I read @AnthropicAI's IPO filing today, and the S-1 says the insurance industry might not even be big enough to cover its worst day. If the insurers can't price the risk, I'm not sure how retail investors are supposed to do that either. sec.gov/Archives/edgar/data/…

How do you know an AI actually represents an expert? There are four pillars: accuracy, groundedness, persona, and judgment. At @OnixHQ, we call it the expert fidelity score.

At @OnixHQ, we're creating the official, verified version of an expert's AI. The master copy. Everything else is the bootleg. The Napster, the LimeWire.

At @OnixHQ, we evaluate our expert AI on persona too. Some experts are bubbly and upbeat. Others are dry and witty. That's part of their practice. It's almost like the bedside manner of a doctor.

My toaster has more rules than AI does. We need some basic guardrails around what it means to build AI systems.

Generative AI is a probabilistic process. There's literally no way to get hallucinations to 0%. What you can control is what the model reaches for.

Our first official Tiny Labs event at @OnixHQ in Montréal, with Jeremy Pinto. Let's not just do everything with these mega models, but focus on classification, on decisions. Thank you so much, everyone.

A dietitian, a doctor: these people can only see so many patients in a day. They're already booked. Lined up out the door, around the corner, for months on end.

What does it look like when you have 10,000 AI agents that move faster than a human defender, attacking critical infrastructure?

We're at a time where you don't have to trade your privacy for utility. We can build companies whose incentives are aligned with the users, where the users aren't a product.

Thursday at Onix Tiny Labs in Montréal: Jeremy Pinto on Jev and small decision models. luma.com/an3bsgov

Big god models are trained on everything from the history of Japan to how to code. At @OnixHQ, we're domain constrained to the specific expertise of each expert.

As CTO at a data labeling company, I found my cringe posts from when I was 14 in AI data sets. I'm not getting compensated for this. I didn't give anybody permission for this.

My Oura ring is always telling me I'm not sleeping well enough, that I'm stressed. It doesn't have the context that I have two young kids.

We're so excited to build small language models here in Montréal, pushing the frontiers of privacy, safety and true expert-led AI. Thank you so much to everyone that met up with @OnixHQ yesterday at the @Mila_Quebec Career Fair. See you around. #ThinkSmall

The core belief for us at @OnixHQ: expertise is factual knowledge plus a point of view, a perspective and judgment. Not just the statistical average.

When I deploy AI agents, is it aligned to me personally? Or is it actually aligned to some big tech company and their incentives?

Oct 15 at Onix Tiny Labs in Montréal: Jeremy Pinto on why @typesafeai's Jev is interesting, and a demo of how to build and run small decision models. luma.com/an3bsgov

Shared a few thoughts with ITPro on the data center backlash. Most of the concerns are real. For companies buying AI: measure cost and energy per task, size the model to the job, and run inference close to the edge. itpro.com/infrastructure/the…

Our experts are with us every day, even at snack time. 🥜 Today's office pantry note comes from @ashleykoff's onix at @OnixHQ. Keep asking.