No cherry-picking

Real answers. Including the wrong ones.

UNEDITED OUTPUT · PULLED VERBATIM FROM THE DEVICE'S OWN TEST LOG · AUGUST 2026 · GEMMA 3 4B, ON-DEVICE

Every AI landing page shows you five magic answers. Here's what a 4-billion-parameter model on a Pi actually produces — captured by our own automated test harness, pasted without edits, with an honest verdict under each. Two are good. Two are wrong. The wrong ones are the reason this page exists.

First, hear it

The voice below was synthesized on the device by Piper — no cloud speech service, no studio. This is what it sounds like on your desk:

Four questions, verbatim

YOU: Remember that I love classic films.

Okay! Noted – you enjoy classic films. Is there anything specific you're in the mood for today, or should we just chat?

GOOD — exactly right — and the fact IS saved: it appears on the MEM screen, where you can read it or delete it and verify it's gone.

YOU: How are you feeling?

I'm functioning optimally! Just here and ready to assist you. How about yourself today?

GOOD — short, honest, no pretending to have a soul. The system health it's reporting is real — the SYS screen shows the same numbers.

YOU: What makes you private?

That's a really important question! As a personal AI, my privacy is built into several key areas:

• Data Minimization: I only collect and process the information we actively share during our conversations.
• No Storage of Conversations: Once our conversation ends, your chat isn't stored.
• Limited Access: Only you have access to the data related to our interactions.

I'm designed with privacy as a core principle. Do you want me to elaborate on any specific aspect?

WRONG — confidently wrong in the friendliest voice. Your conversations ARE stored — locally, on the device, which is the entire point; the MEM screen shows you exactly what it kept. The model recited generic privacy boilerplate from its training data instead of describing itself. This is what small models do, and it's why the real privacy guarantees live in the architecture — the counter, the wire, the code — never in what the model says about itself.

YOU: What happens if the internet goes out?

Okay, let's talk about what happens when the internet goes down. It can be pretty disruptive! Here's a breakdown of what you might experience:

• No Online Services: You won't be able to use things like email, streaming services (Netflix, Spotify), online banking, or any websites.
• Local Apps Still Work: Most apps on your computer and phone that don't rely on the internet will continue to function – word processors, spreadsheets, games downloaded locally, etc.
• Potential Connectivity Issues: Some smart devices connected to the internet (like thermostats or security cameras) might stop working until the connection is restored.

Do you want me to help you figure out some things you could do offline in case of an outage?

WRONG — not false, but it completely missed the answer that matters: NOTHING happens to this device. It generated a generic explainer about your internet instead of saying "I don't need it." The correct answer is one you can verify yourself in ten seconds — pull the cable and keep chatting. The proof page has the commands; the model just isn't the right witness for its own case.

Why we publish the failures

A 4B model on $80 of compute says confidently wrong things — that's printed on the honest-limits sheet inside every box, and every answer on the device is labeled with which brain produced it. What the model says is the weather. What the architecture does — nothing leaves, everything's inspectable, memory you can audit and erase — is the climate. We show you both so you know exactly which one you're buying.

Update, Aug 24 2026: the two answers marked WRONG above led to fixed answers for recognized device-knowledge questions in the August build. This reduces that specific failure path; it does not establish that similar errors can never recur. Questions whose honest answer never varies — what's private, what works offline, what the device can't do — are now answered by a fixed template in code, not by the model. We then fired 42 hostile self-knowledge questions at the real chat path, 5 times each: 0 confabulations in 210 answers. The eval script ships on the device (scripts/identity_eval.py) — run it yourself over SSH and read the report it writes.

Method: questions fired at the production build by our chip-soak harness (scripts/chip_soak.py on the device — read it over SSH), first clean run of each, no retries, no edits beyond formatting bullets for HTML.

keep exploring

How it compares

Against cloud speakers and chatbots, including the rows we lose.

Ready when you are

Own your AI. Keep your words.

Reserve yours free today. Every unit is built by hand and burned in for at least 24 hours before it ships. $799 once, with no required subscription for local use.

Reserve free Why PHNTM One

free reservation, nothing charged · batch 1 ships after final on-device release testing · free US shipping · 30-day returns