this domain it is not 'meh'.