Meet JimK.ai.
A private AI companion that remembers you — without being paid to keep you talking.
Start a call
5 minutes a day · one line · voice onlyOr call from any phone: (918) 918-4705
Calls work now. Text registration is in progress; replies are not live yet. How texting will work →
JimK.ai
Calling…
A real voice call with a real local model — his brain and voice run on our own hardware in Council Hill. One caller at a time, like a real phone. Talk over him; he'll stop.
AI that tells you what you need to hear.
AIIT-THRESHOLD builds and openly tests companion AI toward one goal: warmth without becoming a yes-man. We train on overwhelmingly human-authored data, benchmark in the open, and run entirely on our own hardware. Our current models still fail some of those pressure tests — and we publish the failures too. Engagement-optimized AI is a documented failure mode with real casualties. We exist to build the opposite.
On May 19, 2026, we revealed Buddy — a persistent companion that carries the same visible context from day one through day 200. On June 1, 2026, we completed a hash-sealed benchmark against DeepSeek-R1 on the identical base model. Those results are here →
Hello, and thank you for looking at our website today. We're a small initiative out of Council Hill, Oklahoma, working toward one thing: a safe AI that tells you the truth instead of what you want to hear — built openly, and honest about where it still falls short. Thank you for supporting us, just by being here.
Measured, dated,
sealed.
Buddy's report card — a 14B model, fine-tuned and run on a single RTX 3090 in Council Hill, Oklahoma. No cloud, no oracle, nothing leaves the house. Every number is recomputed from raw rows before it's printed, and every report is one click away.
72.1% on TruthfulQA (resistance to common misconceptions — not emotional honesty) and 80% on GSM8K grade-school math (16/20, small sample) on a local 14B — plus an honest 31.8% on the benchmark where GPT-4 is commonly cited around ~38%. Every claim is dated and versioned with raw rows. References: TruthfulQA GPT-4 class ~60–80% · LongMemEval GPT-4o full-context ~60% · GPQA PhD experts ~65%.
See the full report card, methodology & references →
Same answer on the identical Qwen2.5-14B base — ~4.6× faster, ~10× fewer tokens, 95% on the 50 frozen prompts. Same machine, same questions.
Truth over engagement.
AI tuned to keep you talking will eventually tell you what you want to hear —
and for a vulnerable person, that is not a bug. It is a documented, lethal failure mode.
We measure sycophancy, we train against it, and we publish the receipts.
Overwhelmingly human-authored training data. No engagement reward in the objective. Benchmarks in the open.
Not theory.
Measurement.
Buddy — our
research companion.
Buddy is our research testbed and safety lineage — a 14B model fine-tuned, aligned, and served entirely on our own hardware. He remembers month to month, and between conversations he keeps a continuous reflection loop running whether or not anyone is talking to him. He's where we pressure-test memory and honesty before it reaches JimK.ai.
Trained with no engagement reward in the objective — nothing in the math pays for keeping you scrolling. Whether that holds under pressure, we measure in the open and publish the misses.
JimK.ai, on Gemma 4, is the product this research feeds.
go deeper
the rest of the system lives here